sys/emulation/linux/linux_socket.c

   1 /*-
   2  * Copyright (c) 1995 Søren Schmidt
   3  * 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  *    in this position and unchanged.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  * 3. The name of the author may not be used to endorse or promote products
  15  *    derived from this software withough specific prior written permission
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  27  *
  28  * $FreeBSD: src/sys/compat/linux/linux_socket.c,v 1.19.2.8 2001/11/07 20:33:55 marcel Exp $
  29  * $DragonFly: src/sys/emulation/linux/linux_socket.c,v 1.28 2008/07/10 00:19:27 aggelos Exp $
  30  */
  31
  32 #include <sys/param.h>
  33 #include <sys/proc.h>
  34 #include <sys/systm.h>
  35 #include <sys/sysproto.h>
  36 #include <sys/fcntl.h>
  37 #include <sys/file.h>
  38 #include <sys/kern_syscall.h>
  39 #include <sys/socket.h>
  40 #include <sys/socketvar.h>
  41 #include <sys/uio.h>
  42 #include <sys/malloc.h>
  43 #include <sys/mbuf.h>
  44
  45 #include <sys/mplock2.h>
  46
  47 #include <netinet/in.h>
  48 #include <netinet/in_systm.h>
  49 #include <netinet/ip.h>
  50
  51 #include <arch_linux/linux.h>
  52 #include <arch_linux/linux_proto.h>
  53 #include "linux_socket.h"
  54 #include "linux_util.h"
  55
  56 /*
  57  * Copyin a sockaddr structure provided by a Linux binary.  Linux uses
  58  * the 4.3BSD sockaddr structure which has no sa_len field.  We must
  59  * pass 4.4BSD sockaddr structures when we call native functions in the
  60  * BSD kernel.  This function does the conversion for us.
  61  *
  62  * Also, our socket calls require the sockaddr structure length to agree
  63  * with the address family.  Linux does not, so we must force it.
  64  *
  65  * This function should only need to be called from linux_connect()
  66  * and linux_bind().
  67  */
  68 static int
  69 linux_getsockaddr(struct sockaddr **namp, struct sockaddr *uaddr, size_t len)
  70 {
  71         struct sockaddr *sa;
  72         uint16_t family;        /* XXX: must match Linux sockaddr */
  73         int error;
  74         int sa_len;
  75
  76         *namp = NULL;
  77
  78         if (len > SOCK_MAXADDRLEN)
  79                 return ENAMETOOLONG;
  80         error = copyin(uaddr, &family, sizeof(family));
  81         if (error)
  82                 return (error);
  83
  84         /*
  85          * Force the sa_len field to match the address family.
  86          */
  87         switch (family) {
  88         case AF_INET:
  89                 sa_len = sizeof(struct sockaddr_in);
  90                 break;
  91         case AF_INET6:
  92                 sa_len = sizeof(struct sockaddr_in6);
  93                 break;
  94         default:
  95                 /*
  96                  * This is the default behavior of the old
  97                  * linux_to_bsd_namelen() function.  NOTE!  The
  98                  * minimum length we allocate must cover sa->sa_len and
  99                  * sa->sa_family.
 100                  */
 101                 sa_len = offsetof(struct sockaddr, sa_data[0]);
 102                 if (sa_len < len)
 103                         sa_len = len;
 104                 break;
 105         }
 106
 107         MALLOC(sa, struct sockaddr *, sa_len, M_SONAME, M_WAITOK);
 108         error = copyin(uaddr, sa, sa_len);
 109         if (error) {
 110                 FREE(sa, M_SONAME);
 111         } else {
 112                 /*
 113                  * Convert to the 4.4BSD sockaddr structure.
 114                  */
 115                 sa->sa_family = *(sa_family_t *)sa;
 116                 sa->sa_len = sa_len;
 117                 *namp = sa;
 118         }
 119
 120         return (error);
 121 }
 122
 123 /*
 124  * Transform a 4.4BSD sockaddr structure into a Linux sockaddr structure
 125  * and copy it out to a user address.
