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 #ifndef PIPSIZ 973 #define PIPSIZ 57344 974 #endif 975 static u_long unpst_sendspace = PIPSIZ; 976 static u_long unpst_recvspace = PIPSIZ; 977 static u_long unpdg_sendspace = 2*1024; /* really max datagram size */ 978 static u_long unpdg_recvspace = 4*1024; 979 980 SYSCTL_DECL(_net_local_seqpacket); 981 SYSCTL_DECL(_net_local_stream); 982 SYSCTL_INT(_net_local_stream, OID_AUTO, sendspace, CTLFLAG_RW, 983 &unpst_sendspace, 0, "Size of stream socket send buffer"); 984 SYSCTL_INT(_net_local_stream, OID_AUTO, recvspace, CTLFLAG_RW, 985 &unpst_recvspace, 0, "Size of stream socket receive buffer"); 986 987 SYSCTL_DECL(_net_local_dgram); 988 SYSCTL_INT(_net_local_dgram, OID_AUTO, maxdgram, CTLFLAG_RW, 989 &unpdg_sendspace, 0, "Max datagram socket size"); 990 SYSCTL_INT(_net_local_dgram, OID_AUTO, recvspace, CTLFLAG_RW, 991 &unpdg_recvspace, 0, "Size of datagram socket receive buffer"); 992 993 static int 994 unp_attach(struct socket *so, struct pru_attach_info *ai) 995 { 996 struct unp_global_head *head; 997 struct unpcb *unp; 998 int error; 999 1000 lwkt_gettoken(&unp_token); 1001 1002 if (so->so_snd.ssb_hiwat == 0 || so->so_rcv.ssb_hiwat == 0) { 1003 switch (so->so_type) { 1004 case SOCK_STREAM: 1005 case SOCK_SEQPACKET: 1006 error = soreserve(so, unpst_sendspace, unpst_recvspace, 1007 ai->sb_rlimit); 1008 break; 1009 1010 case SOCK_DGRAM: 1011 error = soreserve(so, unpdg_sendspace, unpdg_recvspace, 1012 ai->sb_rlimit); 1013 break; 1014 1015 default: 1016 panic("unp_attach"); 1017 } 1018 if (error) 1019 goto failed; 1020 } 1021 1022 /* 1023 * In order to support sendfile we have to set either SSB_STOPSUPP 1024 * or SSB_PREALLOC. Unix domain sockets use the SSB_STOP flow 1025 * control mechanism. 1026 */ 1027 if (so->so_type == SOCK_STREAM) { 1028 atomic_set_int(&so->so_rcv.ssb_flags, SSB_STOPSUPP); 1029 atomic_set_int(&so->so_snd.ssb_flags, SSB_STOPSUPP); 1030 } 1031 1032 unp = kmalloc(sizeof(*unp), M_UNPCB, M_WAITOK | M_ZERO | M_NULLOK); 1033 if (unp == NULL) { 1034 error = ENOBUFS; 1035 goto failed; 1036 } 1037 unp->unp_refcnt = 1; 1038 unp->unp_gencnt = ++unp_gencnt; 1039 LIST_INIT(&unp->unp_refs); 1040 unp->unp_socket = so; 1041 unp->unp_rvnode = ai->fd_rdir; /* jail cruft XXX JH */ 1042 so->so_pcb = (caddr_t)unp; 1043 soreference(so); 1044 1045 head = unp_globalhead(so->so_type); 1046 TAILQ_INSERT_TAIL(&head->list, unp, unp_link); 1047 head->count++; 1048 error = 0; 1049 failed: 1050 lwkt_reltoken(&unp_token); 1051 return error; 1052 } 1053 1054 static void 1055 unp_detach(struct unpcb *unp) 1056 { 1057 struct socket *so; 1058 1059 lwkt_gettoken(&unp_token); 1060 lwkt_getpooltoken(unp); 1061 1062 so = unp->unp_socket; 1063 1064 unp->unp_gencnt = ++unp_gencnt; 1065 if (unp->unp_vnode) { 1066 unp->unp_vnode->v_socket = NULL; 1067 vrele(unp->unp_vnode); 1068 unp->unp_vnode = NULL; 1069 } 1070 soisdisconnected(so); 1071 KKASSERT(so->so_pcb == unp); 1072 so->so_pcb = NULL; /* both tokens required */ 1073 unp->unp_socket = NULL; 1074 1075 lwkt_relpooltoken(unp); 1076 lwkt_reltoken(&unp_token); 1077 1078 sofree(so); 1079 1080 KASSERT(unp->unp_conn == NULL, ("unp is still connected")); 1081 KASSERT(LIST_EMPTY(&unp->unp_refs), ("unp still has references")); 1082 1083 if (unp->unp_addr) 1084 kfree(unp->unp_addr, M_SONAME); 1085 kfree(unp, M_UNPCB); 1086 1087 if (unp_rights) 1088 taskqueue_enqueue(unp_taskqueue, &unp_gc_task); 1089 } 1090 1091 static int 1092 unp_bind(struct unpcb *unp, struct sockaddr *nam, struct thread *td) 1093 { 1094 struct proc *p = td->td_proc; 1095 struct sockaddr_un *soun = (struct sockaddr_un *)nam; 1096 struct vnode *vp; 1097 struct vattr vattr; 1098 int error, namelen; 1099 struct nlookupdata nd; 1100 char buf[SOCK_MAXADDRLEN]; 1101 1102 ASSERT_LWKT_TOKEN_HELD(&unp_token); 1103 UNP_ASSERT_TOKEN_HELD(unp); 1104 1105 if (unp->unp_vnode != NULL) 1106 return EINVAL; 1107 1108 namelen = soun->sun_len - offsetof(struct sockaddr_un, sun_path); 1109 if (namelen <= 0) 1110 return EINVAL; 1111 strncpy(buf, soun->sun_path, namelen); 1112 buf[namelen] = 0; /* null-terminate the string */ 1113 error = nlookup_init(&nd, buf, UIO_SYSSPACE, 1114 NLC_LOCKVP | NLC_CREATE | NLC_REFDVP); 1115 if (error == 0) 1116 error = nlookup(&nd); 1117 if (error == 0 && nd.nl_nch.ncp->nc_vp != NULL) 1118 error = EADDRINUSE; 1119 if (error) 1120 goto done; 1121 1122 VATTR_NULL(&vattr); 1123 vattr.va_type = VSOCK; 1124 vattr.va_mode = (ACCESSPERMS & ~p->p_fd->fd_cmask); 1125 error = VOP_NCREATE(&nd.nl_nch, nd.nl_dvp, &vp, nd.nl_cred, &vattr); 1126 if (error == 0) { 1127 if (unp->unp_vnode == NULL) { 1128 vp->v_socket = unp->unp_socket; 1129 unp->unp_vnode = vp; 1130 unp->unp_addr = (struct sockaddr_un *)dup_sockaddr(nam); 1131 vn_unlock(vp); 1132 } else { 1133 vput(vp); /* late race */ 1134 error = EINVAL; 1135 } 1136 } 1137 done: 1138 nlookup_done(&nd); 1139 return (error); 1140 } 1141 1142 static int 1143 unp_connect(struct socket *so, struct sockaddr *nam, struct thread *td) 1144 { 1145 struct unpcb *unp, *unp2; 1146 int error, flags = 0; 1147 1148 lwkt_gettoken(&unp_token); 1149 1150 unp = unp_getsocktoken(so); 1151 if (!UNP_ISATTACHED(unp)) { 1152 error = EINVAL; 1153 goto failed; 1154 } 1155 1156 if ((unp->unp_flags & UNP_CONNECTING) || unp->unp_conn != NULL) { 1157 error = EISCONN; 1158 goto failed; 1159 } 1160 1161 flags = UNP_CONNECTING; 1162 unp_setflags(unp, flags); 1163 1164 error = unp_find_lockref(nam, td, so->so_type, &unp2); 1165 if (error) 1166 goto failed; 1167 /* 1168 * NOTE: 1169 * unp2 is locked and referenced. 