sys/kern/sysv_msg.c

   1 /* $FreeBSD: src/sys/kern/sysv_msg.c,v 1.23.2.5 2002/12/31 08:54:53 maxim Exp $ */
   2
   3 /*
   4  * Implementation of SVID messages
   5  *
   6  * Author:  Daniel Boulet
   7  *
   8  * Copyright 1993 Daniel Boulet and RTMX Inc.
   9  *
  10  * This system call was implemented by Daniel Boulet under contract from RTMX.
  11  *
  12  * Redistribution and use in source forms, with and without modification,
  13  * are permitted provided that this entire comment appears intact.
  14  *
  15  * Redistribution in binary form may occur without any restrictions.
  16  * Obviously, it would be nice if you gave credit where credit is due
  17  * but requiring it would be too onerous.
  18  *
  19  * This software is provided ``AS IS'' without any warranties of any kind.
  20  */
  21
  22 #include "opt_sysvipc.h"
  23
  24 #include <sys/param.h>
  25 #include <sys/systm.h>
  26 #include <sys/sysproto.h>
  27 #include <sys/kernel.h>
  28 #include <sys/proc.h>
  29 #include <sys/priv.h>
  30 #include <sys/msg.h>
  31 #include <sys/sysent.h>
  32 #include <sys/sysctl.h>
  33 #include <sys/malloc.h>
  34 #include <sys/jail.h>
  35
  36 #include <sys/mplock2.h>
  37
  38 static MALLOC_DEFINE(M_MSG, "msg", "SVID compatible message queues");
  39
  40 static void msginit (void *);
  41
  42 #define MSG_DEBUG
  43 #undef MSG_DEBUG_OK
  44
  45 static void msg_freehdr (struct msg *msghdr);
  46
  47 struct msg {
  48         struct  msg *msg_next;  /* next msg in the chain */
  49         long    msg_type;       /* type of this message */
  50                                 /* >0 -> type of this message */
  51                                 /* 0 -> free header */
  52         u_short msg_ts;         /* size of this message */
  53         short   msg_spot;       /* location of start of msg in buffer */
  54 };
  55
  56
  57 #ifndef MSGSSZ
  58 #define MSGSSZ  8               /* Each segment must be 2^N long */
  59 #endif
  60 #ifndef MSGSEG
  61 #define MSGSEG  2048            /* must be less than 32767 */
  62 #endif
  63 #define MSGMAX  (MSGSSZ*MSGSEG)
  64 #ifndef MSGMNB
  65 #define MSGMNB  2048            /* max # of bytes in a queue */
  66 #endif
  67 #ifndef MSGMNI
  68 #define MSGMNI  40
  69 #endif
  70 #ifndef MSGTQL
  71 #define MSGTQL  40
  72 #endif
  73
  74 /*
  75  * Based on the configuration parameters described in an SVR2 (yes, two)
  76  * config(1m) man page.
  77  *
  78  * Each message is broken up and stored in segments that are msgssz bytes
  79  * long.  For efficiency reasons, this should be a power of two.  Also,
  80  * it doesn't make sense if it is less than 8 or greater than about 256.
  81  * Consequently, msginit in kern/sysv_msg.c checks that msgssz is a power of
  82  * two between 8 and 1024 inclusive (and panic's if it isn't).
  83  */
  84 struct msginfo msginfo = {
  85                 MSGMAX,         /* max chars in a message */
  86                 MSGMNI,         /* # of message queue identifiers */
  87                 MSGMNB,         /* max chars in a queue */
  88                 MSGTQL,         /* max messages in system */
  89                 MSGSSZ,         /* size of a message segment */
  90                                 /* (must be small power of 2 greater than 4) */
  91                 MSGSEG          /* number of message segments */
  92 };
  93
  94 /*
  95  * macros to convert between msqid_ds's and msqid's.
  96  * (specific to this implementation)
  97  */
  98 #define MSQID(ix,ds)    ((ix) & 0xffff | (((ds).msg_perm.seq << 16) & 0xffff0000))
  99 #define MSQID_IX(id)    ((id) & 0xffff)
 100 #define MSQID_SEQ(id)   (((id) >> 16) & 0xffff)
 101
 102 /*
 103  * The rest of this file is specific to this particular implementation.
 104  */
 105
 106 struct msgmap {
 107         short   next;           /* next segment in buffer */
 108                                 /* -1 -> available */
 109                                 /* 0..(MSGSEG-1) -> index of next segment */
 110 };
 111
 112 #define MSG_LOCKED      01000   /* Is this msqid_ds locked? */
 113
 114 static int nfree_msgmaps;       /* # of free map entries */
 115 static short free_msgmaps;      /* head of linked list of free map entries */
 116 static struct msg *free_msghdrs;/* list of free msg headers */
 117 static char *msgpool;           /* MSGMAX byte long msg buffer pool */
 118 static struct msgmap *msgmaps;  /* MSGSEG msgmap structures */
 119 static struct msg *msghdrs;     /* MSGTQL msg headers */
 120 static struct msqid_ds *msqids; /* MSGMNI msqid_ds struct's */
 121
 122 static void
 123 msginit(void *dummy)
 124 {
 125         int i;
 126
 127         msginfo.msgmax = msginfo.msgseg * msginfo.msgssz;
 128         msgpool = kmalloc(msginfo.msgmax, M_MSG, M_WAITOK);
 129         msgmaps = kmalloc(sizeof(struct msgmap) * msginfo.msgseg, M_MSG, M_WAITOK);
 130         msghdrs = kmalloc(sizeof(struct msg) * msginfo.msgtql, M_MSG, M_WAITOK);
 131         msqids = kmalloc(sizeof(struct msqid_ds) * msginfo.msgmni, M_MSG, M_WAITOK);
 132
 133         /*
 134          * msginfo.msgssz should be a power of two for efficiency reasons.
