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 /* $DragonFly: src/sys/kern/sysv_msg.c,v 1.17 2006/12/23 23:47:54 swildner Exp $ */
   3
   4 /*
   5  * Implementation of SVID messages
   6  *
   7  * Author:  Daniel Boulet
   8  *
   9  * Copyright 1993 Daniel Boulet and RTMX Inc.
  10  *
  11  * This system call was implemented by Daniel Boulet under contract from RTMX.
  12  *
  13  * Redistribution and use in source forms, with and without modification,
  14  * are permitted provided that this entire comment appears intact.
  15  *
  16  * Redistribution in binary form may occur without any restrictions.
  17  * Obviously, it would be nice if you gave credit where credit is due
  18  * but requiring it would be too onerous.
  19  *
  20  * This software is provided ``AS IS'' without any warranties of any kind.
  21  */
  22
  23 #include "opt_sysvipc.h"
  24
  25 #include <sys/param.h>
  26 #include <sys/systm.h>
  27 #include <sys/sysproto.h>
  28 #include <sys/kernel.h>
  29 #include <sys/proc.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 static MALLOC_DEFINE(M_MSG, "msg", "SVID compatible message queues");
  37
  38 static void msginit (void *);
  39
  40 #define MSG_DEBUG
  41 #undef MSG_DEBUG_OK
  42
  43 static void msg_freehdr (struct msg *msghdr);
  44
  45 /* XXX casting to (sy_call_t *) is bogus, as usual. */
  46 static sy_call_t *msgcalls[] = {
  47         (sy_call_t *)sys_msgctl, (sy_call_t *)sys_msgget,
  48         (sy_call_t *)sys_msgsnd, (sy_call_t *)sys_msgrcv
  49 };
  50
  51 struct msg {
  52         struct  msg *msg_next;  /* next msg in the chain */
  53         long    msg_type;       /* type of this message */
  54                                 /* >0 -> type of this message */
  55                                 /* 0 -> free header */
  56         u_short msg_ts;         /* size of this message */
  57         short   msg_spot;       /* location of start of msg in buffer */
  58 };
  59
  60
  61 #ifndef MSGSSZ
  62 #define MSGSSZ  8               /* Each segment must be 2^N long */
  63 #endif
  64 #ifndef MSGSEG
  65 #define MSGSEG  2048            /* must be less than 32767 */
  66 #endif
  67 #define MSGMAX  (MSGSSZ*MSGSEG)
  68 #ifndef MSGMNB
  69 #define MSGMNB  2048            /* max # of bytes in a queue */
  70 #endif
  71 #ifndef MSGMNI
  72 #define MSGMNI  40
  73 #endif
  74 #ifndef MSGTQL
  75 #define MSGTQL  40
  76 #endif
  77
  78 /*
  79  * Based on the configuration parameters described in an SVR2 (yes, two)
  80  * config(1m) man page.
  81  *
  82  * Each message is broken up and stored in segments that are msgssz bytes
  83  * long.  For efficiency reasons, this should be a power of two.  Also,
  84  * it doesn't make sense if it is less than 8 or greater than about 256.
  85  * Consequently, msginit in kern/sysv_msg.c checks that msgssz is a power of
  86  * two between 8 and 1024 inclusive (and panic's if it isn't).
  87  */
  88 struct msginfo msginfo = {
  89                 MSGMAX,         /* max chars in a message */
  90                 MSGMNI,         /* # of message queue identifiers */
  91                 MSGMNB,         /* max chars in a queue */
  92                 MSGTQL,         /* max messages in system */
  93                 MSGSSZ,         /* size of a message segment */
  94                                 /* (must be small power of 2 greater than 4) */
  95                 MSGSEG          /* number of message segments */
  96 };
  97
  98 /*
  99  * macros to convert between msqid_ds's and msqid's.
 100  * (specific to this implementation)
 101  */
 102 #define MSQID(ix,ds)    ((ix) & 0xffff | (((ds).msg_perm.seq << 16) & 0xffff0000))
 103 #define MSQID_IX(id)    ((id) & 0xffff)
 104 #define MSQID_SEQ(id)   (((id) >> 16) & 0xffff)
 105
 106 /*
 107  * The rest of this file is specific to this particular implementation.