1170 */ 1171 1172 if (so->so_proto->pr_flags & PR_CONNREQUIRED) { 1173 struct socket *so2, *so3; 1174 struct unpcb *unp3; 1175 1176 so2 = unp2->unp_socket; 1177 if (!(so2->so_options & SO_ACCEPTCONN) || 1178 (so3 = sonewconn_faddr(so2, 0, NULL, 1179 TRUE /* keep ref */)) == NULL) { 1180 error = ECONNREFUSED; 1181 goto done; 1182 } 1183 /* so3 has a socket reference. */ 1184 1185 unp3 = unp_getsocktoken(so3); 1186 if (!UNP_ISATTACHED(unp3)) { 1187 unp_reltoken(unp3); 1188 /* 1189 * Already aborted; we only need to drop the 1190 * socket reference held by sonewconn_faddr(). 1191 */ 1192 sofree(so3); 1193 error = ECONNREFUSED; 1194 goto done; 1195 } 1196 unp_reference(unp3); 1197 /* 1198 * NOTE: 1199 * unp3 is locked and referenced. 1200 */ 1201 1202 /* 1203 * Release so3 socket reference held by sonewconn_faddr(). 1204 * Since we have referenced unp3, neither unp3 nor so3 will 1205 * be destroyed here. 1206 */ 1207 sofree(so3); 1208 1209 if (unp2->unp_addr != NULL) { 1210 unp3->unp_addr = (struct sockaddr_un *) 1211 dup_sockaddr((struct sockaddr *)unp2->unp_addr); 1212 } 1213 1214 /* 1215 * unp_peercred management: 1216 * 1217 * The connecter's (client's) credentials are copied 1218 * from its process structure at the time of connect() 1219 * (which is now). 1220 */ 1221 cru2x(td->td_proc->p_ucred, &unp3->unp_peercred); 1222 unp_setflags(unp3, UNP_HAVEPC); 1223 /* 1224 * The receiver's (server's) credentials are copied 1225 * from the unp_peercred member of socket on which the 1226 * former called listen(); unp_listen() cached that 1227 * process's credentials at that time so we can use 1228 * them now. 1229 */ 1230 KASSERT(unp2->unp_flags & UNP_HAVEPCCACHED, 1231 ("unp_connect: listener without cached peercred")); 1232 memcpy(&unp->unp_peercred, &unp2->unp_peercred, 1233 sizeof(unp->unp_peercred)); 1234 unp_setflags(unp, UNP_HAVEPC); 1235 1236 error = unp_connect_pair(unp, unp3); 1237 if (error) 1238 soabort_direct(so3); 1239 1240 /* Done with unp3 */ 1241 unp_free(unp3); 1242 unp_reltoken(unp3); 1243 } else { 1244 error = unp_connect_pair(unp, unp2); 1245 } 1246 done: 1247 unp_free(unp2); 1248 unp_reltoken(unp2); 1249 failed: 1250 if (flags) 1251 unp_clrflags(unp, flags); 1252 unp_reltoken(unp); 1253 1254 lwkt_reltoken(&unp_token); 1255 return (error); 1256 } 1257 1258 /* 1259 * Connect two unix domain sockets together. 1260 * 1261 * NOTE: Semantics for any change to unp_conn requires that the per-unp 1262 * pool token also be held. 1263 */ 1264 int 1265 unp_connect2(struct socket *so, struct socket *so2) 1266 { 1267 struct unpcb *unp, *unp2; 1268 int error; 1269 1270 lwkt_gettoken(&unp_token); 1271 if (so2->so_type != so->so_type) { 1272 lwkt_reltoken(&unp_token); 1273 return (EPROTOTYPE); 1274 } 1275 unp = unp_getsocktoken(so); 1276 unp2 = unp_getsocktoken(so2); 1277 1278 if (!UNP_ISATTACHED(unp)) { 1279 error = EINVAL; 1280 goto done; 1281 } 1282 if (!UNP_ISATTACHED(unp2)) { 1283 error = ECONNREFUSED; 1284 goto done; 1285 } 1286 1287 if (unp->unp_conn != NULL) { 1288 error = EISCONN; 1289 goto done; 1290 } 1291 if ((so->so_type == SOCK_STREAM || so->so_type == SOCK_SEQPACKET) && 1292 unp2->unp_conn != NULL) { 1293 error = EISCONN; 1294 goto done; 1295 } 1296 1297 error = unp_connect_pair(unp, unp2); 1298 done: 1299 unp_reltoken(unp2); 1300 unp_reltoken(unp); 1301 lwkt_reltoken(&unp_token); 1302 return (error); 1303 } 1304 1305 /* 1306 * Disconnect a unix domain socket pair. 1307 * 1308 * NOTE: Semantics for any change to unp_conn requires that the per-unp 1309 * pool token also be held. 1310 */ 1311 static void 1312 unp_disconnect(struct unpcb *unp, int error) 1313 { 1314 struct socket *so = unp->unp_socket; 1315 struct unpcb *unp2; 1316 1317 ASSERT_LWKT_TOKEN_HELD(&unp_token); 1318 UNP_ASSERT_TOKEN_HELD(unp); 1319 1320 if (error) 1321 so->so_error = error; 1322 1323 while ((unp2 = unp->unp_conn) != NULL) { 1324 lwkt_getpooltoken(unp2); 1325 if (unp2 == unp->unp_conn) 1326 break; 1327 lwkt_relpooltoken(unp2); 1328 } 1329 if (unp2 == NULL) 1330 return; 1331 /* unp2 is locked. */ 1332 1333 KASSERT((unp2->unp_flags & UNP_DROPPED) == 0, ("unp2 was dropped")); 1334 1335 unp->unp_conn = NULL; 1336 1337 switch (so->so_type) { 1338 case SOCK_DGRAM: 1339 LIST_REMOVE(unp, unp_reflink); 1340 soclrstate(so, SS_ISCONNECTED); 1341 break; 1342 1343 case SOCK_STREAM: 1344 case SOCK_SEQPACKET: 1345 /* 1346 * Keep a reference before clearing the unp_conn 1347 * to avoid racing uipc_detach()/uipc_abort() in 1348 * other thread. 1349 */ 1350 unp_reference(unp2); 1351 KASSERT(unp2->unp_conn == unp, ("unp_conn mismatch")); 1352 unp2->unp_conn = NULL; 1353 1354 soisdisconnected(so); 1355 soisdisconnected(unp2->unp_socket); 1356 1357 unp_free(unp2); 1358 break; 1359 } 1360 1361 lwkt_relpooltoken(unp2); 1362 } 1363 1364 #ifdef notdef 1365 void 1366 unp_abort(struct unpcb *unp) 1367 { 1368 lwkt_gettoken(&unp_token); 1369 unp_free(unp); 1370 lwkt_reltoken(&unp_token); 1371 } 1372 #endif 1373 1374 static int 1375 prison_unpcb(struct thread *td, struct unpcb *unp) 1376 { 1377 struct proc *p; 1378 1379 if (td == NULL) 1380 return (0); 1381 if ((p = td->td_proc) == NULL) 1382 return (0); 1383 if (!p->p_ucred->cr_prison) 1384 return (0); 1385 if (p->p_fd->fd_rdir == unp->unp_rvnode) 1386 return (0); 1387 return (1); 1388 } 1389 1390 static int 1391 unp_pcblist(SYSCTL_HANDLER_ARGS) 1392 { 1393 struct unp_global_head *head = arg1; 1394 int error, i, n; 1395 struct unpcb *unp, *marker; 1396 1397 KKASSERT(curproc != NULL); 1398 1399 /* 1400 * The process of preparing the PCB list is too time-consuming and 1401 * resource-intensive to repeat twice on every request. 