 135          * It is also pretty silly if msginfo.msgssz is less than 8
 136          * or greater than about 256 so ...
 137          */
 138
 139         i = 8;
 140         while (i < 1024 && i != msginfo.msgssz)
 141                 i <<= 1;
 142         if (i != msginfo.msgssz) {
 143                 kprintf("msginfo.msgssz=%d (0x%x)\n", msginfo.msgssz,
 144                     msginfo.msgssz);
 145                 panic("msginfo.msgssz not a small power of 2");
 146         }
 147
 148         if (msginfo.msgseg > 32767) {
 149                 kprintf("msginfo.msgseg=%d\n", msginfo.msgseg);
 150                 panic("msginfo.msgseg > 32767");
 151         }
 152
 153         for (i = 0; i < msginfo.msgseg; i++) {
 154                 if (i > 0)
 155                         msgmaps[i-1].next = i;
 156                 msgmaps[i].next = -1;   /* implies entry is available */
 157         }
 158         free_msgmaps = 0;
 159         nfree_msgmaps = msginfo.msgseg;
 160
 161         for (i = 0; i < msginfo.msgtql; i++) {
 162                 msghdrs[i].msg_type = 0;
 163                 if (i > 0)
 164                         msghdrs[i-1].msg_next = &msghdrs[i];
 165                 msghdrs[i].msg_next = NULL;
 166         }
 167         free_msghdrs = &msghdrs[0];
 168
 169         for (i = 0; i < msginfo.msgmni; i++) {
 170                 msqids[i].msg_qbytes = 0;       /* implies entry is available */
 171                 msqids[i].msg_perm.seq = 0;     /* reset to a known value */
 172                 msqids[i].msg_perm.mode = 0;
 173         }
 174 }
 175 SYSINIT(sysv_msg, SI_SUB_SYSV_MSG, SI_ORDER_FIRST, msginit, NULL);
 176
 177 static void
 178 msg_freehdr(struct msg *msghdr)
 179 {
 180         while (msghdr->msg_ts > 0) {
 181                 short next;
 182                 if (msghdr->msg_spot < 0 || msghdr->msg_spot >= msginfo.msgseg)
 183                         panic("msghdr->msg_spot out of range");
 184                 next = msgmaps[msghdr->msg_spot].next;
 185                 msgmaps[msghdr->msg_spot].next = free_msgmaps;
 186                 free_msgmaps = msghdr->msg_spot;
 187                 nfree_msgmaps++;
 188                 msghdr->msg_spot = next;
 189                 if (msghdr->msg_ts >= msginfo.msgssz)
 190                         msghdr->msg_ts -= msginfo.msgssz;
 191                 else
 192                         msghdr->msg_ts = 0;
 193         }
 194         if (msghdr->msg_spot != -1)
 195                 panic("msghdr->msg_spot != -1");
 196         msghdr->msg_next = free_msghdrs;
 197         free_msghdrs = msghdr;
 198 }
 199
 200 /*
 201  * MPALMOSTSAFE
 202  */
 203 int
 204 sys_msgctl(struct msgctl_args *uap)
 205 {
 206         struct thread *td = curthread;
 207         struct proc *p = td->td_proc;
 208         int msqid = uap->msqid;
 209         int cmd = uap->cmd;
 210         struct msqid_ds *user_msqptr = uap->buf;
 211         int rval, eval;
 212         struct msqid_ds msqbuf;
 213         struct msqid_ds *msqptr;
 214
 215 #ifdef MSG_DEBUG_OK
 216         kprintf("call to msgctl(%d, %d, 0x%x)\n", msqid, cmd, user_msqptr);
 217 #endif
 218
 219         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 220                 return (ENOSYS);
 221
 222         get_mplock();
 223         msqid = IPCID_TO_IX(msqid);
 224
 225         if (msqid < 0 || msqid >= msginfo.msgmni) {
 226 #ifdef MSG_DEBUG_OK
 227                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 228                     msginfo.msgmni);
 229 #endif
 230                 eval = EINVAL;
 231                 goto done;
 232         }
 233
 234         msqptr = &msqids[msqid];
 235
 236         if (msqptr->msg_qbytes == 0) {
 237 #ifdef MSG_DEBUG_OK
 238                 kprintf("no such msqid\n");
 239 #endif
 240                 eval = EINVAL;
 241                 goto done;
 242         }
 243         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 244 #ifdef MSG_DEBUG_OK
 245                 kprintf("wrong sequence number\n");
 246 #endif
 247                 eval = EINVAL;
 248                 goto done;
 249         }
 250
 251         rval = 0;
 252
 253         switch (cmd) {
 254         case IPC_RMID:
 255         {
 256                 struct msg *msghdr;
 257                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)) != 0)
 258                         break;
 259                 /* Free the message headers */
 260                 msghdr = msqptr->msg_first;
 261                 while (msghdr != NULL) {
 262                         struct msg *msghdr_tmp;
 263
 264                         /* Free the segments of each message */
 265                         msqptr->msg_cbytes -= msghdr->msg_ts;
 266                         msqptr->msg_qnum--;
 267                         msghdr_tmp = msghdr;
 268                         msghdr = msghdr->msg_next;
 269                         msg_freehdr(msghdr_tmp);
 270                 }
 271
 272                 if (msqptr->msg_cbytes != 0)
