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 /* XXX casting to (sy_call_t *) is bogus, as usual. */
  48 static sy_call_t *msgcalls[] = {
  49         (sy_call_t *)sys_msgctl, (sy_call_t *)sys_msgget,
  50         (sy_call_t *)sys_msgsnd, (sy_call_t *)sys_msgrcv
  51 };
  52
  53 struct msg {
  54         struct  msg *msg_next;  /* next msg in the chain */
  55         long    msg_type;       /* type of this message */
  56                                 /* >0 -> type of this message */
  57                                 /* 0 -> free header */
  58         u_short msg_ts;         /* size of this message */
  59         short   msg_spot;       /* location of start of msg in buffer */
  60 };
  61
  62
  63 #ifndef MSGSSZ
  64 #define MSGSSZ  8               /* Each segment must be 2^N long */
  65 #endif
  66 #ifndef MSGSEG
  67 #define MSGSEG  2048            /* must be less than 32767 */
  68 #endif
  69 #define MSGMAX  (MSGSSZ*MSGSEG)
  70 #ifndef MSGMNB
  71 #define MSGMNB  2048            /* max # of bytes in a queue */
  72 #endif
  73 #ifndef MSGMNI
  74 #define MSGMNI  40
  75 #endif
  76 #ifndef MSGTQL
  77 #define MSGTQL  40
  78 #endif
  79
  80 /*
  81  * Based on the configuration parameters described in an SVR2 (yes, two)
  82  * config(1m) man page.
  83  *
  84  * Each message is broken up and stored in segments that are msgssz bytes
  85  * long.  For efficiency reasons, this should be a power of two.  Also,
  86  * it doesn't make sense if it is less than 8 or greater than about 256.
  87  * Consequently, msginit in kern/sysv_msg.c checks that msgssz is a power of
  88  * two between 8 and 1024 inclusive (and panic's if it isn't).
  89  */
  90 struct msginfo msginfo = {
  91                 MSGMAX,         /* max chars in a message */
  92                 MSGMNI,         /* # of message queue identifiers */
  93                 MSGMNB,         /* max chars in a queue */
  94                 MSGTQL,         /* max messages in system */
  95                 MSGSSZ,         /* size of a message segment */
  96                                 /* (must be small power of 2 greater than 4) */
  97                 MSGSEG          /* number of message segments */
  98 };
  99
 100 /*
 101  * macros to convert between msqid_ds's and msqid's.
 102  * (specific to this implementation)
 103  */
 104 #define MSQID(ix,ds)    ((ix) & 0xffff | (((ds).msg_perm.seq << 16) & 0xffff0000))
 105 #define MSQID_IX(id)    ((id) & 0xffff)
 106 #define MSQID_SEQ(id)   (((id) >> 16) & 0xffff)
 107
 108 /*
 109  * The rest of this file is specific to this particular implementation.
 110  */
 111
 112 struct msgmap {
 113         short   next;           /* next segment in buffer */
 114                                 /* -1 -> available */
 115                                 /* 0..(MSGSEG-1) -> index of next segment */
 116 };
 117
 118 #define MSG_LOCKED      01000   /* Is this msqid_ds locked? */
 119
 120 static int nfree_msgmaps;       /* # of free map entries */
 121 static short free_msgmaps;      /* head of linked list of free map entries */
 122 static struct msg *free_msghdrs;/* list of free msg headers */
 123 static char *msgpool;           /* MSGMAX byte long msg buffer pool */
 124 static struct msgmap *msgmaps;  /* MSGSEG msgmap structures */
 125 static struct msg *msghdrs;     /* MSGTQL msg headers */
 126 static struct msqid_ds *msqids; /* MSGMNI msqid_ds struct's */
 127
 128 static void
 129 msginit(void *dummy)
 130 {
 131         int i;
 132
 133         msginfo.msgmax = msginfo.msgseg * msginfo.msgssz;
 134         msgpool = kmalloc(msginfo.msgmax, M_MSG, M_WAITOK);
 135         msgmaps = kmalloc(sizeof(struct msgmap) * msginfo.msgseg, M_MSG, M_WAITOK);
 136         msghdrs = kmalloc(sizeof(struct msg) * msginfo.msgtql, M_MSG, M_WAITOK);
 137         msqids = kmalloc(sizeof(struct msqid_ds) * msginfo.msgmni, M_MSG, M_WAITOK);
 138
 139         /*
 140          * msginfo.msgssz should be a power of two for efficiency reasons.
 141          * It is also pretty silly if msginfo.msgssz is less than 8
 142          * or greater than about 256 so ...