1402 */ 1403 if (req->oldptr == NULL) { 1404 n = head->count; 1405 req->oldidx = (n + n/8) * sizeof(struct xunpcb); 1406 return 0; 1407 } 1408 1409 if (req->newptr != NULL) 1410 return EPERM; 1411 1412 marker = kmalloc(sizeof(*marker), M_UNPCB, M_WAITOK | M_ZERO); 1413 marker->unp_flags |= UNP_MARKER; 1414 1415 lwkt_gettoken(&unp_token); 1416 1417 n = head->count; 1418 i = 0; 1419 error = 0; 1420 1421 TAILQ_INSERT_HEAD(&head->list, marker, unp_link); 1422 while ((unp = TAILQ_NEXT(marker, unp_link)) != NULL && i < n) { 1423 struct xunpcb xu; 1424 1425 TAILQ_REMOVE(&head->list, marker, unp_link); 1426 TAILQ_INSERT_AFTER(&head->list, unp, marker, unp_link); 1427 1428 if (unp->unp_flags & UNP_MARKER) 1429 continue; 1430 if (prison_unpcb(req->td, unp)) 1431 continue; 1432 1433 xu.xu_len = sizeof(xu); 1434 xu.xu_unpp = unp; 1435 1436 /* 1437 * NOTE: 1438 * unp->unp_addr and unp->unp_conn are protected by 1439 * unp_token. So if we want to get rid of unp_token 1440 * or reduce the coverage of unp_token, care must be 1441 * taken. 1442 */ 1443 if (unp->unp_addr) { 1444 bcopy(unp->unp_addr, &xu.xu_addr, 1445 unp->unp_addr->sun_len); 1446 } 1447 if (unp->unp_conn && unp->unp_conn->unp_addr) { 1448 bcopy(unp->unp_conn->unp_addr, 1449 &xu.xu_caddr, 1450 unp->unp_conn->unp_addr->sun_len); 1451 } 1452 bcopy(unp, &xu.xu_unp, sizeof(*unp)); 1453 sotoxsocket(unp->unp_socket, &xu.xu_socket); 1454 1455 /* NOTE: This could block and temporarily release unp_token */ 1456 error = SYSCTL_OUT(req, &xu, sizeof(xu)); 1457 if (error) 1458 break; 1459 ++i; 1460 } 1461 TAILQ_REMOVE(&head->list, marker, unp_link); 1462 1463 lwkt_reltoken(&unp_token); 1464 1465 kfree(marker, M_UNPCB); 1466 return error; 1467 } 1468 1469 SYSCTL_PROC(_net_local_dgram, OID_AUTO, pcblist, CTLFLAG_RD, 1470 &unp_dgram_head, 0, unp_pcblist, "S,xunpcb", 1471 "List of active local datagram sockets"); 1472 SYSCTL_PROC(_net_local_stream, OID_AUTO, pcblist, CTLFLAG_RD, 1473 &unp_stream_head, 0, unp_pcblist, "S,xunpcb", 1474 "List of active local stream sockets"); 1475 SYSCTL_PROC(_net_local_seqpacket, OID_AUTO, pcblist, CTLFLAG_RD, 1476 &unp_seqpkt_head, 0, unp_pcblist, "S,xunpcb", 1477 "List of active local seqpacket sockets"); 1478 1479 static void 1480 unp_shutdown(struct unpcb *unp) 1481 { 1482 struct socket *so; 1483 1484 if ((unp->unp_socket->so_type == SOCK_STREAM || 1485 unp->unp_socket->so_type == SOCK_SEQPACKET) && 1486 unp->unp_conn != NULL && (so = unp->unp_conn->unp_socket)) { 1487 socantrcvmore(so); 1488 } 1489 } 1490 1491 #ifdef notdef 1492 void 1493 unp_drain(void) 1494 { 1495 lwkt_gettoken(&unp_token); 1496 lwkt_reltoken(&unp_token); 1497 } 1498 #endif 1499 1500 int 1501 unp_externalize(struct mbuf *rights, int flags) 1502 { 1503 struct thread *td = curthread; 1504 struct proc *p = td->td_proc; /* XXX */ 1505 struct lwp *lp = td->td_lwp; 1506 struct cmsghdr *cm = mtod(rights, struct cmsghdr *); 1507 int *fdp; 1508 int i; 1509 struct file **rp; 1510 struct file *fp; 1511 int newfds = (cm->cmsg_len - (CMSG_DATA(cm) - (u_char *)cm)) 1512 / sizeof(struct file *); 1513 int f; 1514 1515 lwkt_gettoken(&unp_rights_token); 1516 1517 /* 1518 * if the new FD's will not fit, then we free them all 1519 */ 1520 if (!fdavail(p, newfds)) { 1521 rp = (struct file **)CMSG_DATA(cm); 1522 for (i = 0; i < newfds; i++) { 1523 fp = *rp; 1524 /* 1525 * zero the pointer before calling unp_discard, 1526 * since it may end up in unp_gc().. 1527 */ 1528 *rp++ = NULL; 1529 unp_discard(fp, NULL); 1530 } 1531 lwkt_reltoken(&unp_rights_token); 1532 return (EMSGSIZE); 1533 } 1534 1535 /* 1536 * now change each pointer to an fd in the global table to 1537 * an integer that is the index to the local fd table entry 1538 * that we set up to point to the global one we are transferring. 1539 * Since the sizeof(struct file *) is bigger than or equal to 1540 * the sizeof(int), we do it in forward order. In that case, 1541 * an integer will always come in the same place or before its 1542 * corresponding struct file pointer. 1543 * 1544 * Hold revoke_token in 'shared' mode, so that we won't miss 1545 * the FREVOKED update on fps being externalized (fsetfd). 1546 */ 1547 lwkt_gettoken_shared(&revoke_token); 1548 fdp = (int *)CMSG_DATA(cm); 1549 rp = (struct file **)CMSG_DATA(cm); 1550 for (i = 0; i < newfds; i++) { 1551 if (fdalloc(p, 0, &f)) { 1552 int j; 1553 1554 /* 1555 * Previous fdavail() can't garantee 1556 * fdalloc() success due to SMP race. 1557 * Just clean up and return the same 1558 * error value as if fdavail() failed. 1559 */ 1560 lwkt_reltoken(&revoke_token); 1561 1562 /* Close externalized files */ 1563 for (j = 0; j < i; j++) 1564 kern_close(fdp[j]); 1565 /* Discard the rest of internal files */ 1566 for (; i < newfds; i++) 1567 unp_discard(rp[i], NULL); 1568 /* Wipe out the control message */ 1569 for (i = 0; i < newfds; i++) 1570 rp[i] = NULL; 1571 1572 lwkt_reltoken(&unp_rights_token); 1573 return (EMSGSIZE); 1574 } 1575 fp = rp[i]; 1576 unp_fp_externalize(lp, fp, f, flags); 1577 fdp[i] = f; 1578 } 1579 lwkt_reltoken(&revoke_token); 1580 1581 lwkt_reltoken(&unp_rights_token); 1582 1583 /* 1584 * Adjust length, in case sizeof(struct file *) and sizeof(int) 1585 * differs. 