 273                         panic("msg_cbytes is screwed up");
 274                 if (msqptr->msg_qnum != 0)
 275                         panic("msg_qnum is screwed up");
 276
 277                 msqptr->msg_qbytes = 0; /* Mark it as free */
 278
 279                 wakeup((caddr_t)msqptr);
 280         }
 281
 282                 break;
 283
 284         case IPC_SET:
 285                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)) != 0)
 286                         break;
 287                 if ((eval = copyin(user_msqptr, &msqbuf, sizeof(msqbuf))) != 0)
 288                         break;
 289                 if (msqbuf.msg_qbytes > msqptr->msg_qbytes) {
 290                         eval = priv_check(td, PRIV_ROOT);
 291                         if (eval)
 292                                 break;
 293                 }
 294                 if (msqbuf.msg_qbytes > msginfo.msgmnb) {
 295 #ifdef MSG_DEBUG_OK
 296                         kprintf("can't increase msg_qbytes beyond %d (truncating)\n",
 297                             msginfo.msgmnb);
 298 #endif
 299                         msqbuf.msg_qbytes = msginfo.msgmnb;     /* silently restrict qbytes to system limit */
 300                 }
 301                 if (msqbuf.msg_qbytes == 0) {
 302 #ifdef MSG_DEBUG_OK
 303                         kprintf("can't reduce msg_qbytes to 0\n");
 304 #endif
 305                         eval = EINVAL;          /* non-standard errno! */
 306                         break;
 307                 }
 308                 msqptr->msg_perm.uid = msqbuf.msg_perm.uid;     /* change the owner */
 309                 msqptr->msg_perm.gid = msqbuf.msg_perm.gid;     /* change the owner */
 310                 msqptr->msg_perm.mode = (msqptr->msg_perm.mode & ~0777) |
 311                                         (msqbuf.msg_perm.mode & 0777);
 312                 msqptr->msg_qbytes = msqbuf.msg_qbytes;
 313                 msqptr->msg_ctime = time_second;
 314                 break;
 315
 316         case IPC_STAT:
 317                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
 318 #ifdef MSG_DEBUG_OK
 319                         kprintf("requester doesn't have read access\n");
 320 #endif
 321                         eval = EINVAL;
 322                         break;
 323                 }
 324                 eval = copyout(msqptr, user_msqptr, sizeof(struct msqid_ds));
 325                 break;
 326
 327         default:
 328 #ifdef MSG_DEBUG_OK
 329                 kprintf("invalid command %d\n", cmd);
 330 #endif
 331                 eval = EINVAL;
 332                 break;
 333         }
 334 done:
 335         rel_mplock();
 336         if (eval == 0)
 337                 uap->sysmsg_result = rval;
 338         return(eval);
 339 }
 340
 341 /*
 342  * MPALMOSTSAFE
 343  */
 344 int
 345 sys_msgget(struct msgget_args *uap)
 346 {
 347         struct thread *td = curthread;
 348         int msqid, eval;
 349         int key = uap->key;
 350         int msgflg = uap->msgflg;
 351         struct ucred *cred = td->td_ucred;
 352         struct msqid_ds *msqptr = NULL;
 353
 354 #ifdef MSG_DEBUG_OK
 355         kprintf("msgget(0x%x, 0%o)\n", key, msgflg);
 356 #endif
 357         if (!jail_sysvipc_allowed && cred->cr_prison != NULL)
 358                 return (ENOSYS);
 359
 360         eval = 0;
 361         get_mplock();
 362
 363         if (key != IPC_PRIVATE) {
 364                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 365                         msqptr = &msqids[msqid];
 366                         if (msqptr->msg_qbytes != 0 &&
 367                             msqptr->msg_perm.key == key)
 368                                 break;
 369                 }
 370                 if (msqid < msginfo.msgmni) {
 371 #ifdef MSG_DEBUG_OK
 372                         kprintf("found public key\n");
 373 #endif
 374                         if ((msgflg & IPC_CREAT) && (msgflg & IPC_EXCL)) {
 375 #ifdef MSG_DEBUG_OK
 376                                 kprintf("not exclusive\n");
 377 #endif
 378                                 eval = EEXIST;
 379                                 goto done;
 380                         }
 381                         if ((eval = ipcperm(td->td_proc, &msqptr->msg_perm, msgflg & 0700 ))) {
 382 #ifdef MSG_DEBUG_OK
 383                                 kprintf("requester doesn't have 0%o access\n",
 384                                     msgflg & 0700);
 385 #endif
 386                                 goto done;
 387                         }
 388                         goto done;
 389                 }
 390         }
 391
 392 #ifdef MSG_DEBUG_OK
 393         kprintf("need to allocate the msqid_ds\n");
 394 #endif
 395         if (key == IPC_PRIVATE || (msgflg & IPC_CREAT)) {
 396                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 397                         /*
 398                          * Look for an unallocated and unlocked msqid_ds.