 126  */
 127 static int
 128 linux_copyout_sockaddr(struct sockaddr *sa, struct sockaddr *uaddr, int sa_len)
 129 {
 130         int error;
 131
 132         if (sa_len < (int)sizeof(u_short))
 133                 return (EINVAL);
 134
 135         *(u_short *)sa = sa->sa_family;
 136         error = copyout(sa, uaddr, sa_len);
 137
 138         return (error);
 139 }
 140
 141 static int
 142 linux_to_bsd_domain(int domain)
 143 {
 144
 145         switch (domain) {
 146         case LINUX_AF_UNSPEC:
 147                 return (AF_UNSPEC);
 148         case LINUX_AF_UNIX:
 149                 return (AF_LOCAL);
 150         case LINUX_AF_INET:
 151                 return (AF_INET);
 152         case LINUX_AF_AX25:
 153                 return (AF_CCITT);
 154         case LINUX_AF_IPX:
 155                 return (AF_IPX);
 156         case LINUX_AF_APPLETALK:
 157                 return (AF_APPLETALK);
 158         }
 159         return (-1);
 160 }
 161
 162 static int
 163 linux_to_bsd_sockopt_level(int level)
 164 {
 165
 166         switch (level) {
 167         case LINUX_SOL_SOCKET:
 168                 return (SOL_SOCKET);
 169         }
 170         return (level);
 171 }
 172
 173 static int
 174 linux_to_bsd_ip_sockopt(int opt)
 175 {
 176
 177         switch (opt) {
 178         case LINUX_IP_TOS:
 179                 return (IP_TOS);
 180         case LINUX_IP_TTL:
 181                 return (IP_TTL);
 182         case LINUX_IP_OPTIONS:
 183                 return (IP_OPTIONS);
 184         case LINUX_IP_MULTICAST_IF:
 185                 return (IP_MULTICAST_IF);
 186         case LINUX_IP_MULTICAST_TTL:
 187                 return (IP_MULTICAST_TTL);
 188         case LINUX_IP_MULTICAST_LOOP:
 189                 return (IP_MULTICAST_LOOP);
 190         case LINUX_IP_ADD_MEMBERSHIP:
 191                 return (IP_ADD_MEMBERSHIP);
 192         case LINUX_IP_DROP_MEMBERSHIP:
 193                 return (IP_DROP_MEMBERSHIP);
 194         case LINUX_IP_HDRINCL:
 195                 return (IP_HDRINCL);
 196         }
 197         return (-1);
 198 }
 199
 200 static int
 201 linux_to_bsd_so_sockopt(int opt)
 202 {
 203
 204         switch (opt) {
 205         case LINUX_SO_DEBUG:
 206                 return (SO_DEBUG);
 207         case LINUX_SO_REUSEADDR:
 208                 return (SO_REUSEADDR);
 209         case LINUX_SO_TYPE:
 210                 return (SO_TYPE);
 211         case LINUX_SO_ERROR:
 212                 return (SO_ERROR);
 213         case LINUX_SO_DONTROUTE:
 214                 return (SO_DONTROUTE);
 215         case LINUX_SO_BROADCAST:
 216                 return (SO_BROADCAST);
 217         case LINUX_SO_SNDBUF:
 218                 return (SO_SNDBUF);
 219         case LINUX_SO_RCVBUF:
 220                 return (SO_RCVBUF);
 221         case LINUX_SO_KEEPALIVE:
 222                 return (SO_KEEPALIVE);
 223         case LINUX_SO_OOBINLINE:
 224                 return (SO_OOBINLINE);
 225         case LINUX_SO_LINGER:
 226                 return (SO_LINGER);
 227         }
 228         return (-1);
 229 }
 230
 231 static int
 232 linux_to_bsd_msg_flags(int flags)
 233 {
 234         int ret_flags = 0;
 235
 236         if (flags & LINUX_MSG_OOB)
 237                 ret_flags |= MSG_OOB;
 238         if (flags & LINUX_MSG_PEEK)
 239                 ret_flags |= MSG_PEEK;
 240         if (flags & LINUX_MSG_DONTROUTE)
 241                 ret_flags |= MSG_DONTROUTE;
 242         if (flags & LINUX_MSG_CTRUNC)
 243                 ret_flags |= MSG_CTRUNC;
 244         if (flags & LINUX_MSG_TRUNC)
 245                 ret_flags |= MSG_TRUNC;
 246         if (flags & LINUX_MSG_DONTWAIT)
 247                 ret_flags |= MSG_DONTWAIT;
 248         if (flags & LINUX_MSG_EOR)
 249                 ret_flags |= MSG_EOR;
 250         if (flags & LINUX_MSG_WAITALL)
 251                 ret_flags |= MSG_WAITALL;
 252 #if 0 /* not handled */
 253         if (flags & LINUX_MSG_PROXY)
 254                 ;
 255         if (flags & LINUX_MSG_FIN)
 256                 ;
 257         if (flags & LINUX_MSG_SYN)
 258                 ;
 259         if (flags & LINUX_MSG_CONFIRM)
 260                 ;
 261         if (flags & LINUX_MSG_RST)
 262                 ;
 263         if (flags & LINUX_MSG_ERRQUEUE)
 264                 ;
 265         if (flags & LINUX_MSG_NOSIGNAL)
 266                 ;
 267 #endif
 268         return ret_flags;
 269 }
 270
 271 struct linux_socket_args {
 272         int domain;
 273         int type;
 274         int protocol;
 275 };
 276
 277 static int
 278 linux_socket(struct linux_socket_args *args, int *res)
 279 {
 280         struct linux_socket_args linux_args;
 281         struct sockopt sopt;
 282         int error, domain, optval;
 283
 284         error = copyin(args, &linux_args, sizeof(linux_args));
 285         if (error)
 286                 return (error);
 287
 288         domain = linux_to_bsd_domain(linux_args.domain);
 289         if (domain == -1)
 290                 return (EINVAL);
 291
 292         error = kern_socket(domain, linux_args.type, linux_args.protocol, res);
 293
 294         /* Copy back the return value from socket() */
 295         if (error == 0 && linux_args.type == SOCK_RAW &&
 296             (linux_args.protocol == IPPROTO_RAW || linux_args.protocol == 0) &&
 297             linux_args.domain == AF_INET) {
 298                 /* It's a raw IP socket: set the IP_HDRINCL option. */
 299                 optval = 1;
 300                 sopt.sopt_dir = SOPT_SET;