1586 */ 1587 cm->cmsg_len = CMSG_LEN(newfds * sizeof(int)); 1588 rights->m_len = cm->cmsg_len; 1589 1590 return (0); 1591 } 1592 1593 static void 1594 unp_fp_externalize(struct lwp *lp, struct file *fp, int fd, int flags) 1595 { 1596 if (lp) { 1597 struct filedesc *fdp = lp->lwp_proc->p_fd; 1598 1599 KKASSERT(fd >= 0); 1600 if (fp->f_flag & FREVOKED) { 1601 struct file *fx; 1602 int error; 1603 1604 kprintf("Warning: revoked fp exiting unix socket\n"); 1605 error = falloc(lp, &fx, NULL); 1606 if (error == 0) { 1607 if (flags & MSG_CMSG_CLOEXEC) 1608 fdp->fd_files[fd].fileflags |= UF_EXCLOSE; 1609 fsetfd(fdp, fx, fd); 1610 fdrop(fx); 1611 } else { 1612 fsetfd(fdp, NULL, fd); 1613 } 1614 } else { 1615 if (flags & MSG_CMSG_CLOEXEC) 1616 fdp->fd_files[fd].fileflags |= UF_EXCLOSE; 1617 fsetfd(fdp, fp, fd); 1618 } 1619 } 1620 unp_del_right(fp); 1621 fdrop(fp); 1622 } 1623 1624 void 1625 unp_init(void) 1626 { 1627 TAILQ_INIT(&unp_stream_head.list); 1628 TAILQ_INIT(&unp_dgram_head.list); 1629 TAILQ_INIT(&unp_seqpkt_head.list); 1630 1631 SLIST_INIT(&unp_defdiscard_head); 1632 spin_init(&unp_defdiscard_spin, "unpdisc"); 1633 TASK_INIT(&unp_defdiscard_task, 0, unp_defdiscard_taskfunc, NULL); 1634 1635 /* 1636 * This implies that only one gc can be in-progress at any 1637 * given moment. 1638 */ 1639 TASK_INIT(&unp_gc_task, 0, unp_gc, NULL); 1640 1641 unp_gc_marker = kmalloc(sizeof(*unp_gc_marker), M_UNPCB, 1642 M_WAITOK | M_ZERO); 1643 unp_gc_marker->unp_flags |= UNP_MARKER; 1644 1645 /* 1646 * Create taskqueue for defered discard, and stick it to 1647 * the last CPU. 1648 */ 1649 unp_taskqueue = taskqueue_create("unp_taskq", M_WAITOK, 1650 taskqueue_thread_enqueue, &unp_taskqueue); 1651 taskqueue_start_threads(&unp_taskqueue, 1, TDPRI_KERN_DAEMON, 1652 ncpus - 1, "unp taskq"); 1653 } 1654 1655 static int 1656 unp_internalize(struct mbuf *control, struct thread *td) 1657 { 1658 struct proc *p = td->td_proc; 1659 struct filedesc *fdescp; 1660 struct cmsghdr *cm = mtod(control, struct cmsghdr *); 1661 struct file **rp; 1662 struct file *fp; 1663 int i, fd, *fdp; 1664 struct cmsgcred *cmcred; 1665 int oldfds; 1666 u_int newlen; 1667 int error; 1668 1669 KKASSERT(p); 1670 1671 if ((cm->cmsg_type != SCM_RIGHTS && cm->cmsg_type != SCM_CREDS) || 1672 cm->cmsg_level != SOL_SOCKET || 1673 CMSG_ALIGN(cm->cmsg_len) != control->m_len) 1674 return EINVAL; 1675 1676 /* 1677 * Fill in credential information. 1678 */ 1679 if (cm->cmsg_type == SCM_CREDS) { 1680 cmcred = (struct cmsgcred *)CMSG_DATA(cm); 1681 cmcred->cmcred_pid = p->p_pid; 1682 cmcred->cmcred_uid = p->p_ucred->cr_ruid; 1683 cmcred->cmcred_gid = p->p_ucred->cr_rgid; 1684 cmcred->cmcred_euid = p->p_ucred->cr_uid; 1685 cmcred->cmcred_ngroups = MIN(p->p_ucred->cr_ngroups, 1686 CMGROUP_MAX); 1687 for (i = 0; i < cmcred->cmcred_ngroups; i++) 1688 cmcred->cmcred_groups[i] = p->p_ucred->cr_groups[i]; 1689 return 0; 1690 } 1691 1692 /* 1693 * cmsghdr may not be aligned, do not allow calculation(s) to 1694 * go negative. 1695 */ 1696 if (cm->cmsg_len < CMSG_LEN(0)) 1697 return EINVAL; 1698 1699 oldfds = (cm->cmsg_len - CMSG_LEN(0)) / sizeof(int); 1700 1701 /* 1702 * Now replace the integer FDs with pointers to 1703 * the associated global file table entry.. 1704 * Allocate a bigger buffer as necessary. But if an cluster is not 1705 * enough, return E2BIG. 1706 */ 1707 newlen = CMSG_LEN(oldfds * sizeof(struct file *)); 1708 if (newlen > MCLBYTES) 1709 return E2BIG; 1710 if (newlen - control->m_len > M_TRAILINGSPACE(control)) { 1711 if (control->m_flags & M_EXT) 1712 return E2BIG; 1713 MCLGET(control, M_WAITOK); 1714 if (!(control->m_flags & M_EXT)) 1715 return ENOBUFS; 1716 1717 /* copy the data to the cluster */ 1718 memcpy(mtod(control, char *), cm, cm->cmsg_len); 1719 cm = mtod(control, struct cmsghdr *); 1720 } 1721 1722 lwkt_gettoken(&unp_rights_token); 1723 1724 fdescp = p->p_fd; 1725 spin_lock_shared(&fdescp->fd_spin); 1726 1727 /* 1728 * check that all the FDs passed in refer to legal OPEN files 1729 * If not, reject the entire operation. 1730 */ 1731 fdp = (int *)CMSG_DATA(cm); 1732 for (i = 0; i < oldfds; i++) { 1733 fd = *fdp++; 1734 if ((unsigned)fd >= fdescp->fd_nfiles || 1735 fdescp->fd_files[fd].fp == NULL) { 1736 error = EBADF; 1737 goto done; 1738 } 1739 if (fdescp->fd_files[fd].fp->f_type == DTYPE_KQUEUE) { 1740 error = EOPNOTSUPP; 1741 goto done; 1742 } 1743 } 1744 1745 /* 1746 * Adjust length, in case sizeof(struct file *) and sizeof(int) 1747 * differs. 1748 */ 1749 cm->cmsg_len = newlen; 1750 control->m_len = CMSG_ALIGN(newlen); 1751 1752 /* 1753 * Transform the file descriptors into struct file pointers. 1754 * Since the sizeof(struct file *) is bigger than or equal to 1755 * the sizeof(int), we do it in reverse order so that the int 1756 * won't get trashed until we're done. 1757 */ 1758 fdp = (int *)CMSG_DATA(cm) + oldfds - 1; 1759 rp = (struct file **)CMSG_DATA(cm) + oldfds - 1; 1760 for (i = 0; i < oldfds; i++) { 1761 fp = fdescp->fd_files[*fdp--].fp; 1762 *rp-- = fp; 1763 fhold(fp); 1764 unp_add_right(fp); 1765 } 1766 error = 0; 1767 done: 1768 spin_unlock_shared(&fdescp->fd_spin); 1769 lwkt_reltoken(&unp_rights_token); 1770 return error; 1771 } 1772 1773 #ifdef UNP_GC_ALLFILES 1774 1775 /* 1776 * Garbage collect in-transit file descriptors that get lost due to 1777 * loops (i.e. when a socket is sent to another process over itself, 1778 * and more complex situations). 