 143          */
 144
 145         i = 8;
 146         while (i < 1024 && i != msginfo.msgssz)
 147                 i <<= 1;
 148         if (i != msginfo.msgssz) {
 149                 kprintf("msginfo.msgssz=%d (0x%x)\n", msginfo.msgssz,
 150                     msginfo.msgssz);
 151                 panic("msginfo.msgssz not a small power of 2");
 152         }
 153
 154         if (msginfo.msgseg > 32767) {
 155                 kprintf("msginfo.msgseg=%d\n", msginfo.msgseg);
 156                 panic("msginfo.msgseg > 32767");
 157         }
 158
 159         if (msgmaps == NULL)
 160                 panic("msgmaps is NULL");
 161
 162         for (i = 0; i < msginfo.msgseg; i++) {
 163                 if (i > 0)
 164                         msgmaps[i-1].next = i;
 165                 msgmaps[i].next = -1;   /* implies entry is available */
 166         }
 167         free_msgmaps = 0;
 168         nfree_msgmaps = msginfo.msgseg;
 169
 170         if (msghdrs == NULL)
 171                 panic("msghdrs is NULL");
 172
 173         for (i = 0; i < msginfo.msgtql; i++) {
 174                 msghdrs[i].msg_type = 0;
 175                 if (i > 0)
 176                         msghdrs[i-1].msg_next = &msghdrs[i];
 177                 msghdrs[i].msg_next = NULL;
 178         }
 179         free_msghdrs = &msghdrs[0];
 180
 181         if (msqids == NULL)
 182                 panic("msqids is NULL");
 183
 184         for (i = 0; i < msginfo.msgmni; i++) {
 185                 msqids[i].msg_qbytes = 0;       /* implies entry is available */
 186                 msqids[i].msg_perm.seq = 0;     /* reset to a known value */
 187                 msqids[i].msg_perm.mode = 0;
 188         }
 189 }
 190 SYSINIT(sysv_msg, SI_SUB_SYSV_MSG, SI_ORDER_FIRST, msginit, NULL)
 191
 192 /*
 193  * Entry point for all MSG calls
 194  *
 195  * msgsys_args(int which, int a2, ...) (VARARGS)
 196  *
 197  * MPALMOSTSAFE
 198  */
 199 int
 200 sys_msgsys(struct msgsys_args *uap)
 201 {
 202         struct thread *td = curthread;
 203         unsigned int which = (unsigned int)uap->which;
 204         int error;
 205
 206         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 207                 return (ENOSYS);
 208
 209         if (which >= NELEM(msgcalls))
 210                 return (EINVAL);
 211         bcopy(&uap->a2, &uap->which,
 212               sizeof(struct msgsys_args) - offsetof(struct msgsys_args, a2));
 213         get_mplock();
 214         error = (*msgcalls[which])(uap);
 215         rel_mplock();
 216         return (error);
 217 }
 218
 219 static void
 220 msg_freehdr(struct msg *msghdr)
 221 {
 222         while (msghdr->msg_ts > 0) {
 223                 short next;
 224                 if (msghdr->msg_spot < 0 || msghdr->msg_spot >= msginfo.msgseg)
 225                         panic("msghdr->msg_spot out of range");
 226                 next = msgmaps[msghdr->msg_spot].next;
 227                 msgmaps[msghdr->msg_spot].next = free_msgmaps;
 228                 free_msgmaps = msghdr->msg_spot;
 229                 nfree_msgmaps++;
 230                 msghdr->msg_spot = next;
 231                 if (msghdr->msg_ts >= msginfo.msgssz)
 232                         msghdr->msg_ts -= msginfo.msgssz;
 233                 else
 234                         msghdr->msg_ts = 0;
 235         }
 236         if (msghdr->msg_spot != -1)
 237                 panic("msghdr->msg_spot != -1");
 238         msghdr->msg_next = free_msghdrs;
 239         free_msghdrs = msghdr;
 240 }
 241
 242 /*
 243  * MPALMOSTSAFE
 244  */
 245 int
 246 sys_msgctl(struct msgctl_args *uap)
 247 {
 248         struct thread *td = curthread;
 249         struct proc *p = td->td_proc;
 250         int msqid = uap->msqid;
 251         int cmd = uap->cmd;
 252         struct msqid_ds *user_msqptr = uap->buf;
 253         int rval, eval;
 254         struct msqid_ds msqbuf;
 255         struct msqid_ds *msqptr;
 256
 257 #ifdef MSG_DEBUG_OK
 258         kprintf("call to msgctl(%d, %d, 0x%x)\n", msqid, cmd, user_msqptr);
 259 #endif
 260
 261         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 262                 return (ENOSYS);
 263
 264         get_mplock();
 265         msqid = IPCID_TO_IX(msqid);
 266
 267         if (msqid < 0 || msqid >= msginfo.msgmni) {
 268 #ifdef MSG_DEBUG_OK
 269                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 270                     msginfo.msgmni);
 271 #endif
 272                 eval = EINVAL;
 273                 goto done;
 274         }
 275
 276         msqptr = &msqids[msqid];
 277
 278         if (msqptr->msg_qbytes == 0) {
 279 #ifdef MSG_DEBUG_OK
 280                 kprintf("no such msqid\n");
 281 #endif
 282                 eval = EINVAL;
 283                 goto done;
 284         }
 285         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 286 #ifdef MSG_DEBUG_OK
 287                 kprintf("wrong sequence number\n");
 288 #endif
 289                 eval = EINVAL;
 290                 goto done;
 291         }
 292
 293         rval = 0;
 294
 295         switch (cmd) {
 296         case IPC_RMID:
 297         {
 298                 struct msg *msghdr;
 299                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)) != 0)
 300                         break;
 301                 /* Free the message headers */
 302                 msghdr = msqptr->msg_first;
 303                 while (msghdr != NULL) {