 301                 sopt.sopt_level = IPPROTO_IP;
 302                 sopt.sopt_name = IP_HDRINCL;
 303                 sopt.sopt_val = &optval;
 304                 sopt.sopt_valsize = sizeof(optval);
 305                 sopt.sopt_td = NULL;
 306
 307                 /* We ignore any error returned by setsockopt() */
 308                 kern_setsockopt(*res, &sopt);
 309         }
 310
 311         return (error);
 312 }
 313
 314 struct linux_bind_args {
 315         int s;
 316         struct sockaddr *name;
 317         int namelen;
 318 };
 319
 320 static int
 321 linux_bind(struct linux_bind_args *args, int *res)
 322 {
 323         struct linux_bind_args linux_args;
 324         struct sockaddr *sa;
 325         int error;
 326
 327         error = copyin(args, &linux_args, sizeof(linux_args));
 328         if (error)
 329                 return (error);
 330         error = linux_getsockaddr(&sa, linux_args.name, linux_args.namelen);
 331         if (error)
 332                 return (error);
 333
 334         error = kern_bind(linux_args.s, sa);
 335         FREE(sa, M_SONAME);
 336
 337         return (error);
 338 }
 339
 340 struct linux_connect_args {
 341         int s;
 342         struct sockaddr * name;
 343         int namelen;
 344 };
 345
 346 static int
 347 linux_connect(struct linux_connect_args *args, int *res)
 348 {
 349         struct thread *td = curthread;  /* XXX */
 350         struct proc *p = td->td_proc;
 351         struct linux_connect_args linux_args;
 352         struct sockaddr *sa;
 353         struct socket *so;
 354         struct file *fp;
 355         int error;
 356
 357         KKASSERT(p);
 358
 359         error = copyin(args, &linux_args, sizeof(linux_args));
 360         if (error)
 361                 return (error);
 362         error = linux_getsockaddr(&sa, linux_args.name, linux_args.namelen);
 363         if (error)
 364                 return (error);
 365
 366         error = kern_connect(linux_args.s, 0, sa);
 367         FREE(sa, M_SONAME);
 368
 369         if (error != EISCONN)
 370                 return (error);
 371
 372         /*
 373          * Linux doesn't return EISCONN the first time it occurs,
 374          * when on a non-blocking socket. Instead it returns the
 375          * error getsockopt(SOL_SOCKET, SO_ERROR) would return on BSD.
 376          */
 377         error = holdsock(p->p_fd, linux_args.s, &fp);
 378         if (error)
 379                 return (error);
 380         error = EISCONN;
 381         if (fp->f_flag & FNONBLOCK) {
 382                 so = (struct socket *)fp->f_data;
 383                 if (so->so_emuldata == 0)
 384                         error = so->so_error;
 385                 so->so_emuldata = (void *)1;
 386         }
 387         fdrop(fp);
 388         return (error);
 389 }
 390
 391 struct linux_listen_args {
 392         int s;
 393         int backlog;
 394 };
 395
 396 static int
 397 linux_listen(struct linux_listen_args *args, int *res)
 398 {
 399         struct linux_listen_args linux_args;
 400         int error;
 401
 402         error = copyin(args, &linux_args, sizeof(linux_args));
 403         if (error)
 404                 return (error);
 405
 406         error = kern_listen(linux_args.s, linux_args.backlog);
 407
 408         return(error);
 409 }
 410
 411 struct linux_accept_args {
 412         int s;
 413         struct sockaddr *addr;
 414         int *namelen;
 415 };
 416
 417 static int
 418 linux_accept(struct linux_accept_args *args, int *res)
 419 {
 420         struct thread *td = curthread;
 421         struct linux_accept_args linux_args;
 422         struct sockaddr *sa = NULL;
 423         union fcntl_dat dat = { 0 };
 424         int error, sa_len;
 425
 426         error = copyin(args, &linux_args, sizeof(linux_args));
 427         if (error)
 428                 return (error);
 429
 430         if (linux_args.addr) {
 431                 error = copyin(linux_args.namelen, &sa_len, sizeof(sa_len));
 432                 if (error)
 433                         return (error);
 434
 435                 error = kern_accept(linux_args.s, 0, &sa, &sa_len, res);
 436
 437                 if (error) {
 438                         /*
 439                          * Return a namelen of zero for older code which
 440                          * might ignore the return value from accept().
 441                          */
 442                         sa_len = 0;
 443                         copyout(&sa_len, linux_args.namelen,
 444                             sizeof(*linux_args.namelen));
 445                 } else {
 446                         error = linux_copyout_sockaddr(sa, linux_args.addr,
 447                             sa_len);
 448                         if (error == 0) {
 449                                 error = copyout(&sa_len, linux_args.namelen,
 450                                     sizeof(*linux_args.namelen));
 451                         }
 452                 }
 453                 if (sa)
 454                         FREE(sa, M_SONAME);
 455         } else {
 456                 error = kern_accept(linux_args.s, 0, NULL, 0, res);
 457         }
 458
 459         if (error)
 460                 return (error);
 461
 462         /*
 463          * linux appears not to copy flags from the parent socket to the
 464          * accepted one, so we must clear the flags in the new descriptor.
 465          * Ignore any errors, because we already have an open fd.