1779 * 1780 * NOT MPSAFE - TODO socket flush code and maybe fdrop. Rest is MPSAFE. 1781 */ 1782 1783 struct unp_gc_info { 1784 struct file **extra_ref; 1785 struct file *locked_fp; 1786 int defer; 1787 int index; 1788 int maxindex; 1789 }; 1790 1791 static void 1792 unp_gc(void *arg __unused, int pending __unused) 1793 { 1794 struct unp_gc_info info; 1795 struct file **fpp; 1796 int i; 1797 1798 lwkt_gettoken(&unp_rights_token); 1799 1800 /* 1801 * Before going through all this, set all FDs to be NOT defered 1802 * and NOT externally accessible (not marked). During the scan 1803 * a fd can be marked externally accessible but we may or may not 1804 * be able to immediately process it (controlled by FDEFER). 1805 * 1806 * If we loop sleep a bit. The complexity of the topology can cause 1807 * multiple loops. Also failure to acquire the socket's so_rcv 1808 * token can cause us to loop. 1809 */ 1810 allfiles_scan_exclusive(unp_gc_clearmarks, NULL); 1811 do { 1812 info.defer = 0; 1813 allfiles_scan_exclusive(unp_gc_checkmarks, &info); 1814 if (info.defer) 1815 tsleep(&info, 0, "gcagain", 1); 1816 } while (info.defer); 1817 1818 /* 1819 * We grab an extra reference to each of the file table entries 1820 * that are not otherwise accessible and then free the rights 1821 * that are stored in messages on them. 1822 * 1823 * The bug in the orginal code is a little tricky, so I'll describe 1824 * what's wrong with it here. 1825 * 1826 * It is incorrect to simply unp_discard each entry for f_msgcount 1827 * times -- consider the case of sockets A and B that contain 1828 * references to each other. On a last close of some other socket, 1829 * we trigger a gc since the number of outstanding rights (unp_rights) 1830 * is non-zero. If during the sweep phase the gc code unp_discards, 1831 * we end up doing a (full) fdrop on the descriptor. A fdrop on A 1832 * results in the following chain. Closef calls soo_close, which 1833 * calls soclose. Soclose calls first (through the switch 1834 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply 1835 * returns because the previous instance had set unp_gcing, and 1836 * we return all the way back to soclose, which marks the socket 1837 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush 1838 * to free up the rights that are queued in messages on the socket A, 1839 * i.e., the reference on B. The sorflush calls via the dom_dispose 1840 * switch unp_dispose, which unp_scans with unp_discard. This second 1841 * instance of unp_discard just calls fdrop on B. 1842 * 1843 * Well, a similar chain occurs on B, resulting in a sorflush on B, 1844 * which results in another fdrop on A. Unfortunately, A is already 1845 * being closed, and the descriptor has already been marked with 1846 * SS_NOFDREF, and soclose panics at this point. 1847 * 1848 * Here, we first take an extra reference to each inaccessible 1849 * descriptor. Then, we call sorflush ourself, since we know 1850 * it is a Unix domain socket anyhow. After we destroy all the 1851 * rights carried in messages, we do a last fdrop to get rid 1852 * of our extra reference. This is the last close, and the 1853 * unp_detach etc will shut down the socket. 1854 * 1855 * 91/09/19, bsy@cs.cmu.edu 1856 */ 1857 info.extra_ref = kmalloc(256 * sizeof(struct file *), M_FILE, M_WAITOK); 1858 info.maxindex = 256; 1859 1860 do { 1861 /* 1862 * Look for matches 1863 */ 1864 info.index = 0; 1865 allfiles_scan_exclusive(unp_gc_checkrefs, &info); 1866 1867 /* 1868 * For each FD on our hit list, do the following two things 1869 */ 1870 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp) { 1871 struct file *tfp = *fpp; 1872 if (tfp->f_type == DTYPE_SOCKET && tfp->f_data != NULL) 1873 sorflush((struct socket *)(tfp->f_data)); 1874 } 1875 for (i = info.index, fpp = info.extra_ref; --i >= 0; ++fpp) 1876 fdrop(*fpp); 1877 } while (info.index == info.maxindex); 1878 1879 kfree((caddr_t)info.extra_ref, M_FILE); 1880 1881 lwkt_reltoken(&unp_rights_token); 1882 } 1883 1884 /* 1885 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1886 */ 1887 static int 1888 unp_gc_checkrefs(struct file *fp, void *data) 1889 { 1890 struct unp_gc_info *info = data; 1891 1892 if (fp->f_count == 0) 1893 return(0); 1894 if (info->index == info->maxindex) 1895 return(-1); 1896 1897 /* 1898 * If all refs are from msgs, and it's not marked accessible 1899 * then it must be referenced from some unreachable cycle 1900 * of (shut-down) FDs, so include it in our 1901 * list of FDs to remove 1902 */ 1903 if (fp->f_count == fp->f_msgcount && !(fp->f_flag & FMARK)) { 1904 info->extra_ref[info->index++] = fp; 1905 fhold(fp); 1906 } 1907 return(0); 1908 } 1909 1910 /* 1911 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1912 */ 1913 static int 1914 unp_gc_clearmarks(struct file *fp, void *data __unused) 1915 { 1916 atomic_clear_int(&fp->f_flag, FMARK | FDEFER); 1917 return(0); 1918 } 1919 1920 /* 1921 * MPSAFE - NOTE: filehead list and file pointer spinlocked on entry 1922 */ 1923 static int 1924 unp_gc_checkmarks(struct file *fp, void *data) 1925 { 1926 struct unp_gc_info *info = data; 1927 struct socket *so; 1928 1929 /* 1930 * If the file is not open, skip it. Make sure it isn't marked 1931 * defered or we could loop forever, in case we somehow race 1932 * something. 