 304                         struct msg *msghdr_tmp;
 305
 306                         /* Free the segments of each message */
 307                         msqptr->msg_cbytes -= msghdr->msg_ts;
 308                         msqptr->msg_qnum--;
 309                         msghdr_tmp = msghdr;
 310                         msghdr = msghdr->msg_next;
 311                         msg_freehdr(msghdr_tmp);
 312                 }
 313
 314                 if (msqptr->msg_cbytes != 0)
 315                         panic("msg_cbytes is screwed up");
 316                 if (msqptr->msg_qnum != 0)
 317                         panic("msg_qnum is screwed up");
 318
 319                 msqptr->msg_qbytes = 0; /* Mark it as free */
 320
 321                 wakeup((caddr_t)msqptr);
 322         }
 323
 324                 break;
 325
 326         case IPC_SET:
 327                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)) != 0)
 328                         break;
 329                 if ((eval = copyin(user_msqptr, &msqbuf, sizeof(msqbuf))) != 0)
 330                         break;
 331                 if (msqbuf.msg_qbytes > msqptr->msg_qbytes) {
 332                         eval = priv_check(td, PRIV_ROOT);
 333                         if (eval)
 334                                 break;
 335                 }
 336                 if (msqbuf.msg_qbytes > msginfo.msgmnb) {
 337 #ifdef MSG_DEBUG_OK
 338                         kprintf("can't increase msg_qbytes beyond %d (truncating)\n",
 339                             msginfo.msgmnb);
 340 #endif
 341                         msqbuf.msg_qbytes = msginfo.msgmnb;     /* silently restrict qbytes to system limit */
 342                 }
 343                 if (msqbuf.msg_qbytes == 0) {
 344 #ifdef MSG_DEBUG_OK
 345                         kprintf("can't reduce msg_qbytes to 0\n");
 346 #endif
 347                         eval = EINVAL;          /* non-standard errno! */
 348                         break;
 349                 }
 350                 msqptr->msg_perm.uid = msqbuf.msg_perm.uid;     /* change the owner */
 351                 msqptr->msg_perm.gid = msqbuf.msg_perm.gid;     /* change the owner */
 352                 msqptr->msg_perm.mode = (msqptr->msg_perm.mode & ~0777) |
 353                                         (msqbuf.msg_perm.mode & 0777);
 354                 msqptr->msg_qbytes = msqbuf.msg_qbytes;
 355                 msqptr->msg_ctime = time_second;
 356                 break;
 357
 358         case IPC_STAT:
 359                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
 360 #ifdef MSG_DEBUG_OK
 361                         kprintf("requester doesn't have read access\n");
 362 #endif
 363                         eval = EINVAL;
 364                         break;
 365                 }
 366                 eval = copyout(msqptr, user_msqptr, sizeof(struct msqid_ds));
 367                 break;
 368
 369         default:
 370 #ifdef MSG_DEBUG_OK
 371                 kprintf("invalid command %d\n", cmd);
 372 #endif
 373                 eval = EINVAL;
 374                 break;
 375         }
 376 done:
 377         rel_mplock();
 378         if (eval == 0)
 379                 uap->sysmsg_result = rval;
 380         return(eval);
 381 }
 382
 383 /*
 384  * MPALMOSTSAFE
 385  */
 386 int
 387 sys_msgget(struct msgget_args *uap)
 388 {
 389         struct thread *td = curthread;
 390         int msqid, eval;
 391         int key = uap->key;
 392         int msgflg = uap->msgflg;
 393         struct ucred *cred = td->td_ucred;
 394         struct msqid_ds *msqptr = NULL;
 395
 396 #ifdef MSG_DEBUG_OK
 397         kprintf("msgget(0x%x, 0%o)\n", key, msgflg);
 398 #endif
 399         if (!jail_sysvipc_allowed && cred->cr_prison != NULL)
 400                 return (ENOSYS);
 401
 402         eval = 0;
 403         get_mplock();
 404
 405         if (key != IPC_PRIVATE) {
 406                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 407                         msqptr = &msqids[msqid];
 408                         if (msqptr->msg_qbytes != 0 &&
 409                             msqptr->msg_perm.key == key)
 410                                 break;
 411                 }
 412                 if (msqid < msginfo.msgmni) {
 413 #ifdef MSG_DEBUG_OK
 414                         kprintf("found public key\n");
 415 #endif
 416                         if ((msgflg & IPC_CREAT) && (msgflg & IPC_EXCL)) {
 417 #ifdef MSG_DEBUG_OK
 418                                 kprintf("not exclusive\n");
 419 #endif
 420                                 eval = EEXIST;
 421                                 goto done;
 422                         }
 423                         if ((eval = ipcperm(td->td_proc, &msqptr->msg_perm, msgflg & 0700 ))) {
 424 #ifdef MSG_DEBUG_OK
 425                                 kprintf("requester doesn't have 0%o access\n",
 426                                     msgflg & 0700);
 427 #endif
 428                                 goto done;
 429                         }
 430                         goto done;
 431                 }
 432         }
 433
 434 #ifdef MSG_DEBUG_OK
 435         kprintf("need to allocate the msqid_ds\n");
 436 #endif
 437         if (key == IPC_PRIVATE || (msgflg & IPC_CREAT)) {
 438                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 439                         /*
 440                          * Look for an unallocated and unlocked msqid_ds.