 108  */
 109
 110 struct msgmap {
 111         short   next;           /* next segment in buffer */
 112                                 /* -1 -> available */
 113                                 /* 0..(MSGSEG-1) -> index of next segment */
 114 };
 115
 116 #define MSG_LOCKED      01000   /* Is this msqid_ds locked? */
 117
 118 static int nfree_msgmaps;       /* # of free map entries */
 119 static short free_msgmaps;      /* head of linked list of free map entries */
 120 static struct msg *free_msghdrs;/* list of free msg headers */
 121 static char *msgpool;           /* MSGMAX byte long msg buffer pool */
 122 static struct msgmap *msgmaps;  /* MSGSEG msgmap structures */
 123 static struct msg *msghdrs;     /* MSGTQL msg headers */
 124 static struct msqid_ds *msqids; /* MSGMNI msqid_ds struct's */
 125
 126 static void
 127 msginit(void *dummy)
 128 {
 129         int i;
 130
 131         msginfo.msgmax = msginfo.msgseg * msginfo.msgssz;
 132         msgpool = kmalloc(msginfo.msgmax, M_MSG, M_WAITOK);
 133         if (msgpool == NULL)
 134                 panic("msgpool is NULL");
 135         msgmaps = kmalloc(sizeof(struct msgmap) * msginfo.msgseg, M_MSG, M_WAITOK);
 136         if (msgmaps == NULL)
 137                 panic("msgmaps is NULL");
 138         msghdrs = kmalloc(sizeof(struct msg) * msginfo.msgtql, M_MSG, M_WAITOK);
 139         if (msghdrs == NULL)
 140                 panic("msghdrs is NULL");
 141         msqids = kmalloc(sizeof(struct msqid_ds) * msginfo.msgmni, M_MSG, M_WAITOK);
 142         if (msqids == NULL)
 143                 panic("msqids is NULL");
 144
 145         /*
 146          * msginfo.msgssz should be a power of two for efficiency reasons.
 147          * It is also pretty silly if msginfo.msgssz is less than 8
 148          * or greater than about 256 so ...
 149          */
 150
 151         i = 8;
 152         while (i < 1024 && i != msginfo.msgssz)
 153                 i <<= 1;
 154         if (i != msginfo.msgssz) {
 155                 kprintf("msginfo.msgssz=%d (0x%x)\n", msginfo.msgssz,
 156                     msginfo.msgssz);
 157                 panic("msginfo.msgssz not a small power of 2");
 158         }
 159
 160         if (msginfo.msgseg > 32767) {
 161                 kprintf("msginfo.msgseg=%d\n", msginfo.msgseg);
 162                 panic("msginfo.msgseg > 32767");
 163         }
 164
 165         if (msgmaps == NULL)
 166                 panic("msgmaps is NULL");
 167
 168         for (i = 0; i < msginfo.msgseg; i++) {
 169                 if (i > 0)
 170                         msgmaps[i-1].next = i;
 171                 msgmaps[i].next = -1;   /* implies entry is available */
 172         }
 173         free_msgmaps = 0;
 174         nfree_msgmaps = msginfo.msgseg;
 175
 176         if (msghdrs == NULL)
 177                 panic("msghdrs is NULL");
 178
 179         for (i = 0; i < msginfo.msgtql; i++) {
 180                 msghdrs[i].msg_type = 0;
 181                 if (i > 0)
 182                         msghdrs[i-1].msg_next = &msghdrs[i];
 183                 msghdrs[i].msg_next = NULL;
 184         }
 185         free_msghdrs = &msghdrs[0];
 186
 187         if (msqids == NULL)
 188                 panic("msqids is NULL");
 189
 190         for (i = 0; i < msginfo.msgmni; i++) {
 191                 msqids[i].msg_qbytes = 0;       /* implies entry is available */
 192                 msqids[i].msg_perm.seq = 0;     /* reset to a known value */
 193                 msqids[i].msg_perm.mode = 0;
 194         }
 195 }
 196 SYSINIT(sysv_msg, SI_SUB_SYSV_MSG, SI_ORDER_FIRST, msginit, NULL)
 197
 198 /*
 199  * Entry point for all MSG calls
 200  *
 201  * msgsys_args(int which, int a2, ...) (VARARGS)
 202  */
 203 int
 204 sys_msgsys(struct msgsys_args *uap)
 205 {
 206         struct proc *p = curproc;
 207         unsigned int which = (unsigned int)uap->which;
 208
 209         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 210                 return (ENOSYS);
 211
 212         if (which >= sizeof(msgcalls)/sizeof(msgcalls[0]))
 213                 return (EINVAL);
 214         bcopy(&uap->a2, &uap->which,
 215             sizeof(struct msgsys_args) - offsetof(struct msgsys_args, a2));
 216         return ((*msgcalls[which])(uap));
 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 int
 243 sys_msgctl(struct msgctl_args *uap)
 244 {
 245         struct thread *td = curthread;
 246         struct proc *p = td->td_proc;
 247         int msqid = uap->msqid;
 248         int cmd = uap->cmd;
 249         struct msqid_ds *user_msqptr = uap->buf;
 250         int rval, eval;