1933 */ 1934 if (fp->f_count == 0) { 1935 if (fp->f_flag & FDEFER) 1936 atomic_clear_int(&fp->f_flag, FDEFER); 1937 return(0); 1938 } 1939 /* 1940 * If we already marked it as 'defer' in a 1941 * previous pass, then try process it this time 1942 * and un-mark it 1943 */ 1944 if (fp->f_flag & FDEFER) { 1945 atomic_clear_int(&fp->f_flag, FDEFER); 1946 } else { 1947 /* 1948 * if it's not defered, then check if it's 1949 * already marked.. if so skip it 1950 */ 1951 if (fp->f_flag & FMARK) 1952 return(0); 1953 /* 1954 * If all references are from messages 1955 * in transit, then skip it. it's not 1956 * externally accessible. 1957 */ 1958 if (fp->f_count == fp->f_msgcount) 1959 return(0); 1960 /* 1961 * If it got this far then it must be 1962 * externally accessible. 1963 */ 1964 atomic_set_int(&fp->f_flag, FMARK); 1965 } 1966 1967 /* 1968 * either it was defered, or it is externally 1969 * accessible and not already marked so. 1970 * Now check if it is possibly one of OUR sockets. 1971 */ 1972 if (fp->f_type != DTYPE_SOCKET || 1973 (so = (struct socket *)fp->f_data) == NULL) { 1974 return(0); 1975 } 1976 if (so->so_proto->pr_domain != &localdomain || 1977 !(so->so_proto->pr_flags & PR_RIGHTS)) { 1978 return(0); 1979 } 1980 1981 /* 1982 * So, Ok, it's one of our sockets and it IS externally accessible 1983 * (or was defered). Now we look to see if we hold any file 1984 * descriptors in its message buffers. Follow those links and mark 1985 * them as accessible too. 1986 * 1987 * We are holding multiple spinlocks here, if we cannot get the 1988 * token non-blocking defer until the next loop. 1989 */ 1990 info->locked_fp = fp; 1991 if (lwkt_trytoken(&so->so_rcv.ssb_token)) { 1992 unp_scan(so->so_rcv.ssb_mb, unp_mark, info); 1993 lwkt_reltoken(&so->so_rcv.ssb_token); 1994 } else { 1995 atomic_set_int(&fp->f_flag, FDEFER); 1996 ++info->defer; 1997 } 1998 return (0); 1999 } 2000 2001 /* 2002 * Mark visibility. info->defer is recalculated on every pass. 2003 */ 2004 static void 2005 unp_mark(struct file *fp, void *data) 2006 { 2007 struct unp_gc_info *info = data; 2008 2009 if ((fp->f_flag & FMARK) == 0) { 2010 ++info->defer; 2011 atomic_set_int(&fp->f_flag, FMARK | FDEFER); 2012 } else if (fp->f_flag & FDEFER) { 2013 ++info->defer; 2014 } 2015 } 2016 2017 #else /* !UNP_GC_ALLFILES */ 2018 2019 /* 2020 * They are thread local and do not require explicit synchronization. 2021 */ 2022 static int unp_marked; 2023 static int unp_unreachable; 2024 2025 static void 2026 unp_accessable(struct file *fp, void *data __unused) 2027 { 2028 struct unpcb *unp; 2029 2030 if ((unp = unp_fp2unpcb(fp)) == NULL) 2031 return; 2032 if (unp->unp_gcflags & UNPGC_REF) 2033 return; 2034 unp->unp_gcflags &= ~UNPGC_DEAD; 2035 unp->unp_gcflags |= UNPGC_REF; 2036 unp_marked++; 2037 } 2038 2039 static void 2040 unp_gc_process(struct unpcb *unp) 2041 { 2042 struct file *fp; 2043 2044 /* Already processed. */ 2045 if (unp->unp_gcflags & UNPGC_SCANNED) 2046 return; 2047 fp = unp->unp_fp; 2048 2049 /* 2050 * Check for a socket potentially in a cycle. It must be in a 2051 * queue as indicated by msgcount, and this must equal the file 2052 * reference count. Note that when msgcount is 0 the file is NULL. 2053 */ 2054 if ((unp->unp_gcflags & UNPGC_REF) == 0 && fp && 2055 unp->unp_msgcount != 0 && fp->f_count == unp->unp_msgcount) { 2056 unp->unp_gcflags |= UNPGC_DEAD; 2057 unp_unreachable++; 2058 return; 2059 } 2060 2061 /* 2062 * Mark all sockets we reference with RIGHTS. 2063 */ 2064 if (UNP_ISATTACHED(unp)) { 2065 struct signalsockbuf *ssb = &unp->unp_socket->so_rcv; 2066 2067 unp_reference(unp); 2068 lwkt_gettoken(&ssb->ssb_token); 2069 /* 2070 * unp_token would be temporarily dropped, if getting 2071 * so_rcv token blocks, so we need to check unp state 2072 * here again. 2073 */ 2074 if (UNP_ISATTACHED(unp)) 2075 unp_scan(ssb->ssb_mb, unp_accessable, NULL); 2076 lwkt_reltoken(&ssb->ssb_token); 2077 unp->unp_gcflags |= UNPGC_SCANNED; 2078 unp_free(unp); 2079 } else { 2080 unp->unp_gcflags |= UNPGC_SCANNED; 2081 } 2082 } 2083 2084 static void 2085 unp_gc(void *arg __unused, int pending __unused) 2086 { 2087 struct unp_global_head *head; 2088 int h, filemax, fileidx, filetot; 2089 struct file **unref; 2090 struct unpcb *unp; 2091 2092 lwkt_gettoken(&unp_rights_token); 2093 lwkt_gettoken(&unp_token); 2094 2095 /* 2096 * First clear all gc flags from previous runs. 2097 */ 2098 for (h = 0; unp_heads[h] != NULL; ++h) { 2099 /* 2100 * NOTE: This loop does not block, so it is safe 2101 * to use TAILQ_FOREACH here. 2102 */ 2103 head = unp_heads[h]; 2104 TAILQ_FOREACH(unp, &head->list, unp_link) 2105 unp->unp_gcflags = 0; 2106 } 2107 2108 /* 2109 * Scan marking all reachable sockets with UNPGC_REF. Once a socket 2110 * is reachable all of the sockets it references are reachable. 2111 * Stop the scan once we do a complete loop without discovering 2112 * a new reachable socket. 