 399                          * msqid_ds's can be locked by msgsnd or msgrcv while
 400                          * they are copying the message in/out.  We can't
 401                          * re-use the entry until they release it.
 402                          */
 403                         msqptr = &msqids[msqid];
 404                         if (msqptr->msg_qbytes == 0 &&
 405                             (msqptr->msg_perm.mode & MSG_LOCKED) == 0)
 406                                 break;
 407                 }
 408                 if (msqid == msginfo.msgmni) {
 409 #ifdef MSG_DEBUG_OK
 410                         kprintf("no more msqid_ds's available\n");
 411 #endif
 412                         eval = ENOSPC;
 413                         goto done;
 414                 }
 415 #ifdef MSG_DEBUG_OK
 416                 kprintf("msqid %d is available\n", msqid);
 417 #endif
 418                 msqptr->msg_perm.key = key;
 419                 msqptr->msg_perm.cuid = cred->cr_uid;
 420                 msqptr->msg_perm.uid = cred->cr_uid;
 421                 msqptr->msg_perm.cgid = cred->cr_gid;
 422                 msqptr->msg_perm.gid = cred->cr_gid;
 423                 msqptr->msg_perm.mode = (msgflg & 0777);
 424                 /* Make sure that the returned msqid is unique */
 425                 msqptr->msg_perm.seq = (msqptr->msg_perm.seq + 1) & 0x7fff;
 426                 msqptr->msg_first = NULL;
 427                 msqptr->msg_last = NULL;
 428                 msqptr->msg_cbytes = 0;
 429                 msqptr->msg_qnum = 0;
 430                 msqptr->msg_qbytes = msginfo.msgmnb;
 431                 msqptr->msg_lspid = 0;
 432                 msqptr->msg_lrpid = 0;
 433                 msqptr->msg_stime = 0;
 434                 msqptr->msg_rtime = 0;
 435                 msqptr->msg_ctime = time_second;
 436         } else {
 437 #ifdef MSG_DEBUG_OK
 438                 kprintf("didn't find it and wasn't asked to create it\n");
 439 #endif
 440                 eval = ENOENT;
 441         }
 442
 443 done:
 444         rel_mplock();
 445         /* Construct the unique msqid */
 446         if (eval == 0)
 447                 uap->sysmsg_result = IXSEQ_TO_IPCID(msqid, msqptr->msg_perm);
 448         return(eval);
 449 }
 450
 451 /*
 452  * MPALMOSTSAFE
 453  */
 454 int
 455 sys_msgsnd(struct msgsnd_args *uap)
 456 {
 457         struct thread *td = curthread;
 458         int msqid = uap->msqid;
 459         const void *user_msgp = uap->msgp;
 460         size_t msgsz = uap->msgsz;
 461         int msgflg = uap->msgflg;
 462         int segs_needed, eval;
 463         struct msqid_ds *msqptr;
 464         struct msg *msghdr;
 465         short next;
 466
 467 #ifdef MSG_DEBUG_OK
 468         kprintf("call to msgsnd(%d, 0x%x, %d, %d)\n", msqid, user_msgp, msgsz,
 469             msgflg);
 470 #endif
 471
 472         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 473                 return (ENOSYS);
 474
 475         get_mplock();
 476         msqid = IPCID_TO_IX(msqid);
 477
 478         if (msqid < 0 || msqid >= msginfo.msgmni) {
 479 #ifdef MSG_DEBUG_OK
 480                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 481                     msginfo.msgmni);
 482 #endif
 483                 eval = EINVAL;
 484                 goto done;
 485         }
 486
 487         msqptr = &msqids[msqid];
 488         if (msqptr->msg_qbytes == 0) {
 489 #ifdef MSG_DEBUG_OK
 490                 kprintf("no such message queue id\n");
 491 #endif
 492                 eval = EINVAL;
 493                 goto done;
 494         }
 495         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 496 #ifdef MSG_DEBUG_OK
 497                 kprintf("wrong sequence number\n");
 498 #endif
 499                 eval = EINVAL;
 500                 goto done;
 501         }
 502
 503         if ((eval = ipcperm(td->td_proc, &msqptr->msg_perm, IPC_W))) {
 504 #ifdef MSG_DEBUG_OK
 505                 kprintf("requester doesn't have write access\n");
 506 #endif
 507                 eval = EINVAL;
 508                 goto done;
 509         }
 510
 511         segs_needed = (msgsz + msginfo.msgssz - 1) / msginfo.msgssz;
 512 #ifdef MSG_DEBUG_OK
 513         kprintf("msgsz=%d, msgssz=%d, segs_needed=%d\n", msgsz, msginfo.msgssz,
 514             segs_needed);
 515 #endif
 516         for (;;) {
 517                 int need_more_resources = 0;
 518
 519                 /*
 520                  * check msgsz
 521                  * (inside this loop in case msg_qbytes changes while we sleep)
 522                  */
 523
 524                 if (msgsz > msqptr->msg_qbytes) {
 525 #ifdef MSG_DEBUG_OK
 526                         kprintf("msgsz > msqptr->msg_qbytes\n");
 527 #endif
 528                         eval = EINVAL;
 529                         goto done;
 530                 }
 531
 532                 if (msqptr->msg_perm.mode & MSG_LOCKED) {
 533 #ifdef MSG_DEBUG_OK
 534                         kprintf("msqid is locked\n");
 535 #endif
 536                         need_more_resources = 1;
 537                 }
 538                 if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes) {
 539 #ifdef MSG_DEBUG_OK
 540                         kprintf("msgsz + msg_cbytes > msg_qbytes\n");
 541 #endif
 542                         need_more_resources = 1;
 543                 }
 544                 if (segs_needed > nfree_msgmaps) {
 545 #ifdef MSG_DEBUG_OK
 546                         kprintf("segs_needed > nfree_msgmaps\n");
 547 #endif
 548                         need_more_resources = 1;
 549                 }
 550                 if (free_msghdrs == NULL) {
 551 #ifdef MSG_DEBUG_OK