 441                          * msqid_ds's can be locked by msgsnd or msgrcv while
 442                          * they are copying the message in/out.  We can't
 443                          * re-use the entry until they release it.
 444                          */
 445                         msqptr = &msqids[msqid];
 446                         if (msqptr->msg_qbytes == 0 &&
 447                             (msqptr->msg_perm.mode & MSG_LOCKED) == 0)
 448                                 break;
 449                 }
 450                 if (msqid == msginfo.msgmni) {
 451 #ifdef MSG_DEBUG_OK
 452                         kprintf("no more msqid_ds's available\n");
 453 #endif
 454                         eval = ENOSPC;
 455                         goto done;
 456                 }
 457 #ifdef MSG_DEBUG_OK
 458                 kprintf("msqid %d is available\n", msqid);
 459 #endif
 460                 msqptr->msg_perm.key = key;
 461                 msqptr->msg_perm.cuid = cred->cr_uid;
 462                 msqptr->msg_perm.uid = cred->cr_uid;
 463                 msqptr->msg_perm.cgid = cred->cr_gid;
 464                 msqptr->msg_perm.gid = cred->cr_gid;
 465                 msqptr->msg_perm.mode = (msgflg & 0777);
 466                 /* Make sure that the returned msqid is unique */
 467                 msqptr->msg_perm.seq = (msqptr->msg_perm.seq + 1) & 0x7fff;
 468                 msqptr->msg_first = NULL;
 469                 msqptr->msg_last = NULL;
 470                 msqptr->msg_cbytes = 0;
 471                 msqptr->msg_qnum = 0;
 472                 msqptr->msg_qbytes = msginfo.msgmnb;
 473                 msqptr->msg_lspid = 0;
 474                 msqptr->msg_lrpid = 0;
 475                 msqptr->msg_stime = 0;
 476                 msqptr->msg_rtime = 0;
 477                 msqptr->msg_ctime = time_second;
 478         } else {
 479 #ifdef MSG_DEBUG_OK
 480                 kprintf("didn't find it and wasn't asked to create it\n");
 481 #endif
 482                 eval = ENOENT;
 483         }
 484
 485 done:
 486         rel_mplock();
 487         /* Construct the unique msqid */
 488         if (eval == 0)
 489                 uap->sysmsg_result = IXSEQ_TO_IPCID(msqid, msqptr->msg_perm);
 490         return(eval);
 491 }
 492
 493 /*
 494  * MPALMOSTSAFE
 495  */
 496 int
 497 sys_msgsnd(struct msgsnd_args *uap)
 498 {
 499         struct thread *td = curthread;
 500         int msqid = uap->msqid;
 501         void *user_msgp = uap->msgp;
 502         size_t msgsz = uap->msgsz;
 503         int msgflg = uap->msgflg;
 504         int segs_needed, eval;
 505         struct msqid_ds *msqptr;
 506         struct msg *msghdr;
 507         short next;
 508
 509 #ifdef MSG_DEBUG_OK
 510         kprintf("call to msgsnd(%d, 0x%x, %d, %d)\n", msqid, user_msgp, msgsz,
 511             msgflg);
 512 #endif
 513
 514         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 515                 return (ENOSYS);
 516
 517         get_mplock();
 518         msqid = IPCID_TO_IX(msqid);
 519
 520         if (msqid < 0 || msqid >= msginfo.msgmni) {
 521 #ifdef MSG_DEBUG_OK
 522                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 523                     msginfo.msgmni);
 524 #endif
 525                 eval = EINVAL;
 526                 goto done;
 527         }
 528
 529         msqptr = &msqids[msqid];
 530         if (msqptr->msg_qbytes == 0) {
 531 #ifdef MSG_DEBUG_OK
 532                 kprintf("no such message queue id\n");
 533 #endif
 534                 eval = EINVAL;
 535                 goto done;
 536         }
 537         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 538 #ifdef MSG_DEBUG_OK
 539                 kprintf("wrong sequence number\n");
 540 #endif
 541                 eval = EINVAL;
 542                 goto done;
 543         }
 544
 545         if ((eval = ipcperm(td->td_proc, &msqptr->msg_perm, IPC_W))) {
 546 #ifdef MSG_DEBUG_OK
 547                 kprintf("requester doesn't have write access\n");
 548 #endif
 549                 eval = eval;
 550                 goto done;
 551         }
 552
 553         segs_needed = (msgsz + msginfo.msgssz - 1) / msginfo.msgssz;
 554 #ifdef MSG_DEBUG_OK