2113 */ 2114 do { 2115 unp_unreachable = 0; 2116 unp_marked = 0; 2117 for (h = 0; unp_heads[h] != NULL; ++h) { 2118 head = unp_heads[h]; 2119 TAILQ_INSERT_HEAD(&head->list, unp_gc_marker, unp_link); 2120 while ((unp = TAILQ_NEXT(unp_gc_marker, unp_link)) 2121 != NULL) { 2122 TAILQ_REMOVE(&head->list, unp_gc_marker, 2123 unp_link); 2124 TAILQ_INSERT_AFTER(&head->list, unp, 2125 unp_gc_marker, unp_link); 2126 2127 if (unp->unp_flags & UNP_MARKER) 2128 continue; 2129 unp_gc_process(unp); 2130 } 2131 TAILQ_REMOVE(&head->list, unp_gc_marker, unp_link); 2132 } 2133 } while (unp_marked); 2134 2135 if (unp_unreachable == 0) 2136 goto done; 2137 2138 /* 2139 * We grab an extra reference to each of the file table entries 2140 * that are not otherwise accessible and then free the rights 2141 * that are stored in messages on them. 2142 * 2143 * The bug in the orginal code is a little tricky, so I'll describe 2144 * what's wrong with it here. 2145 * 2146 * It is incorrect to simply unp_discard each entry for f_msgcount 2147 * times -- consider the case of sockets A and B that contain 2148 * references to each other. On a last close of some other socket, 2149 * we trigger a gc since the number of outstanding rights (unp_rights) 2150 * is non-zero. If during the sweep phase the gc code unp_discards, 2151 * we end up doing a (full) fdrop on the descriptor. A fdrop on A 2152 * results in the following chain. Closef calls soo_close, which 2153 * calls soclose. Soclose calls first (through the switch 2154 * uipc_usrreq) unp_detach, which re-invokes unp_gc. Unp_gc simply 2155 * returns because the previous instance had set unp_gcing, and 2156 * we return all the way back to soclose, which marks the socket 2157 * with SS_NOFDREF, and then calls sofree. Sofree calls sorflush 2158 * to free up the rights that are queued in messages on the socket A, 2159 * i.e., the reference on B. The sorflush calls via the dom_dispose 2160 * switch unp_dispose, which unp_scans with unp_discard. This second 2161 * instance of unp_discard just calls fdrop on B. 2162 * 2163 * Well, a similar chain occurs on B, resulting in a sorflush on B, 2164 * which results in another fdrop on A. Unfortunately, A is already 2165 * being closed, and the descriptor has already been marked with 2166 * SS_NOFDREF, and soclose panics at this point. 2167 * 2168 * Here, we first take an extra reference to each inaccessible 2169 * descriptor. Then, we call sorflush ourself, since we know 2170 * it is a Unix domain socket anyhow. After we destroy all the 2171 * rights carried in messages, we do a last fdrop to get rid 2172 * of our extra reference. This is the last close, and the 2173 * unp_detach etc will shut down the socket. 2174 * 2175 * 91/09/19, bsy@cs.cmu.edu 2176 */ 2177 2178 filemax = unp_unreachable; 2179 if (filemax > UNP_GCFILE_MAX) 2180 filemax = UNP_GCFILE_MAX; 2181 unref = kmalloc(filemax * sizeof(struct file *), M_TEMP, M_WAITOK); 2182 2183 filetot = 0; 2184 do { 2185 int i; 2186 2187 /* 2188 * Iterate looking for sockets which have been specifically 2189 * marked as as unreachable and store them locally. 2190 */ 2191 fileidx = 0; 2192 for (h = 0; unp_heads[h] != NULL; ++h) { 2193 /* 2194 * NOTE: This loop does not block, so it is safe 2195 * to use TAILQ_FOREACH here. 2196 */ 2197 head = unp_heads[h]; 2198 TAILQ_FOREACH(unp, &head->list, unp_link) { 2199 struct file *fp; 2200 2201 if ((unp->unp_gcflags & UNPGC_DEAD) == 0) 2202 continue; 2203 unp->unp_gcflags &= ~UNPGC_DEAD; 2204 2205 fp = unp->unp_fp; 2206 if (unp->unp_msgcount == 0 || fp == NULL || 2207 fp->f_count != unp->unp_msgcount) 2208 continue; 2209 fhold(fp); 2210 2211 KASSERT(fileidx < filemax, 2212 ("invalid fileidx %d, filemax %d", 2213 fileidx, filemax)); 2214 unref[fileidx++] = fp; 2215 2216 KASSERT(filetot < unp_unreachable, 2217 ("invalid filetot %d and " 2218 "unp_unreachable %d", 2219 filetot, unp_unreachable)); 2220 ++filetot; 2221 2222 if (fileidx == filemax || 2223 filetot == unp_unreachable) 2224 goto dogc; 2225 } 2226 } 2227 dogc: 2228 /* 2229 * For each Unix domain socket on our hit list, do the 2230 * following two things. 2231 */ 2232 for (i = 0; i < fileidx; ++i) 2233 sorflush(unref[i]->f_data); 2234 for (i = 0; i < fileidx; ++i) 2235 fdrop(unref[i]); 2236 } while (fileidx == filemax && filetot < unp_unreachable); 2237 kfree(unref, M_TEMP); 2238 done: 2239 lwkt_reltoken(&unp_token); 2240 lwkt_reltoken(&unp_rights_token); 2241 } 2242 2243 #endif /* UNP_GC_ALLFILES */ 2244 2245 /* 2246 * Dispose of the fp's stored in a mbuf. 2247 * 2248 * The dds loop can cause additional fps to be entered onto the 2249 * list while it is running, flattening out the operation and avoiding 2250 * a deep kernel stack recursion. 2251 */ 2252 void 2253 unp_dispose(struct mbuf *m) 2254 { 2255 lwkt_gettoken(&unp_rights_token); 2256 if (m) 2257 unp_scan(m, unp_discard, NULL); 2258 lwkt_reltoken(&unp_rights_token); 2259 } 2260 2261 static int 2262 unp_listen(struct unpcb *unp, struct thread *td) 2263 { 2264 struct proc *p = td->td_proc; 2265 2266 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2267 UNP_ASSERT_TOKEN_HELD(unp); 2268 2269 KKASSERT(p); 2270 cru2x(p->p_ucred, &unp->unp_peercred); 2271 unp_setflags(unp, UNP_HAVEPCCACHED); 2272 return (0); 2273 } 2274 2275 static void 2276 unp_scan(struct mbuf *m0, void (*op)(struct file *, void *), void *data) 2277 { 2278 struct mbuf *m; 2279 struct file **rp; 2280 struct cmsghdr *cm; 2281 int i; 2282 int qfds; 2283 2284 while (m0) { 2285 for (m = m0; m; m = m->m_next) { 2286 if (m->m_type == MT_CONTROL && 2287 m->m_len >= sizeof(*cm)) { 2288 cm = mtod(m, struct cmsghdr *); 2289 if (cm->cmsg_level != SOL_SOCKET || 2290 cm->cmsg_type != SCM_RIGHTS) 2291 continue; 2292 qfds = (cm->cmsg_len - CMSG_LEN(0)) / 2293 sizeof(void *); 2294 rp = (struct file **)CMSG_DATA(cm); 2295 for (i = 0; i < qfds; i++) 2296 (*op)(*rp++, data); 2297 break; /* XXX, but saves time */ 2298 } 2299 } 2300 m0 = m0->m_nextpkt; 2301 } 2302 } 2303 2304 /* 2305 * Discard a fp previously held in a unix domain socket mbuf. To 2306 * avoid blowing out the kernel stack due to contrived chain-reactions 2307 * we may have to defer the operation to a dedicated taskqueue. 