 552                         kprintf("no more msghdrs\n");
 553 #endif
 554                         need_more_resources = 1;
 555                 }
 556
 557                 if (need_more_resources) {
 558                         int we_own_it;
 559
 560                         if ((msgflg & IPC_NOWAIT) != 0) {
 561 #ifdef MSG_DEBUG_OK
 562                                 kprintf("need more resources but caller doesn't want to wait\n");
 563 #endif
 564                                 eval = EAGAIN;
 565                                 goto done;
 566                         }
 567
 568                         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0) {
 569 #ifdef MSG_DEBUG_OK
 570                                 kprintf("we don't own the msqid_ds\n");
 571 #endif
 572                                 we_own_it = 0;
 573                         } else {
 574                                 /* Force later arrivals to wait for our
 575                                    request */
 576 #ifdef MSG_DEBUG_OK
 577                                 kprintf("we own the msqid_ds\n");
 578 #endif
 579                                 msqptr->msg_perm.mode |= MSG_LOCKED;
 580                                 we_own_it = 1;
 581                         }
 582 #ifdef MSG_DEBUG_OK
 583                         kprintf("goodnight\n");
 584 #endif
 585                         eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 586 #ifdef MSG_DEBUG_OK
 587                         kprintf("good morning, eval=%d\n", eval);
 588 #endif
 589                         if (we_own_it)
 590                                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 591                         if (eval != 0) {
 592 #ifdef MSG_DEBUG_OK
 593                                 kprintf("msgsnd:  interrupted system call\n");
 594 #endif
 595                                 eval = EINTR;
 596                                 goto done;
 597                         }
 598
 599                         /*
 600                          * Make sure that the msq queue still exists
 601                          */
 602
 603                         if (msqptr->msg_qbytes == 0) {
 604 #ifdef MSG_DEBUG_OK
 605                                 kprintf("msqid deleted\n");
 606 #endif
 607                                 eval = EIDRM;
 608                                 goto done;
 609                         }
 610
 611                 } else {
 612 #ifdef MSG_DEBUG_OK
 613                         kprintf("got all the resources that we need\n");
 614 #endif
 615                         break;
 616                 }
 617         }
 618
 619         /*
 620          * We have the resources that we need.
 621          * Make sure!
 622          */
 623
 624         if (msqptr->msg_perm.mode & MSG_LOCKED)
 625                 panic("msg_perm.mode & MSG_LOCKED");
 626         if (segs_needed > nfree_msgmaps)
 627                 panic("segs_needed > nfree_msgmaps");
 628         if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes)
 629                 panic("msgsz + msg_cbytes > msg_qbytes");
 630         if (free_msghdrs == NULL)
 631                 panic("no more msghdrs");
 632
 633         /*
 634          * Re-lock the msqid_ds in case we page-fault when copying in the
 635          * message
 636          */
 637
 638         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0)
 639                 panic("msqid_ds is already locked");
 640         msqptr->msg_perm.mode |= MSG_LOCKED;
 641
 642         /*
 643          * Allocate a message header
 644          */
 645
 646         msghdr = free_msghdrs;
 647         free_msghdrs = msghdr->msg_next;
 648         msghdr->msg_spot = -1;
 649         msghdr->msg_ts = msgsz;
 650
 651         /*
 652          * Allocate space for the message
 653          */
 654
 655         while (segs_needed > 0) {
 656                 if (nfree_msgmaps <= 0)
 657                         panic("not enough msgmaps");
 658                 if (free_msgmaps == -1)
 659                         panic("nil free_msgmaps");
 660                 next = free_msgmaps;
 661                 if (next <= -1)
 662                         panic("next too low #1");
 663                 if (next >= msginfo.msgseg)
 664                         panic("next out of range #1");
 665 #ifdef MSG_DEBUG_OK
 666                 kprintf("allocating segment %d to message\n", next);
 667 #endif
 668                 free_msgmaps = msgmaps[next].next;
 669                 nfree_msgmaps--;
 670                 msgmaps[next].next = msghdr->msg_spot;
 671                 msghdr->msg_spot = next;
 672                 segs_needed--;
 673         }
 674
 675         /*
 676          * Copy in the message type
 677          */
 678
 679         if ((eval = copyin(user_msgp, &msghdr->msg_type,
 680             sizeof(msghdr->msg_type))) != 0) {
 681 #ifdef MSG_DEBUG_OK
 682                 kprintf("error %d copying the message type\n", eval);
 683 #endif
 684                 msg_freehdr(msghdr);
 685                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 686                 wakeup((caddr_t)msqptr);
 687                 goto done;
 688         }
 689         user_msgp = (const char *)user_msgp + sizeof(msghdr->msg_type);
 690
 691         /*
 692          * Validate the message type
 693          */
 694
 695         if (msghdr->msg_type < 1) {
 696                 msg_freehdr(msghdr);
 697                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 698                 wakeup((caddr_t)msqptr);
 699 #ifdef MSG_DEBUG_OK
 700                 kprintf("mtype (%d) < 1\n", msghdr->msg_type);
 701 #endif
 702                 eval = EINVAL;
 703                 goto done;
 704         }