 555         kprintf("msgsz=%d, msgssz=%d, segs_needed=%d\n", msgsz, msginfo.msgssz,
 556             segs_needed);
 557 #endif
 558         for (;;) {
 559                 int need_more_resources = 0;
 560
 561                 /*
 562                  * check msgsz
 563                  * (inside this loop in case msg_qbytes changes while we sleep)
 564                  */
 565
 566                 if (msgsz > msqptr->msg_qbytes) {
 567 #ifdef MSG_DEBUG_OK
 568                         kprintf("msgsz > msqptr->msg_qbytes\n");
 569 #endif
 570                         eval = EINVAL;
 571                         goto done;
 572                 }
 573
 574                 if (msqptr->msg_perm.mode & MSG_LOCKED) {
 575 #ifdef MSG_DEBUG_OK
 576                         kprintf("msqid is locked\n");
 577 #endif
 578                         need_more_resources = 1;
 579                 }
 580                 if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes) {
 581 #ifdef MSG_DEBUG_OK
 582                         kprintf("msgsz + msg_cbytes > msg_qbytes\n");
 583 #endif
 584                         need_more_resources = 1;
 585                 }
 586                 if (segs_needed > nfree_msgmaps) {
 587 #ifdef MSG_DEBUG_OK
 588                         kprintf("segs_needed > nfree_msgmaps\n");
 589 #endif
 590                         need_more_resources = 1;
 591                 }
 592                 if (free_msghdrs == NULL) {
 593 #ifdef MSG_DEBUG_OK
 594                         kprintf("no more msghdrs\n");
 595 #endif
 596                         need_more_resources = 1;
 597                 }
 598
 599                 if (need_more_resources) {
 600                         int we_own_it;
 601
 602                         if ((msgflg & IPC_NOWAIT) != 0) {
 603 #ifdef MSG_DEBUG_OK
 604                                 kprintf("need more resources but caller doesn't want to wait\n");
 605 #endif
 606                                 eval = EAGAIN;
 607                                 goto done;
 608                         }
 609
 610                         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0) {
 611 #ifdef MSG_DEBUG_OK
 612                                 kprintf("we don't own the msqid_ds\n");
 613 #endif
 614                                 we_own_it = 0;
 615                         } else {
 616                                 /* Force later arrivals to wait for our
 617                                    request */
 618 #ifdef MSG_DEBUG_OK
 619                                 kprintf("we own the msqid_ds\n");
 620 #endif
 621                                 msqptr->msg_perm.mode |= MSG_LOCKED;
 622                                 we_own_it = 1;
 623                         }
 624 #ifdef MSG_DEBUG_OK
 625                         kprintf("goodnight\n");
 626 #endif
 627                         eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 628 #ifdef MSG_DEBUG_OK
 629                         kprintf("good morning, eval=%d\n", eval);
 630 #endif
 631                         if (we_own_it)
 632                                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 633                         if (eval != 0) {
 634 #ifdef MSG_DEBUG_OK
 635                                 kprintf("msgsnd:  interrupted system call\n");
 636 #endif
 637                                 eval = EINTR;
 638                                 goto done;
 639                         }
 640
 641                         /*
 642                          * Make sure that the msq queue still exists
 643                          */
 644
 645                         if (msqptr->msg_qbytes == 0) {
 646 #ifdef MSG_DEBUG_OK
 647                                 kprintf("msqid deleted\n");
 648 #endif
 649                                 eval = EIDRM;
 650                                 goto done;
 651                         }
 652
 653                 } else {
 654 #ifdef MSG_DEBUG_OK
 655                         kprintf("got all the resources that we need\n");
 656 #endif
 657                         break;
 658                 }
 659         }
 660
 661         /*
 662          * We have the resources that we need.
 663          * Make sure!