 466          */
 467         kern_fcntl(*res, F_SETFL, &dat, td->td_ucred);
 468         return (0);
 469 }
 470
 471 struct linux_getsockname_args {
 472         int s;
 473         struct sockaddr *addr;
 474         int *namelen;
 475 };
 476
 477 static int
 478 linux_getsockname(struct linux_getsockname_args *args, int *res)
 479 {
 480         struct linux_getsockname_args linux_args;
 481         struct sockaddr *sa = NULL;
 482         int error, sa_len;
 483
 484
 485         error = copyin(args, &linux_args, sizeof(linux_args));
 486         if (error)
 487                 return (error);
 488         error = copyin(linux_args.namelen, &sa_len, sizeof(sa_len));
 489         if (error)
 490                 return (error);
 491
 492         error = kern_getsockname(linux_args.s, &sa, &sa_len);
 493
 494         if (error == 0)
 495                 error = linux_copyout_sockaddr(sa, linux_args.addr, sa_len);
 496         if (error == 0)
 497                 error = copyout(&sa_len, linux_args.namelen,
 498                     sizeof(*linux_args.namelen));
 499         if (sa)
 500                 FREE(sa, M_SONAME);
 501         return(error);
 502 }
 503
 504 struct linux_getpeername_args {
 505         int s;
 506         struct sockaddr *addr;
 507         int *namelen;
 508 };
 509
 510 static int
 511 linux_getpeername(struct linux_getpeername_args *args, int *res)
 512 {
 513         struct linux_getpeername_args linux_args;
 514         struct sockaddr *sa = NULL;
 515         int error, sa_len;
 516
 517         error = copyin(args, &linux_args, sizeof(linux_args));
 518         if (error)
 519                 return (error);
 520         error = copyin(linux_args.namelen, &sa_len, sizeof(sa_len));
 521         if (error)
 522                 return (error);
 523
 524         error = kern_getpeername(linux_args.s, &sa, &sa_len);
 525
 526         if (error == 0)
 527                 error = linux_copyout_sockaddr(sa, linux_args.addr, sa_len);
 528         if (error == 0)
 529                 error = copyout(&sa_len, linux_args.namelen,
 530                     sizeof(*linux_args.namelen));
 531         if (sa)
 532                 FREE(sa, M_SONAME);
 533         return(error);
 534 }
 535
 536 struct linux_socketpair_args {
 537         int domain;
 538         int type;
 539         int protocol;
 540         int *rsv;
 541 };
 542
 543 static int
 544 linux_socketpair(struct linux_socketpair_args *args, int *res)
 545 {
 546         struct linux_socketpair_args linux_args;
 547         int error, domain, sockv[2];
 548
 549         error = copyin(args, &linux_args, sizeof(linux_args));
 550         if (error)
 551                 return (error);
 552
 553         domain = linux_to_bsd_domain(linux_args.domain);
 554         if (domain == -1)
 555                 return (EINVAL);
 556         error = kern_socketpair(domain, linux_args.type, linux_args.protocol,
 557             sockv);
 558
 559         if (error == 0)
 560                 error = copyout(sockv, linux_args.rsv, sizeof(sockv));
 561         return(error);
 562 }
 563
 564 struct linux_send_args {
 565         int s;
 566         void *msg;
 567         int len;
 568         int flags;
 569 };
 570
 571 static int
 572 linux_send(struct linux_send_args *args, size_t *res)
 573 {
 574         struct linux_send_args linux_args;
 575         struct thread *td = curthread;
 576         struct uio auio;
 577         struct iovec aiov;
 578         int error;
 579
 580         error = copyin(args, &linux_args, sizeof(linux_args));
 581         if (error)
 582                 return (error);
 583
 584         aiov.iov_base = linux_args.msg;
 585         aiov.iov_len = linux_args.len;
 586         auio.uio_iov = &aiov;
 587         auio.uio_iovcnt = 1;
 588         auio.uio_offset = 0;
 589         auio.uio_resid = linux_args.len;
 590         auio.uio_segflg = UIO_USERSPACE;
 591         auio.uio_rw = UIO_WRITE;
 592         auio.uio_td = td;
 593
 594         error = kern_sendmsg(linux_args.s, NULL, &auio, NULL,
 595                              linux_args.flags, res);
 596
 597         return(error);
 598 }
 599
 600 struct linux_recv_args {
 601         int s;
 602         void *msg;
 603         int len;
 604         int flags;
 605 };
 606
 607 static int
 608 linux_recv(struct linux_recv_args *args, size_t *res)
 609 {
 610         struct linux_recv_args linux_args;
 611         struct thread *td = curthread;
 612         struct uio auio;
 613         struct iovec aiov;
 614         int error;
 615
 616         error = copyin(args, &linux_args, sizeof(linux_args));
 617         if (error)
 618                 return (error);
 619
 620         aiov.iov_base = linux_args.msg;
 621         aiov.iov_len = linux_args.len;
 622         auio.uio_iov = &aiov;
 623         auio.uio_iovcnt = 1;
 624         auio.uio_offset = 0;
 625         auio.uio_resid = linux_args.len;
 626         auio.uio_segflg = UIO_USERSPACE;
 627         auio.uio_rw = UIO_READ;
 628         auio.uio_td = td;
 629
 630         error = kern_recvmsg(linux_args.s, NULL, &auio, NULL,
 631                              &linux_args.flags, res);
 632
 633         return(error);
 634 }
 635
 636 struct linux_sendto_args {
 637         int s;
 638         void *msg;
 639         int len;
 640         int flags;
 641         struct sockaddr *to;
 642         int tolen;
 643 };
 644
 645 static int
 646 linux_sendto(struct linux_sendto_args *args, size_t *res)
 647 {
 648         struct linux_sendto_args linux_args;
 649         struct thread *td = curthread;
 650         struct uio auio;
 651         struct iovec aiov;
 652         struct sockopt sopt;
 653         struct sockaddr *sa = NULL;
 654         caddr_t msg = NULL;
 655         int error, optval;
 656
 657         error = copyin(args, &linux_args, sizeof(linux_args));
 658         if (error)
 659                 return (error);
 660
 661         if (linux_args.to) {
 662                 error = linux_getsockaddr(&sa, linux_args.to,
 663                     linux_args.tolen);
 664                 if (error)
 665                         return (error);
 666         }
 667
 668         /*
 669          * Check to see if the IP_HDRINCL option is set.