 705
 706         /*
 707          * Copy in the message body
 708          */
 709
 710         next = msghdr->msg_spot;
 711         while (msgsz > 0) {
 712                 size_t tlen;
 713                 if (msgsz > msginfo.msgssz)
 714                         tlen = msginfo.msgssz;
 715                 else
 716                         tlen = msgsz;
 717                 if (next <= -1)
 718                         panic("next too low #2");
 719                 if (next >= msginfo.msgseg)
 720                         panic("next out of range #2");
 721                 if ((eval = copyin(user_msgp, &msgpool[next * msginfo.msgssz],
 722                     tlen)) != 0) {
 723 #ifdef MSG_DEBUG_OK
 724                         kprintf("error %d copying in message segment\n", eval);
 725 #endif
 726                         msg_freehdr(msghdr);
 727                         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 728                         wakeup((caddr_t)msqptr);
 729                         goto done;
 730                 }
 731                 msgsz -= tlen;
 732                 user_msgp = (const char *)user_msgp + tlen;
 733                 next = msgmaps[next].next;
 734         }
 735         if (next != -1)
 736                 panic("didn't use all the msg segments");
 737
 738         /*
 739          * We've got the message.  Unlock the msqid_ds.
 740          */
 741
 742         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 743
 744         /*
 745          * Make sure that the msqid_ds is still allocated.
 746          */
 747
 748         if (msqptr->msg_qbytes == 0) {
 749                 msg_freehdr(msghdr);
 750                 wakeup((caddr_t)msqptr);
 751                 eval = EIDRM;
 752                 goto done;
 753         }
 754
 755         /*
 756          * Put the message into the queue
 757          */
 758
 759         if (msqptr->msg_first == NULL) {
 760                 msqptr->msg_first = msghdr;
 761                 msqptr->msg_last = msghdr;
 762         } else {
 763                 msqptr->msg_last->msg_next = msghdr;
 764                 msqptr->msg_last = msghdr;
 765         }
 766         msqptr->msg_last->msg_next = NULL;
 767
 768         msqptr->msg_cbytes += msghdr->msg_ts;
 769         msqptr->msg_qnum++;
 770         msqptr->msg_lspid = td->td_proc->p_pid;
 771         msqptr->msg_stime = time_second;
 772
 773         wakeup((caddr_t)msqptr);
 774         eval = 0;
 775 done:
 776         rel_mplock();
 777         if (eval == 0)
 778                 uap->sysmsg_result = 0;
 779         return (eval);
 780 }
 781
 782 /*
 783  * MPALMOSTSAFE
 784  */
 785 int
 786 sys_msgrcv(struct msgrcv_args *uap)
 787 {
 788         struct thread *td = curthread;
 789         int msqid = uap->msqid;
 790         void *user_msgp = uap->msgp;
 791         size_t msgsz = uap->msgsz;
 792         long msgtyp = uap->msgtyp;
 793         int msgflg = uap->msgflg;
 794         size_t len;
 795         struct msqid_ds *msqptr;
 796         struct msg *msghdr;
 797         int eval;
 798         short next;
 799
 800 #ifdef MSG_DEBUG_OK
 801         kprintf("call to msgrcv(%d, 0x%x, %d, %ld, %d)\n", msqid, user_msgp,
 802             msgsz, msgtyp, msgflg);
 803 #endif
 804
 805         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 806                 return (ENOSYS);
 807
 808         get_mplock();
 809         msqid = IPCID_TO_IX(msqid);
 810
 811         if (msqid < 0 || msqid >= msginfo.msgmni) {
 812 #ifdef MSG_DEBUG_OK
 813                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 814                     msginfo.msgmni);
 815 #endif
 816                 eval = EINVAL;
 817                 goto done;
 818         }
 819
 820         msqptr = &msqids[msqid];
 821         if (msqptr->msg_qbytes == 0) {
 822 #ifdef MSG_DEBUG_OK
 823                 kprintf("no such message queue id\n");
 824 #endif
 825                 eval = EINVAL;
 826                 goto done;
 827         }
 828         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 829 #ifdef MSG_DEBUG_OK
 830                 kprintf("wrong sequence number\n");
 831 #endif
 832                 eval = EINVAL;
 833                 goto done;
 834         }
 835
 836         if ((eval = ipcperm(td->td_proc, &msqptr->msg_perm, IPC_R))) {
 837 #ifdef MSG_DEBUG_OK
 838                 kprintf("requester doesn't have read access\n");
 839 #endif
 840                 goto done;
 841         }
 842
 843         msghdr = NULL;
 844         while (msghdr == NULL) {
 845                 if (msgtyp == 0) {
 846                         msghdr = msqptr->msg_first;
 847                         if (msghdr != NULL) {
 848                                 if (msgsz < msghdr->msg_ts &&
 849                                     (msgflg & MSG_NOERROR) == 0) {
 850 #ifdef MSG_DEBUG_OK
 851                                         kprintf("first message on the queue is too big (want %d, got %d)\n",
 852                                             msgsz, msghdr->msg_ts);
 853 #endif
 854                                         eval = E2BIG;
 855                                         goto done;
 856                                 }
 857                                 if (msqptr->msg_first == msqptr->msg_last) {
 858                                         msqptr->msg_first = NULL;
 859                                         msqptr->msg_last = NULL;
 860                                 } else {
 861                                         msqptr->msg_first = msghdr->msg_next;
 862                                         if (msqptr->msg_first == NULL)
 863                                                 panic("msg_first/last screwed up #1");