 664          */
 665
 666         if (msqptr->msg_perm.mode & MSG_LOCKED)
 667                 panic("msg_perm.mode & MSG_LOCKED");
 668         if (segs_needed > nfree_msgmaps)
 669                 panic("segs_needed > nfree_msgmaps");
 670         if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes)
 671                 panic("msgsz + msg_cbytes > msg_qbytes");
 672         if (free_msghdrs == NULL)
 673                 panic("no more msghdrs");
 674
 675         /*
 676          * Re-lock the msqid_ds in case we page-fault when copying in the
 677          * message
 678          */
 679
 680         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0)
 681                 panic("msqid_ds is already locked");
 682         msqptr->msg_perm.mode |= MSG_LOCKED;
 683
 684         /*
 685          * Allocate a message header
 686          */
 687
 688         msghdr = free_msghdrs;
 689         free_msghdrs = msghdr->msg_next;
 690         msghdr->msg_spot = -1;
 691         msghdr->msg_ts = msgsz;
 692
 693         /*
 694          * Allocate space for the message
 695          */
 696
 697         while (segs_needed > 0) {
 698                 if (nfree_msgmaps <= 0)
 699                         panic("not enough msgmaps");
 700                 if (free_msgmaps == -1)
 701                         panic("nil free_msgmaps");
 702                 next = free_msgmaps;
 703                 if (next <= -1)
 704                         panic("next too low #1");
 705                 if (next >= msginfo.msgseg)
 706                         panic("next out of range #1");
 707 #ifdef MSG_DEBUG_OK
 708                 kprintf("allocating segment %d to message\n", next);
 709 #endif
 710                 free_msgmaps = msgmaps[next].next;
 711                 nfree_msgmaps--;
 712                 msgmaps[next].next = msghdr->msg_spot;
 713                 msghdr->msg_spot = next;
 714                 segs_needed--;
 715         }
 716
 717         /*
 718          * Copy in the message type
 719          */
 720
 721         if ((eval = copyin(user_msgp, &msghdr->msg_type,
 722             sizeof(msghdr->msg_type))) != 0) {
 723 #ifdef MSG_DEBUG_OK
 724                 kprintf("error %d copying the message type\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         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
 732
 733         /*
 734          * Validate the message type
 735          */
 736
 737         if (msghdr->msg_type < 1) {
 738                 msg_freehdr(msghdr);
 739                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 740                 wakeup((caddr_t)msqptr);
 741 #ifdef MSG_DEBUG_OK
 742                 kprintf("mtype (%d) < 1\n", msghdr->msg_type);
 743 #endif
 744                 eval = EINVAL;
 745                 goto done;
 746         }
 747
 748         /*
 749          * Copy in the message body
 750          */
 751
 752         next = msghdr->msg_spot;
 753         while (msgsz > 0) {
 754                 size_t tlen;
 755                 if (msgsz > msginfo.msgssz)
 756                         tlen = msginfo.msgssz;
 757                 else
 758                         tlen = msgsz;
 759                 if (next <= -1)
 760                         panic("next too low #2");
 761                 if (next >= msginfo.msgseg)
 762                         panic("next out of range #2");
 763                 if ((eval = copyin(user_msgp, &msgpool[next * msginfo.msgssz],
 764                     tlen)) != 0) {
 765 #ifdef MSG_DEBUG_OK
 766                         kprintf("error %d copying in message segment\n", eval);
 767 #endif
 768                         msg_freehdr(msghdr);
 769                         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 770                         wakeup((caddr_t)msqptr);
 771                         goto done;
 772                 }
 773                 msgsz -= tlen;
 774                 user_msgp = (char *)user_msgp + tlen;
 775                 next = msgmaps[next].next;
 776         }
 777         if (next != -1)
 778                 panic("didn't use all the msg segments");
 779
 780         /*
 781          * We've got the message.  Unlock the msqid_ds.
 782          */
 783
 784         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 785
 786         /*
 787          * Make sure that the msqid_ds is still allocated.
 788          */
 789
 790         if (msqptr->msg_qbytes == 0) {
 791                 msg_freehdr(msghdr);
 792                 wakeup((caddr_t)msqptr);
 793                 eval = EIDRM;
 794                 goto done;
 795         }
 796
 797         /*
 798          * Put the message into the queue
 799          */
 800
 801         if (msqptr->msg_first == NULL) {
 802                 msqptr->msg_first = msghdr;
 803                 msqptr->msg_last = msghdr;
 804         } else {
 805                 msqptr->msg_last->msg_next = msghdr;
 806                 msqptr->msg_last = msghdr;
 807         }
 808         msqptr->msg_last->msg_next = NULL;
 809
 810         msqptr->msg_cbytes += msghdr->msg_ts;
 811         msqptr->msg_qnum++;
 812         msqptr->msg_lspid = td->td_proc->p_pid;
 813         msqptr->msg_stime = time_second;
 814
 815         wakeup((caddr_t)msqptr);
 816         eval = 0;
 817 done:
 818         rel_mplock();
 819         if (eval == 0)
 820                 uap->sysmsg_result = 0;
 821         return (eval);
 822 }
 823
 824 /*
 825  * MPALMOSTSAFE
 826  */
 827 int
 828 sys_msgrcv(struct msgrcv_args *uap)
 829 {
 830         struct thread *td = curthread;
 831         int msqid = uap->msqid;
 832         void *user_msgp = uap->msgp;
 833         size_t msgsz = uap->msgsz;
 834         long msgtyp = uap->msgtyp;
 835         int msgflg = uap->msgflg;
 836         size_t len;
 837         struct msqid_ds *msqptr;
 838         struct msg *msghdr;
 839         int eval;
 840         short next;
 841
 842 #ifdef MSG_DEBUG_OK
 843         kprintf("call to msgrcv(%d, 0x%x, %d, %ld, %d)\n", msqid, user_msgp,
 844             msgsz, msgtyp, msgflg);
 845 #endif
 846
 847         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 848                 return (ENOSYS);
 849
 850         get_mplock();
 851         msqid = IPCID_TO_IX(msqid);
 852
 853         if (msqid < 0 || msqid >= msginfo.msgmni) {
 854 #ifdef MSG_DEBUG_OK
 855                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 856                     msginfo.msgmni);
 857 #endif
 858                 eval = EINVAL;
 859                 goto done;
 860         }
 861
 862         msqptr = &msqids[msqid];
 863         if (msqptr->msg_qbytes == 0) {
 864 #ifdef MSG_DEBUG_OK
 865                 kprintf("no such message queue id\n");
 866 #endif
 867                 eval = EINVAL;
 868                 goto done;
 869         }
 870         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 871 #ifdef MSG_DEBUG_OK
 872                 kprintf("wrong sequence number\n");
 873 #endif
 874                 eval = EINVAL;
 875                 goto done;
 876         }
 877
 878         if ((eval = ipcperm(td->td_proc, &msqptr->msg_perm, IPC_R))) {