 670          */
 671         sopt.sopt_dir = SOPT_GET;
 672         sopt.sopt_level = IPPROTO_IP;
 673         sopt.sopt_name = IP_HDRINCL;
 674         sopt.sopt_val = &optval;
 675         sopt.sopt_valsize = sizeof(optval);
 676         sopt.sopt_td = NULL;
 677
 678         if (kern_getsockopt(linux_args.s, &sopt) != 0)
 679                 optval = 0;
 680
 681         if (optval == 0) {
 682                 /*
 683                  * IP_HDRINCL is not set.  Package the message as usual.
 684                  */
 685                 aiov.iov_base = linux_args.msg;
 686                 aiov.iov_len = linux_args.len;
 687                 auio.uio_iov = &aiov;
 688                 auio.uio_iovcnt = 1;
 689                 auio.uio_offset = 0;
 690                 auio.uio_resid = linux_args.len;
 691                 auio.uio_segflg = UIO_USERSPACE;
 692                 auio.uio_rw = UIO_WRITE;
 693                 auio.uio_td = td;
 694         } else {
 695                 /*
 696                  * IP_HDRINCL is set.  We must convert the beginning of
 697                  * the packet header so we can feed it to the BSD kernel.
 698                  */
 699
 700                 /*
 701                  * Check that the packet header is long enough to contain
 702                  * the fields of interest.  This relies on the fact that
 703                  * the fragment offset field comes after the length field.
 704                  */
 705                 if (linux_args.len < offsetof(struct ip, ip_off))
 706                         return (EINVAL);
 707
 708                 MALLOC(msg, caddr_t, linux_args.len, M_LINUX, M_WAITOK);
 709                 error = copyin(linux_args.msg, msg, linux_args.len);
 710                 if (error)
 711                         goto cleanup;
 712
 713                 /* Fix the ip_len and ip_off fields.  */
 714                 ((struct ip *)msg)->ip_len = linux_args.len;
 715                 ((struct ip *)msg)->ip_off = ntohs(((struct ip *)msg)->ip_off);
 716
 717                 aiov.iov_base = msg;
 718                 aiov.iov_len = linux_args.len;
 719                 auio.uio_iov = &aiov;
 720                 auio.uio_iovcnt = 1;
 721                 auio.uio_offset = 0;
 722                 auio.uio_resid = linux_args.len;
 723                 auio.uio_segflg = UIO_SYSSPACE;
 724                 auio.uio_rw = UIO_WRITE;
 725                 auio.uio_td = td;
 726         }
 727
 728         error = kern_sendmsg(linux_args.s, sa, &auio, NULL,
 729                              linux_args.flags, res);
 730
 731 cleanup:
 732         if (sa)
 733                 FREE(sa, M_SONAME);
 734         if (msg)
 735                 FREE(msg, M_LINUX);
 736         return(error);
 737 }
 738
 739 struct linux_recvfrom_args {
 740         int s;
 741         void *buf;
 742         int len;
 743         int flags;
 744         struct sockaddr *from;
 745         int *fromlen;
 746 };
 747
 748 static int
 749 linux_recvfrom(struct linux_recvfrom_args *args, size_t *res)
 750 {
 751         struct linux_recvfrom_args linux_args;
 752         struct thread *td = curthread;
 753         struct uio auio;
 754         struct iovec aiov;
 755         struct sockaddr *sa = NULL;
 756         int error, fromlen, flags;
 757
 758         error = copyin(args, &linux_args, sizeof(linux_args));
 759         if (error)
 760                 return (error);
 761
 762         if (linux_args.from && linux_args.fromlen) {
 763                 error = copyin(linux_args.fromlen, &fromlen, sizeof(fromlen));
 764                 if (error)
 765                         return (error);
 766                 if (fromlen < 0)
 767                         return (EINVAL);
 768         } else {
 769                 fromlen = 0;
 770         }
 771         aiov.iov_base = linux_args.buf;
 772         aiov.iov_len = linux_args.len;
 773         auio.uio_iov = &aiov;
 774         auio.uio_iovcnt = 1;
 775         auio.uio_offset = 0;
 776         auio.uio_resid = linux_args.len;
 777         auio.uio_segflg = UIO_USERSPACE;
 778         auio.uio_rw = UIO_READ;
 779         auio.uio_td = td;
 780
 781         flags = linux_to_bsd_msg_flags(linux_args.flags);
 782
 783         error = kern_recvmsg(linux_args.s, linux_args.from ? &sa : NULL, &auio,
 784                              NULL, &flags, res);
 785
 786         if (error == 0 && linux_args.from) {
 787                 if (sa != NULL) {
 788                         fromlen = MIN(fromlen, sa->sa_len);
 789                         error = linux_copyout_sockaddr(sa, linux_args.from,
 790                                                         fromlen);
 791                 } else
 792                         fromlen = 0;
 793                 if (error == 0)
 794                         copyout(&fromlen, linux_args.fromlen,
 795                             sizeof(fromlen));
 796         }
 797         if (sa)
 798                 FREE(sa, M_SONAME);
 799
 800         return(error);
 801 }
 802
 803 struct linux_sendmsg_args {
 804         int s;
 805         struct msghdr *msg;
 806         int flags;
 807 };
 808
 809 static int
 810 linux_sendmsg(struct linux_sendmsg_args *args, size_t *res)
 811 {
 812         struct linux_sendmsg_args linux_args;
 813         struct thread *td = curthread;
 814         struct msghdr msg;
 815         struct uio auio;
 816         struct iovec aiov[UIO_SMALLIOV], *iov = NULL;
 817         struct sockaddr *sa = NULL;
 818         struct mbuf *control = NULL;
 819         int error;
 820
 821         error = copyin(args, &linux_args, sizeof(linux_args));
 822         if (error)
 823                 return (error);
 824
 825         error = copyin(linux_args.msg, &msg, sizeof(msg));
 826         if (error)
 827                 return (error);
 828
 829         /*
 830          * Conditionally copyin msg.msg_name.