2308 * 2309 * Caller holds unp_rights_token. 2310 */ 2311 static void 2312 unp_discard(struct file *fp, void *data __unused) 2313 { 2314 unp_del_right(fp); 2315 if (unp_fp2unpcb(fp) != NULL) { 2316 struct unp_defdiscard *d; 2317 2318 /* 2319 * This fp is a Unix domain socket itself and fdrop() 2320 * it here directly may cause deep unp_discard() 2321 * recursion, so the fdrop() is defered to the 2322 * dedicated taskqueue. 2323 */ 2324 d = kmalloc(sizeof(*d), M_UNPCB, M_WAITOK); 2325 d->fp = fp; 2326 2327 spin_lock(&unp_defdiscard_spin); 2328 SLIST_INSERT_HEAD(&unp_defdiscard_head, d, next); 2329 spin_unlock(&unp_defdiscard_spin); 2330 2331 taskqueue_enqueue(unp_taskqueue, &unp_defdiscard_task); 2332 } else { 2333 /* This fp is not a Unix domain socket */ 2334 fdrop(fp); 2335 } 2336 } 2337 2338 /* 2339 * NOTE: 2340 * unp_token must be held before calling this function to avoid name 2341 * resolution and v_socket accessing races, especially racing against 2342 * the unp_detach(). 2343 * 2344 * NOTE: 2345 * For anyone caring about unconnected Unix domain socket sending 2346 * performance, other approach could be taken... 2347 */ 2348 static int 2349 unp_find_lockref(struct sockaddr *nam, struct thread *td, short type, 2350 struct unpcb **unp_ret) 2351 { 2352 struct proc *p = td->td_proc; 2353 struct sockaddr_un *soun = (struct sockaddr_un *)nam; 2354 struct vnode *vp = NULL; 2355 struct socket *so; 2356 struct unpcb *unp; 2357 int error, len; 2358 struct nlookupdata nd; 2359 char buf[SOCK_MAXADDRLEN]; 2360 2361 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2362 2363 *unp_ret = NULL; 2364 2365 len = nam->sa_len - offsetof(struct sockaddr_un, sun_path); 2366 if (len <= 0) { 2367 error = EINVAL; 2368 goto failed; 2369 } 2370 strncpy(buf, soun->sun_path, len); 2371 buf[len] = 0; 2372 2373 error = nlookup_init(&nd, buf, UIO_SYSSPACE, NLC_FOLLOW); 2374 if (error == 0) 2375 error = nlookup(&nd); 2376 if (error == 0) 2377 error = cache_vget(&nd.nl_nch, nd.nl_cred, LK_EXCLUSIVE, &vp); 2378 nlookup_done(&nd); 2379 if (error) { 2380 vp = NULL; 2381 goto failed; 2382 } 2383 2384 if (vp->v_type != VSOCK) { 2385 error = ENOTSOCK; 2386 goto failed; 2387 } 2388 error = VOP_EACCESS(vp, VWRITE, p->p_ucred); 2389 if (error) 2390 goto failed; 2391 so = vp->v_socket; 2392 if (so == NULL) { 2393 error = ECONNREFUSED; 2394 goto failed; 2395 } 2396 if (so->so_type != type) { 2397 error = EPROTOTYPE; 2398 goto failed; 2399 } 2400 2401 /* Lock this unp. */ 2402 unp = unp_getsocktoken(so); 2403 if (!UNP_ISATTACHED(unp)) { 2404 unp_reltoken(unp); 2405 error = ECONNREFUSED; 2406 goto failed; 2407 } 2408 /* And keep this unp referenced. */ 2409 unp_reference(unp); 2410 2411 /* Done! */ 2412 *unp_ret = unp; 2413 error = 0; 2414 failed: 2415 if (vp != NULL) 2416 vput(vp); 2417 return error; 2418 } 2419 2420 static int 2421 unp_connect_pair(struct unpcb *unp, struct unpcb *unp2) 2422 { 2423 struct socket *so = unp->unp_socket; 2424 struct socket *so2 = unp2->unp_socket; 2425 2426 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2427 UNP_ASSERT_TOKEN_HELD(unp); 2428 UNP_ASSERT_TOKEN_HELD(unp2); 2429 2430 KASSERT(so->so_type == so2->so_type, 2431 ("socket type mismatch, so %d, so2 %d", so->so_type, so2->so_type)); 2432 2433 if (!UNP_ISATTACHED(unp)) 2434 return EINVAL; 2435 if (!UNP_ISATTACHED(unp2)) 2436 return ECONNREFUSED; 2437 2438 KASSERT(unp->unp_conn == NULL, ("unp is already connected")); 2439 unp->unp_conn = unp2; 2440 2441 switch (so->so_type) { 2442 case SOCK_DGRAM: 2443 LIST_INSERT_HEAD(&unp2->unp_refs, unp, unp_reflink); 2444 soisconnected(so); 2445 break; 2446 2447 case SOCK_STREAM: 2448 case SOCK_SEQPACKET: 2449 KASSERT(unp2->unp_conn == NULL, ("unp2 is already connected")); 2450 unp2->unp_conn = unp; 2451 soisconnected(so); 2452 soisconnected(so2); 2453 break; 2454 2455 default: 2456 panic("unp_connect_pair: unknown socket type %d", so->so_type); 2457 } 2458 return 0; 2459 } 2460 2461 static void 2462 unp_drop(struct unpcb *unp, int error) 2463 { 2464 struct unp_global_head *head; 2465 struct unpcb *unp2; 2466 2467 ASSERT_LWKT_TOKEN_HELD(&unp_token); 2468 UNP_ASSERT_TOKEN_HELD(unp); 2469 2470 KASSERT((unp->unp_flags & (UNP_DETACHED | UNP_DROPPED)) == 0, 2471 ("unp is dropped")); 2472 2473 /* Mark this unp as detached. */ 2474 unp_setflags(unp, UNP_DETACHED); 2475 2476 /* Remove this unp from the global unp list. */ 2477 head = unp_globalhead(unp->unp_socket->so_type); 2478 KASSERT(head->count > 0, ("invalid unp count")); 2479 TAILQ_REMOVE(&head->list, unp, unp_link); 2480 head->count--; 2481 2482 /* Disconnect all. */ 2483 unp_disconnect(unp, error); 2484 while ((unp2 = LIST_FIRST(&unp->unp_refs)) != NULL) { 2485 lwkt_getpooltoken(unp2); 2486 unp_disconnect(unp2, ECONNRESET); 2487 lwkt_relpooltoken(unp2); 2488 } 2489 unp_setflags(unp, UNP_DROPPED); 2490 2491 /* Try freeing this unp. */ 2492 unp_free(unp); 2493 } 2494 2495 static void 2496 unp_defdiscard_taskfunc(void *arg __unused, int pending __unused) 2497 { 2498 struct unp_defdiscard *d; 2499 2500 spin_lock(&unp_defdiscard_spin); 2501 while ((d = SLIST_FIRST(&unp_defdiscard_head)) != NULL) { 2502 SLIST_REMOVE_HEAD(&unp_defdiscard_head, next); 2503 spin_unlock(&unp_defdiscard_spin); 2504 2505 fdrop(d->fp); 2506 kfree(d, M_UNPCB); 2507 2508 spin_lock(&unp_defdiscard_spin); 2509 } 2510 spin_unlock(&unp_defdiscard_spin); 2511 } 2512