 251         struct msqid_ds msqbuf;
 252         struct msqid_ds *msqptr;
 253
 254 #ifdef MSG_DEBUG_OK
 255         kprintf("call to msgctl(%d, %d, 0x%x)\n", msqid, cmd, user_msqptr);
 256 #endif
 257
 258         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 259                 return (ENOSYS);
 260
 261         msqid = IPCID_TO_IX(msqid);
 262
 263         if (msqid < 0 || msqid >= msginfo.msgmni) {
 264 #ifdef MSG_DEBUG_OK
 265                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 266                     msginfo.msgmni);
 267 #endif
 268                 return(EINVAL);
 269         }
 270
 271         msqptr = &msqids[msqid];
 272
 273         if (msqptr->msg_qbytes == 0) {
 274 #ifdef MSG_DEBUG_OK
 275                 kprintf("no such msqid\n");
 276 #endif
 277                 return(EINVAL);
 278         }
 279         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 280 #ifdef MSG_DEBUG_OK
 281                 kprintf("wrong sequence number\n");
 282 #endif
 283                 return(EINVAL);
 284         }
 285
 286         eval = 0;
 287         rval = 0;
 288
 289         switch (cmd) {
 290
 291         case IPC_RMID:
 292         {
 293                 struct msg *msghdr;
 294                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)))
 295                         return(eval);
 296                 /* Free the message headers */
 297                 msghdr = msqptr->msg_first;
 298                 while (msghdr != NULL) {
 299                         struct msg *msghdr_tmp;
 300
 301                         /* Free the segments of each message */
 302                         msqptr->msg_cbytes -= msghdr->msg_ts;
 303                         msqptr->msg_qnum--;
 304                         msghdr_tmp = msghdr;
 305                         msghdr = msghdr->msg_next;
 306                         msg_freehdr(msghdr_tmp);
 307                 }
 308
 309                 if (msqptr->msg_cbytes != 0)
 310                         panic("msg_cbytes is screwed up");
 311                 if (msqptr->msg_qnum != 0)
 312                         panic("msg_qnum is screwed up");
 313
 314                 msqptr->msg_qbytes = 0; /* Mark it as free */
 315
 316                 wakeup((caddr_t)msqptr);
 317         }
 318
 319                 break;
 320
 321         case IPC_SET:
 322                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)))
 323                         return(eval);
 324                 if ((eval = copyin(user_msqptr, &msqbuf, sizeof(msqbuf))) != 0)
 325                         return(eval);
 326                 if (msqbuf.msg_qbytes > msqptr->msg_qbytes) {
 327                         eval = suser(td);
 328                         if (eval)
 329                                 return(eval);
 330                 }
 331                 if (msqbuf.msg_qbytes > msginfo.msgmnb) {
 332 #ifdef MSG_DEBUG_OK
 333                         kprintf("can't increase msg_qbytes beyond %d (truncating)\n",
 334                             msginfo.msgmnb);
 335 #endif
 336                         msqbuf.msg_qbytes = msginfo.msgmnb;     /* silently restrict qbytes to system limit */
 337                 }
 338                 if (msqbuf.msg_qbytes == 0) {
 339 #ifdef MSG_DEBUG_OK
 340                         kprintf("can't reduce msg_qbytes to 0\n");
 341 #endif
 342                         return(EINVAL);         /* non-standard errno! */
 343                 }
 344                 msqptr->msg_perm.uid = msqbuf.msg_perm.uid;     /* change the owner */
 345                 msqptr->msg_perm.gid = msqbuf.msg_perm.gid;     /* change the owner */
 346                 msqptr->msg_perm.mode = (msqptr->msg_perm.mode & ~0777) |
 347                     (msqbuf.msg_perm.mode & 0777);
 348                 msqptr->msg_qbytes = msqbuf.msg_qbytes;
 349                 msqptr->msg_ctime = time_second;
 350                 break;
 351
 352         case IPC_STAT:
 353                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
 354 #ifdef MSG_DEBUG_OK
 355                         kprintf("requester doesn't have read access\n");
 356 #endif
 357                         return(eval);
 358                 }
 359                 eval = copyout((caddr_t)msqptr, user_msqptr,
 360                     sizeof(struct msqid_ds));
 361                 break;
 362
 363         default:
 364 #ifdef MSG_DEBUG_OK
 365                 kprintf("invalid command %d\n", cmd);
 366 #endif
 367                 return(EINVAL);
 368         }
 369
 370         if (eval == 0)
 371                 uap->sysmsg_result = rval;
 372         return(eval);