 831          */
 832         if (msg.msg_name) {
 833                 error = linux_getsockaddr(&sa, msg.msg_name, msg.msg_namelen);
 834                 if (error)
 835                         return (error);
 836         }
 837
 838         /*
 839          * Populate auio.
 840          */
 841         error = iovec_copyin(msg.msg_iov, &iov, aiov, msg.msg_iovlen,
 842                              &auio.uio_resid);
 843         if (error)
 844                 goto cleanup2;
 845         auio.uio_iov = iov;
 846         auio.uio_iovcnt = msg.msg_iovlen;
 847         auio.uio_offset = 0;
 848         auio.uio_segflg = UIO_USERSPACE;
 849         auio.uio_rw = UIO_WRITE;
 850         auio.uio_td = td;
 851
 852         /*
 853          * Conditionally copyin msg.msg_control.
 854          */
 855         if (msg.msg_control) {
 856                 if (msg.msg_controllen < sizeof(struct cmsghdr) ||
 857                     msg.msg_controllen > MLEN) {
 858                         error = EINVAL;
 859                         goto cleanup;
 860                 }
 861                 control = m_get(MB_WAIT, MT_CONTROL);
 862                 if (control == NULL) {
 863                         error = ENOBUFS;
 864                         goto cleanup;
 865                 }
 866                 control->m_len = msg.msg_controllen;
 867                 error = copyin(msg.msg_control, mtod(control, caddr_t),
 868                     msg.msg_controllen);
 869                 if (error) {
 870                         m_free(control);
 871                         goto cleanup;
 872                 }
 873                 /*
 874                  * Linux and BSD both support SCM_RIGHTS.  If a linux binary
 875                  * wants anything else with an option level of SOL_SOCKET,
 876                  * we don't support it.
 877                  */
 878                 if (mtod(control, struct cmsghdr *)->cmsg_level ==
 879                     SOL_SOCKET &&
 880                     mtod(control, struct cmsghdr *)->cmsg_type !=
 881                     SCM_RIGHTS) {
 882                         m_free(control);
 883                         error = EINVAL;
 884                         goto cleanup;
 885                 }
 886         }
 887
 888         error = kern_sendmsg(linux_args.s, sa, &auio, control,
 889                              linux_args.flags, res);
 890
 891 cleanup:
 892         iovec_free(&iov, aiov);
 893 cleanup2:
 894         if (sa)
 895                 FREE(sa, M_SONAME);
 896         return (error);
 897 }
 898
 899 struct linux_recvmsg_args {
 900         int s;
 901         struct msghdr *msg;
 902         int flags;
 903 };
 904
 905 static int
 906 linux_recvmsg(struct linux_recvmsg_args *args, size_t *res)
 907 {
 908         struct linux_recvmsg_args linux_args;
 909         struct thread *td = curthread;
 910         struct msghdr msg;
 911         struct uio auio;
 912         struct iovec aiov[UIO_SMALLIOV], *iov = NULL;
 913         struct mbuf *m, *control = NULL;
 914         struct sockaddr *sa = NULL;
 915         caddr_t ctlbuf;
 916         socklen_t *ufromlenp, *ucontrollenp;
 917         int error, fromlen, controllen, len, flags, *uflagsp;
 918
 919         error = copyin(args, &linux_args, sizeof(linux_args));
 920         if (error)
 921                 return (error);
 922
 923         error = copyin(linux_args.msg, &msg, sizeof(struct msghdr));
 924         if (error)
 925                 return (error);
 926
 927         if (msg.msg_name && msg.msg_namelen < 0)
 928                 return (EINVAL);
 929         if (msg.msg_control && msg.msg_controllen < 0)
 930                 return (EINVAL);
 931
 932         ufromlenp = (socklen_t *)((caddr_t)linux_args.msg +
 933             offsetof(struct msghdr, msg_namelen));
 934         ucontrollenp = (socklen_t *)((caddr_t)linux_args.msg +
 935             offsetof(struct msghdr, msg_controllen));
 936         uflagsp = (int *)((caddr_t)linux_args.msg +
 937             offsetof(struct msghdr, msg_flags));
 938
 939         /*
 940          * Populate auio.
 941          */
 942         error = iovec_copyin(msg.msg_iov, &iov, aiov, msg.msg_iovlen,
 943                              &auio.uio_resid);
 944         if (error)
 945                 return (error);
 946         auio.uio_iov = iov;
 947         auio.uio_iovcnt = msg.msg_iovlen;
 948         auio.uio_offset = 0;
 949         auio.uio_segflg = UIO_USERSPACE;
 950         auio.uio_rw = UIO_READ;
 951         auio.uio_td = td;
 952
 953         flags = linux_to_bsd_msg_flags(linux_args.flags);
 954
 955         error = kern_recvmsg(linux_args.s, msg.msg_name ? &sa : NULL, &auio,
 956                              msg.msg_control ? &control : NULL, &flags, res);
 957
 958         /*
 959          * Copyout msg.msg_name and msg.msg_namelen.