 879 #ifdef MSG_DEBUG_OK
 880                 kprintf("requester doesn't have read access\n");
 881 #endif
 882                 goto done;
 883         }
 884
 885         msghdr = NULL;
 886         while (msghdr == NULL) {
 887                 if (msgtyp == 0) {
 888                         msghdr = msqptr->msg_first;
 889                         if (msghdr != NULL) {
 890                                 if (msgsz < msghdr->msg_ts &&
 891                                     (msgflg & MSG_NOERROR) == 0) {
 892 #ifdef MSG_DEBUG_OK
 893                                         kprintf("first message on the queue is too big (want %d, got %d)\n",
 894                                             msgsz, msghdr->msg_ts);
 895 #endif
 896                                         eval = E2BIG;
 897                                         goto done;
 898                                 }
 899                                 if (msqptr->msg_first == msqptr->msg_last) {
 900                                         msqptr->msg_first = NULL;
 901                                         msqptr->msg_last = NULL;
 902                                 } else {
 903                                         msqptr->msg_first = msghdr->msg_next;
 904                                         if (msqptr->msg_first == NULL)
 905                                                 panic("msg_first/last screwed up #1");
 906                                 }
 907                         }
 908                 } else {
 909                         struct msg *previous;
 910                         struct msg **prev;
 911
 912                         previous = NULL;
 913                         prev = &(msqptr->msg_first);
 914                         while ((msghdr = *prev) != NULL) {
 915                                 /*
 916                                  * Is this message's type an exact match or is
 917                                  * this message's type less than or equal to
 918                                  * the absolute value of a negative msgtyp?
 919                                  * Note that the second half of this test can
 920                                  * NEVER be true if msgtyp is positive since
 921                                  * msg_type is always positive!
 922                                  */
 923
 924                                 if (msgtyp == msghdr->msg_type ||
 925                                     msghdr->msg_type <= -msgtyp) {
 926 #ifdef MSG_DEBUG_OK
 927                                         kprintf("found message type %d, requested %d\n",
 928                                             msghdr->msg_type, msgtyp);
 929 #endif
 930                                         if (msgsz < msghdr->msg_ts &&
 931                                             (msgflg & MSG_NOERROR) == 0) {
 932 #ifdef MSG_DEBUG_OK
 933                                                 kprintf("requested message on the queue is too big (want %d, got %d)\n",
 934                                                     msgsz, msghdr->msg_ts);
 935 #endif
 936                                                 eval = E2BIG;
 937                                                 goto done;
 938                                         }
 939                                         *prev = msghdr->msg_next;
 940                                         if (msghdr == msqptr->msg_last) {
 941                                                 if (previous == NULL) {
 942                                                         if (prev !=
 943                                                             &msqptr->msg_first)
 944                                                                 panic("msg_first/last screwed up #2");
 945                                                         msqptr->msg_first =
 946                                                             NULL;
 947                                                         msqptr->msg_last =
 948                                                             NULL;
 949                                                 } else {
 950                                                         if (prev ==
 951                                                             &msqptr->msg_first)
 952                                                                 panic("msg_first/last screwed up #3");
 953                                                         msqptr->msg_last =
 954                                                             previous;
 955                                                 }
 956                                         }
 957                                         break;
 958                                 }
 959                                 previous = msghdr;
 960                                 prev = &(msghdr->msg_next);
 961                         }
 962                 }
 963
 964                 /*
 965                  * We've either extracted the msghdr for the appropriate
 966                  * message or there isn't one.
 967                  * If there is one then bail out of this loop.
 968                  */
 969
 970                 if (msghdr != NULL)
 971                         break;
 972
 973                 /*
 974                  * Hmph!  No message found.  Does the user want to wait?
 975                  */
 976
 977                 if ((msgflg & IPC_NOWAIT) != 0) {
 978 #ifdef MSG_DEBUG_OK
 979                         kprintf("no appropriate message found (msgtyp=%d)\n",
 980                             msgtyp);
 981 #endif
 982                         /* The SVID says to return ENOMSG. */
 983 #ifdef ENOMSG
 984                         eval = ENOMSG;
 985 #else
 986                         /* Unfortunately, BSD doesn't define that code yet! */
 987                         eval = EAGAIN;
 988 #endif
 989                         goto done;
 990                 }
 991
 992                 /*
 993                  * Wait for something to happen
 994                  */
 995
 996 #ifdef MSG_DEBUG_OK
 997                 kprintf("msgrcv:  goodnight\n");
 998 #endif
 999                 eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
1000 #ifdef MSG_DEBUG_OK
1001                 kprintf("msgrcv:  good morning (eval=%d)\n", eval);
1002 #endif
1003
1004                 if (eval != 0) {
1005 #ifdef MSG_DEBUG_OK
1006                         kprintf("msgsnd:  interrupted system call\n");
1007 #endif
1008                         eval = EINTR;
1009                         goto done;
1010                 }
1011
1012                 /*
1013                  * Make sure that the msq queue still exists
1014                  */
1015
1016                 if (msqptr->msg_qbytes == 0 ||
1017                     msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
1018 #ifdef MSG_DEBUG_OK
1019                         kprintf("msqid deleted\n");
1020 #endif
1021                         eval = EIDRM;
1022                         goto done;
1023                 }
1024         }
1025
1026         /*
1027          * Return the message to the user.