 373 }
 374
 375 int
 376 sys_msgget(struct msgget_args *uap)
 377 {
 378         struct proc *p = curproc;
 379         int msqid, eval;
 380         int key = uap->key;
 381         int msgflg = uap->msgflg;
 382         struct ucred *cred = p->p_ucred;
 383         struct msqid_ds *msqptr = NULL;
 384
 385 #ifdef MSG_DEBUG_OK
 386         kprintf("msgget(0x%x, 0%o)\n", key, msgflg);
 387 #endif
 388
 389         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 390                 return (ENOSYS);
 391
 392         if (key != IPC_PRIVATE) {
 393                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 394                         msqptr = &msqids[msqid];
 395                         if (msqptr->msg_qbytes != 0 &&
 396                             msqptr->msg_perm.key == key)
 397                                 break;
 398                 }
 399                 if (msqid < msginfo.msgmni) {
 400 #ifdef MSG_DEBUG_OK
 401                         kprintf("found public key\n");
 402 #endif
 403                         if ((msgflg & IPC_CREAT) && (msgflg & IPC_EXCL)) {
 404 #ifdef MSG_DEBUG_OK
 405                                 kprintf("not exclusive\n");
 406 #endif
 407                                 return(EEXIST);
 408                         }
 409                         if ((eval = ipcperm(p, &msqptr->msg_perm, msgflg & 0700 ))) {
 410 #ifdef MSG_DEBUG_OK
 411                                 kprintf("requester doesn't have 0%o access\n",
 412                                     msgflg & 0700);
 413 #endif
 414                                 return(eval);
 415                         }
 416                         goto found;
 417                 }
 418         }
 419
 420 #ifdef MSG_DEBUG_OK
 421         kprintf("need to allocate the msqid_ds\n");
 422 #endif
 423         if (key == IPC_PRIVATE || (msgflg & IPC_CREAT)) {
 424                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 425                         /*
 426                          * Look for an unallocated and unlocked msqid_ds.
 427                          * msqid_ds's can be locked by msgsnd or msgrcv while
 428                          * they are copying the message in/out.  We can't
 429                          * re-use the entry until they release it.
 430                          */
 431                         msqptr = &msqids[msqid];
 432                         if (msqptr->msg_qbytes == 0 &&
 433                             (msqptr->msg_perm.mode & MSG_LOCKED) == 0)
 434                                 break;
 435                 }
 436                 if (msqid == msginfo.msgmni) {
 437 #ifdef MSG_DEBUG_OK
 438                         kprintf("no more msqid_ds's available\n");
 439 #endif
 440                         return(ENOSPC);
 441                 }
 442 #ifdef MSG_DEBUG_OK
 443                 kprintf("msqid %d is available\n", msqid);
 444 #endif
 445                 msqptr->msg_perm.key = key;
 446                 msqptr->msg_perm.cuid = cred->cr_uid;
 447                 msqptr->msg_perm.uid = cred->cr_uid;
 448                 msqptr->msg_perm.cgid = cred->cr_gid;
 449                 msqptr->msg_perm.gid = cred->cr_gid;
 450                 msqptr->msg_perm.mode = (msgflg & 0777);
 451                 /* Make sure that the returned msqid is unique */
 452                 msqptr->msg_perm.seq = (msqptr->msg_perm.seq + 1) & 0x7fff;
 453                 msqptr->msg_first = NULL;
 454                 msqptr->msg_last = NULL;
 455                 msqptr->msg_cbytes = 0;
 456                 msqptr->msg_qnum = 0;
 457                 msqptr->msg_qbytes = msginfo.msgmnb;
 458                 msqptr->msg_lspid = 0;
 459                 msqptr->msg_lrpid = 0;
 460                 msqptr->msg_stime = 0;
 461                 msqptr->msg_rtime = 0;
 462                 msqptr->msg_ctime = time_second;
 463         } else {
 464 #ifdef MSG_DEBUG_OK
 465                 kprintf("didn't find it and wasn't asked to create it\n");
 466 #endif
 467                 return(ENOENT);
 468         }
 469
 470 found:
 471         /* Construct the unique msqid */
 472         uap->sysmsg_result = IXSEQ_TO_IPCID(msqid, msqptr->msg_perm);
 473         return(0);
 474 }
 475
 476 int
 477 sys_msgsnd(struct msgsnd_args *uap)
 478 {
 479         struct proc *p = curproc;
 480         int msqid = uap->msqid;
 481         void *user_msgp = uap->msgp;
 482         size_t msgsz = uap->msgsz;
 483         int msgflg = uap->msgflg;
 484         int segs_needed, eval;
 485         struct msqid_ds *msqptr;
 486         struct msg *msghdr;
 487         short next;
 488