 960          */
 961         if (error == 0 && msg.msg_name) {
 962                 if (sa != NULL) {
 963                         fromlen = MIN(msg.msg_namelen, sa->sa_len);
 964                         error = linux_copyout_sockaddr(sa, msg.msg_name,
 965                                                         fromlen);
 966                 } else
 967                         fromlen = 0;
 968                 if (error == 0)
 969                         error = copyout(&fromlen, ufromlenp,
 970                             sizeof(*ufromlenp));
 971         }
 972
 973         /*
 974          * Copyout msg.msg_control and msg.msg_controllen.
 975          */
 976         if (error == 0 && msg.msg_control) {
 977                 /*
 978                  * Linux and BSD both support SCM_RIGHTS.  If a linux binary
 979                  * wants anything else with an option level of SOL_SOCKET,
 980                  * we don't support it.
 981                  */
 982                 if (mtod((struct mbuf *)msg.msg_control,
 983                     struct cmsghdr *)->cmsg_level == SOL_SOCKET &&
 984                     mtod((struct mbuf *)msg.msg_control,
 985                     struct cmsghdr *)->cmsg_type != SCM_RIGHTS) {
 986                         error = EINVAL;
 987                         goto cleanup;
 988                 }
 989
 990                 len = msg.msg_controllen;
 991                 m = control;
 992                 ctlbuf = (caddr_t)msg.msg_control;
 993
 994                 while (m && len > 0) {
 995                         unsigned int tocopy;
 996
 997                         if (len >= m->m_len) {
 998                                 tocopy = m->m_len;
 999                         } else {
1000                                 msg.msg_flags |= MSG_CTRUNC;
1001                                 tocopy = len;
1002                         }
1003
1004                         error = copyout(mtod(m, caddr_t), ctlbuf,
1005                             tocopy);
1006                         if (error)
1007                                 goto cleanup;
1008
1009                         ctlbuf += tocopy;
1010                         len -= tocopy;
1011                         m = m->m_next;
1012                 }
1013                 controllen = ctlbuf - (caddr_t)msg.msg_control;
1014                 error = copyout(&controllen, ucontrollenp,
1015                     sizeof(*ucontrollenp));
1016         }
1017
1018         if (error == 0)
1019                 error = copyout(&flags, uflagsp, sizeof(*uflagsp));
1020
1021 cleanup:
1022         if (sa)
1023                 FREE(sa, M_SONAME);
1024         iovec_free(&iov, aiov);
1025         if (control)
1026                 m_freem(control);
1027         return (error);
1028 }
1029
1030 struct linux_shutdown_args {
1031         int s;
1032         int how;
1033 };
1034
1035 static int
1036 linux_shutdown(struct linux_shutdown_args *args, int *res)
1037 {
1038         struct linux_shutdown_args linux_args;
1039         int error;
1040
1041         error = copyin(args, &linux_args, sizeof(linux_args));
1042         if (error)
1043                 return (error);
1044
1045         error = kern_shutdown(linux_args.s, linux_args.how);
1046
1047         return (error);
1048 }
1049
1050 struct linux_setsockopt_args {
1051         int s;
1052         int level;
1053         int optname;
1054         void *optval;
1055         int optlen;
1056 };
1057
1058 static int
1059 linux_setsockopt(struct linux_setsockopt_args *args, int *res)
1060 {
1061         struct linux_setsockopt_args linux_args;
1062         struct thread *td = curthread;
1063         struct sockopt sopt;
1064         int error, name, level;
1065
1066         error = copyin(args, &linux_args, sizeof(linux_args));
1067         if (error)
1068                 return (error);
1069
1070         level = linux_to_bsd_sockopt_level(linux_args.level);
1071         switch (level) {
1072         case SOL_SOCKET:
1073                 name = linux_to_bsd_so_sockopt(linux_args.optname);
1074                 break;
1075         case IPPROTO_IP:
1076                 name = linux_to_bsd_ip_sockopt(linux_args.optname);
1077                 break;
1078         case IPPROTO_TCP:
1079                 /* Linux TCP option values match BSD's */
1080                 name = linux_args.optname;
1081                 break;
1082         default:
1083                 name = -1;
1084                 break;
1085         }
1086         if (name == -1)
1087                 return (EINVAL);
1088
1089         sopt.sopt_dir = SOPT_SET;
1090         sopt.sopt_level = level;
1091         sopt.sopt_name = name;
1092         sopt.sopt_valsize = linux_args.optlen;
1093         sopt.sopt_td = td;
1094
1095         if (sopt.sopt_valsize < 0 || sopt.sopt_valsize > SOMAXOPT_SIZE)
1096                 return (EINVAL);
1097
1098         if (linux_args.optval) {
1099                 sopt.sopt_val = kmalloc(sopt.sopt_valsize, M_TEMP, M_WAITOK);
1100                 error = copyin(linux_args.optval, sopt.sopt_val, sopt.sopt_valsize);
1101                 if (error)
1102                         goto out;
1103         } else {
1104                 sopt.sopt_val = NULL;
1105         }
1106         error = kern_setsockopt(linux_args.s, &sopt);