1028          *
1029          * First, do the bookkeeping (before we risk being interrupted).
1030          */
1031
1032         msqptr->msg_cbytes -= msghdr->msg_ts;
1033         msqptr->msg_qnum--;
1034         msqptr->msg_lrpid = td->td_proc->p_pid;
1035         msqptr->msg_rtime = time_second;
1036
1037         /*
1038          * Make msgsz the actual amount that we'll be returning.
1039          * Note that this effectively truncates the message if it is too long
1040          * (since msgsz is never increased).
1041          */
1042
1043 #ifdef MSG_DEBUG_OK
1044         kprintf("found a message, msgsz=%d, msg_ts=%d\n", msgsz,
1045             msghdr->msg_ts);
1046 #endif
1047         if (msgsz > msghdr->msg_ts)
1048                 msgsz = msghdr->msg_ts;
1049
1050         /*
1051          * Return the type to the user.
1052          */
1053
1054         eval = copyout((caddr_t)&(msghdr->msg_type), user_msgp,
1055             sizeof(msghdr->msg_type));
1056         if (eval != 0) {
1057 #ifdef MSG_DEBUG_OK
1058                 kprintf("error (%d) copying out message type\n", eval);
1059 #endif
1060                 msg_freehdr(msghdr);
1061                 wakeup((caddr_t)msqptr);
1062                 goto done;
1063         }
1064         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
1065
1066         /*
1067          * Return the segments to the user
1068          */
1069
1070         next = msghdr->msg_spot;
1071         for (len = 0; len < msgsz; len += msginfo.msgssz) {
1072                 size_t tlen;
1073
1074                 if (msgsz - len > msginfo.msgssz)
1075                         tlen = msginfo.msgssz;
1076                 else
1077                         tlen = msgsz - len;
1078                 if (next <= -1)
1079                         panic("next too low #3");
1080                 if (next >= msginfo.msgseg)
1081                         panic("next out of range #3");
1082                 eval = copyout((caddr_t)&msgpool[next * msginfo.msgssz],
1083                     user_msgp, tlen);
1084                 if (eval != 0) {
1085 #ifdef MSG_DEBUG_OK
1086                         kprintf("error (%d) copying out message segment\n",
1087                             eval);
1088 #endif
1089                         msg_freehdr(msghdr);
1090                         wakeup((caddr_t)msqptr);
1091                         goto done;
1092                 }
1093                 user_msgp = (char *)user_msgp + tlen;
1094                 next = msgmaps[next].next;
1095         }
1096
1097         /*
1098          * Done, return the actual number of bytes copied out.
1099          */
1100
1101         msg_freehdr(msghdr);
1102         wakeup((caddr_t)msqptr);
1103         eval = 0;
1104 done:
1105         rel_mplock();
1106         if (eval == 0)
1107                 uap->sysmsg_result = msgsz;
1108         return(eval);
1109 }
1110
1111 static int
1112 sysctl_msqids(SYSCTL_HANDLER_ARGS)
1113 {
1114
1115         return (SYSCTL_OUT(req, msqids,
1116             sizeof(struct msqid_ds) * msginfo.msgmni));
1117 }
1118
1119 TUNABLE_INT("kern.ipc.msgseg", &msginfo.msgseg);
1120 TUNABLE_INT("kern.ipc.msgssz", &msginfo.msgssz);
1121 TUNABLE_INT("kern.ipc.msgmni", &msginfo.msgmni);
1122
1123 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmax, CTLFLAG_RD, &msginfo.msgmax, 0,
1124     "Max characters in message");
1125 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmni, CTLFLAG_RD, &msginfo.msgmni, 0,
1126     "Max message queue identifiers");
1127 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmnb, CTLFLAG_RD, &msginfo.msgmnb, 0,
1128     "Max characters in message queue");
1129 SYSCTL_INT(_kern_ipc, OID_AUTO, msgtql, CTLFLAG_RD, &msginfo.msgtql, 0,
1130     "Max SVID messages in system");
1131 SYSCTL_INT(_kern_ipc, OID_AUTO, msgssz, CTLFLAG_RD, &msginfo.msgssz, 0,
1132     "Power-of-two size of a message segment");
1133 SYSCTL_INT(_kern_ipc, OID_AUTO, msgseg, CTLFLAG_RD, &msginfo.msgseg, 0,
1134     "Number of message segments");
1135 SYSCTL_PROC(_kern_ipc, OID_AUTO, msqids, CTLFLAG_RD,
1136     NULL, 0, sysctl_msqids, "", "Message queue IDs");