 864                                 }
 865                         }
 866                 } else {
 867                         struct msg *previous;
 868                         struct msg **prev;
 869
 870                         previous = NULL;
 871                         prev = &(msqptr->msg_first);
 872                         while ((msghdr = *prev) != NULL) {
 873                                 /*
 874                                  * Is this message's type an exact match or is
 875                                  * this message's type less than or equal to
 876                                  * the absolute value of a negative msgtyp?
 877                                  * Note that the second half of this test can
 878                                  * NEVER be true if msgtyp is positive since
 879                                  * msg_type is always positive!
 880                                  */
 881
 882                                 if (msgtyp == msghdr->msg_type ||
 883                                     msghdr->msg_type <= -msgtyp) {
 884 #ifdef MSG_DEBUG_OK
 885                                         kprintf("found message type %d, requested %d\n",
 886                                             msghdr->msg_type, msgtyp);
 887 #endif
 888                                         if (msgsz < msghdr->msg_ts &&
 889                                             (msgflg & MSG_NOERROR) == 0) {
 890 #ifdef MSG_DEBUG_OK
 891                                                 kprintf("requested message on the queue is too big (want %d, got %d)\n",
 892                                                     msgsz, msghdr->msg_ts);
 893 #endif
 894                                                 eval = E2BIG;
 895                                                 goto done;
 896                                         }
 897                                         *prev = msghdr->msg_next;
 898                                         if (msghdr == msqptr->msg_last) {
 899                                                 if (previous == NULL) {
 900                                                         if (prev !=
 901                                                             &msqptr->msg_first)
 902                                                                 panic("msg_first/last screwed up #2");
 903                                                         msqptr->msg_first =
 904                                                             NULL;
 905                                                         msqptr->msg_last =
 906                                                             NULL;
 907                                                 } else {
 908                                                         if (prev ==
 909                                                             &msqptr->msg_first)
 910                                                                 panic("msg_first/last screwed up #3");
 911                                                         msqptr->msg_last =
 912                                                             previous;
 913                                                 }
 914                                         }
 915                                         break;
 916                                 }
 917                                 previous = msghdr;
 918                                 prev = &(msghdr->msg_next);
 919                         }
 920                 }
 921
 922                 /*
 923                  * We've either extracted the msghdr for the appropriate
 924                  * message or there isn't one.
 925                  * If there is one then bail out of this loop.
 926                  */
 927
 928                 if (msghdr != NULL)
 929                         break;
 930
 931                 /*
 932                  * Hmph!  No message found.  Does the user want to wait?
 933                  */
 934
 935                 if ((msgflg & IPC_NOWAIT) != 0) {
 936 #ifdef MSG_DEBUG_OK
 937                         kprintf("no appropriate message found (msgtyp=%d)\n",
 938                             msgtyp);
 939 #endif
 940                         /* The SVID says to return ENOMSG. */
 941 #ifdef ENOMSG
 942                         eval = ENOMSG;
 943 #else
 944                         /* Unfortunately, BSD doesn't define that code yet! */
 945                         eval = EAGAIN;
 946 #endif
 947                         goto done;
 948                 }
 949
 950                 /*
 951                  * Wait for something to happen
 952                  */
 953
 954 #ifdef MSG_DEBUG_OK
 955                 kprintf("msgrcv:  goodnight\n");
 956 #endif
 957                 eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 958 #ifdef MSG_DEBUG_OK
 959                 kprintf("msgrcv:  good morning (eval=%d)\n", eval);
 960 #endif
 961
 962                 if (eval != 0) {
 963 #ifdef MSG_DEBUG_OK
 964                         kprintf("msgsnd:  interrupted system call\n");
 965 #endif
 966                         eval = EINTR;
 967                         goto done;
 968                 }
 969
 970                 /*
 971                  * Make sure that the msq queue still exists
 972                  */
 973
 974                 if (msqptr->msg_qbytes == 0 ||
 975                     msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 976 #ifdef MSG_DEBUG_OK
 977                         kprintf("msqid deleted\n");
 978 #endif
 979                         eval = EIDRM;
 980                         goto done;
 981                 }
 982         }
 983
 984         /*
 985          * Return the message to the user.