1107         if (error)
1108                 goto out;
1109         if (linux_args.optval)
1110                 error = copyout(sopt.sopt_val, linux_args.optval,
1111                                 sopt.sopt_valsize);
1112 out:
1113         if (linux_args.optval)
1114                 kfree(sopt.sopt_val, M_TEMP);
1115         return(error);
1116 }
1117
1118 struct linux_getsockopt_args {
1119         int s;
1120         int level;
1121         int optname;
1122         void *optval;
1123         int *optlen;
1124 };
1125
1126 static int
1127 linux_getsockopt(struct linux_getsockopt_args *args, int *res)
1128 {
1129         struct linux_getsockopt_args linux_args;
1130         struct thread *td = curthread;
1131         struct sockopt sopt;
1132         int error, name, valsize, level;
1133
1134         error = copyin(args, &linux_args, sizeof(linux_args));
1135         if (error)
1136                 return (error);
1137
1138         if (linux_args.optval) {
1139                 error = copyin(linux_args.optlen, &valsize, sizeof(valsize));
1140                 if (error)
1141                         return (error);
1142                 if (valsize < 0 || valsize > SOMAXOPT_SIZE)
1143                         return (EINVAL);
1144         } else {
1145                 valsize = 0;
1146         }
1147
1148         level = linux_to_bsd_sockopt_level(linux_args.level);
1149         switch (level) {
1150         case SOL_SOCKET:
1151                 name = linux_to_bsd_so_sockopt(linux_args.optname);
1152                 break;
1153         case IPPROTO_IP:
1154                 name = linux_to_bsd_ip_sockopt(linux_args.optname);
1155                 break;
1156         case IPPROTO_TCP:
1157                 /* Linux TCP option values match BSD's */
1158                 name = linux_args.optname;
1159                 break;
1160         default:
1161                 name = -1;
1162                 break;
1163         }
1164         if (name == -1)
1165                 return (EINVAL);
1166
1167         sopt.sopt_dir = SOPT_GET;
1168         sopt.sopt_level = level;
1169         sopt.sopt_name = name;
1170         sopt.sopt_valsize = valsize;
1171         sopt.sopt_td = td;
1172
1173         if (linux_args.optval) {
1174                 sopt.sopt_val = kmalloc(sopt.sopt_valsize, M_TEMP, M_WAITOK);
1175                 error = copyin(linux_args.optval, sopt.sopt_val, sopt.sopt_valsize);
1176                 if (error)
1177                         goto out;
1178         } else {
1179                 sopt.sopt_val = NULL;
1180         }
1181         error = kern_getsockopt(linux_args.s, &sopt);
1182         if (error)
1183                 goto out;
1184         valsize = sopt.sopt_valsize;
1185         error = copyout(&valsize, linux_args.optlen, sizeof(valsize));
1186         if (error)
1187                 goto out;
1188         if (linux_args.optval)
1189                 error = copyout(sopt.sopt_val, linux_args.optval, sopt.sopt_valsize);
1190 out:
1191         if (linux_args.optval)
1192                 kfree(sopt.sopt_val, M_TEMP);
1193         return(error);
1194 }
1195
1196 /*
1197  * MPALMOSTSAFE
1198  */
1199 int
1200 sys_linux_socketcall(struct linux_socketcall_args *args)
1201 {
1202         void *arg = (void *)args->args;
1203
1204         get_mplock();
1205
1206         switch (args->what) {
1207         case LINUX_SOCKET:
1208                 return (linux_socket(arg, &args->sysmsg_result));
1209         case LINUX_BIND:
1210                 return (linux_bind(arg, &args->sysmsg_result));
1211         case LINUX_CONNECT:
1212                 return (linux_connect(arg, &args->sysmsg_result));
1213         case LINUX_LISTEN:
1214                 return (linux_listen(arg, &args->sysmsg_result));
1215         case LINUX_ACCEPT:
1216                 return (linux_accept(arg, &args->sysmsg_result));
1217         case LINUX_GETSOCKNAME:
1218                 return (linux_getsockname(arg, &args->sysmsg_result));
1219         case LINUX_GETPEERNAME:
1220                 return (linux_getpeername(arg, &args->sysmsg_result));
1221         case LINUX_SOCKETPAIR:
1222                 return (linux_socketpair(arg, &args->sysmsg_result));
1223         case LINUX_SEND:
1224                 return (linux_send(arg, &args->sysmsg_szresult));
1225         case LINUX_RECV:
1226                 return (linux_recv(arg, &args->sysmsg_szresult));
1227         case LINUX_SENDTO:
1228                 return (linux_sendto(arg, &args->sysmsg_szresult));
1229         case LINUX_RECVFROM:
1230                 return (linux_recvfrom(arg, &args->sysmsg_szresult));
1231         case LINUX_SHUTDOWN:
1232                 return (linux_shutdown(arg, &args->sysmsg_result));
1233         case LINUX_SETSOCKOPT:
1234                 return (linux_setsockopt(arg, &args->sysmsg_result));
1235         case LINUX_GETSOCKOPT:
1236                 return (linux_getsockopt(arg, &args->sysmsg_result));
1237         case LINUX_SENDMSG:
1238                 return (linux_sendmsg(arg, &args->sysmsg_szresult));
1239         case LINUX_RECVMSG:
1240                 return (linux_recvmsg(arg, &args->sysmsg_szresult));
1241         }
1242         rel_mplock();
1243
1244         uprintf("LINUX: 'socket' typ=%d not implemented\n", args->what);
1245         return (ENOSYS);
1246 }