 986          *
 987          * First, do the bookkeeping (before we risk being interrupted).
 988          */
 989
 990         msqptr->msg_cbytes -= msghdr->msg_ts;
 991         msqptr->msg_qnum--;
 992         msqptr->msg_lrpid = td->td_proc->p_pid;
 993         msqptr->msg_rtime = time_second;
 994
 995         /*
 996          * Make msgsz the actual amount that we'll be returning.
 997          * Note that this effectively truncates the message if it is too long
 998          * (since msgsz is never increased).
 999          */
1000
1001 #ifdef MSG_DEBUG_OK
1002         kprintf("found a message, msgsz=%d, msg_ts=%d\n", msgsz,
1003             msghdr->msg_ts);
1004 #endif
1005         if (msgsz > msghdr->msg_ts)
1006                 msgsz = msghdr->msg_ts;
1007
1008         /*
1009          * Return the type to the user.
1010          */
1011
1012         eval = copyout((caddr_t)&(msghdr->msg_type), user_msgp,
1013             sizeof(msghdr->msg_type));
1014         if (eval != 0) {
1015 #ifdef MSG_DEBUG_OK
1016                 kprintf("error (%d) copying out message type\n", eval);
1017 #endif
1018                 msg_freehdr(msghdr);
1019                 wakeup((caddr_t)msqptr);
1020                 goto done;
1021         }
1022         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
1023
1024         /*
1025          * Return the segments to the user
1026          */
1027
1028         next = msghdr->msg_spot;
1029         for (len = 0; len < msgsz; len += msginfo.msgssz) {
1030                 size_t tlen;
1031
1032                 if (msgsz - len > msginfo.msgssz)
1033                         tlen = msginfo.msgssz;
1034                 else
1035                         tlen = msgsz - len;
1036                 if (next <= -1)
1037                         panic("next too low #3");
1038                 if (next >= msginfo.msgseg)
1039                         panic("next out of range #3");
1040                 eval = copyout((caddr_t)&msgpool[next * msginfo.msgssz],
1041                     user_msgp, tlen);
1042                 if (eval != 0) {
1043 #ifdef MSG_DEBUG_OK
1044                         kprintf("error (%d) copying out message segment\n",
1045                             eval);
1046 #endif
1047                         msg_freehdr(msghdr);
1048                         wakeup((caddr_t)msqptr);
1049                         goto done;
1050                 }
1051                 user_msgp = (char *)user_msgp + tlen;
1052                 next = msgmaps[next].next;
1053         }
1054
1055         /*
1056          * Done, return the actual number of bytes copied out.
1057          */
1058
1059         msg_freehdr(msghdr);
1060         wakeup((caddr_t)msqptr);
1061         eval = 0;
1062 done:
1063         rel_mplock();
1064         if (eval == 0)
1065                 uap->sysmsg_result = msgsz;
1066         return(eval);
1067 }
1068
1069 static int
1070 sysctl_msqids(SYSCTL_HANDLER_ARGS)
1071 {
1072
1073         return (SYSCTL_OUT(req, msqids,
1074             sizeof(struct msqid_ds) * msginfo.msgmni));
1075 }
1076
1077 TUNABLE_INT("kern.ipc.msgseg", &msginfo.msgseg);
1078 TUNABLE_INT("kern.ipc.msgssz", &msginfo.msgssz);
1079 TUNABLE_INT("kern.ipc.msgmni", &msginfo.msgmni);
1080
1081 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmax, CTLFLAG_RD, &msginfo.msgmax, 0,
1082     "Max characters in message");
1083 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmni, CTLFLAG_RD, &msginfo.msgmni, 0,
1084     "Max message queue identifiers");
1085 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmnb, CTLFLAG_RD, &msginfo.msgmnb, 0,
1086     "Max characters in message queue");
1087 SYSCTL_INT(_kern_ipc, OID_AUTO, msgtql, CTLFLAG_RD, &msginfo.msgtql, 0,
1088     "Max SVID messages in system");
1089 SYSCTL_INT(_kern_ipc, OID_AUTO, msgssz, CTLFLAG_RD, &msginfo.msgssz, 0,
1090     "Power-of-two size of a message segment");
1091 SYSCTL_INT(_kern_ipc, OID_AUTO, msgseg, CTLFLAG_RD, &msginfo.msgseg, 0,
1092     "Number of message segments");
1093 SYSCTL_PROC(_kern_ipc, OID_AUTO, msqids, CTLFLAG_RD,
1094     NULL, 0, sysctl_msqids, "", "Message queue IDs");