 489 #ifdef MSG_DEBUG_OK
 490         kprintf("call to msgsnd(%d, 0x%x, %d, %d)\n", msqid, user_msgp, msgsz,
 491             msgflg);
 492 #endif
 493
 494         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 495                 return (ENOSYS);
 496
 497         msqid = IPCID_TO_IX(msqid);
 498
 499         if (msqid < 0 || msqid >= msginfo.msgmni) {
 500 #ifdef MSG_DEBUG_OK
 501                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 502                     msginfo.msgmni);
 503 #endif
 504                 return(EINVAL);
 505         }
 506
 507         msqptr = &msqids[msqid];
 508         if (msqptr->msg_qbytes == 0) {
 509 #ifdef MSG_DEBUG_OK
 510                 kprintf("no such message queue id\n");
 511 #endif
 512                 return(EINVAL);
 513         }
 514         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 515 #ifdef MSG_DEBUG_OK
 516                 kprintf("wrong sequence number\n");
 517 #endif
 518                 return(EINVAL);
 519         }
 520
 521         if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_W))) {
 522 #ifdef MSG_DEBUG_OK
 523                 kprintf("requester doesn't have write access\n");
 524 #endif
 525                 return(eval);
 526         }
 527
 528         segs_needed = (msgsz + msginfo.msgssz - 1) / msginfo.msgssz;
 529 #ifdef MSG_DEBUG_OK
 530         kprintf("msgsz=%d, msgssz=%d, segs_needed=%d\n", msgsz, msginfo.msgssz,
 531             segs_needed);
 532 #endif
 533         for (;;) {
 534                 int need_more_resources = 0;
 535
 536                 /*
 537                  * check msgsz
 538                  * (inside this loop in case msg_qbytes changes while we sleep)
 539                  */
 540
 541                 if (msgsz > msqptr->msg_qbytes) {
 542 #ifdef MSG_DEBUG_OK
 543                         kprintf("msgsz > msqptr->msg_qbytes\n");
 544 #endif
 545                         return(EINVAL);
 546                 }
 547
 548                 if (msqptr->msg_perm.mode & MSG_LOCKED) {
 549 #ifdef MSG_DEBUG_OK
 550                         kprintf("msqid is locked\n");
 551 #endif
 552                         need_more_resources = 1;
 553                 }
 554                 if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes) {
 555 #ifdef MSG_DEBUG_OK
 556                         kprintf("msgsz + msg_cbytes > msg_qbytes\n");
 557 #endif
 558                         need_more_resources = 1;
 559                 }
 560                 if (segs_needed > nfree_msgmaps) {
 561 #ifdef MSG_DEBUG_OK
 562                         kprintf("segs_needed > nfree_msgmaps\n");
 563 #endif
 564                         need_more_resources = 1;
 565                 }
 566                 if (free_msghdrs == NULL) {
 567 #ifdef MSG_DEBUG_OK
 568                         kprintf("no more msghdrs\n");
 569 #endif
 570                         need_more_resources = 1;
 571                 }
 572
 573                 if (need_more_resources) {
 574                         int we_own_it;
 575
 576                         if ((msgflg & IPC_NOWAIT) != 0) {
 577 #ifdef MSG_DEBUG_OK
 578                                 kprintf("need more resources but caller doesn't want to wait\n");
 579 #endif
 580                                 return(EAGAIN);
 581                         }
 582
 583                         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0) {
 584 #ifdef MSG_DEBUG_OK
 585                                 kprintf("we don't own the msqid_ds\n");
 586 #endif
 587                                 we_own_it = 0;
 588                         } else {
 589                                 /* Force later arrivals to wait for our
 590                                    request */
 591 #ifdef MSG_DEBUG_OK
 592                                 kprintf("we own the msqid_ds\n");
 593 #endif
 594                                 msqptr->msg_perm.mode |= MSG_LOCKED;
 595                                 we_own_it = 1;
 596                         }
 597 #ifdef MSG_DEBUG_OK
 598                         kprintf("goodnight\n");
 599 #endif
 600                         eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 601 #ifdef MSG_DEBUG_OK
 602                         kprintf("good morning, eval=%d\n", eval);
 603 #endif
 604                         if (we_own_it)
 605                                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 606                         if (eval != 0) {
 607 #ifdef MSG_DEBUG_OK
 608                                 kprintf("msgsnd:  interrupted system call\n");
 609 #endif
 610                                 return(EINTR);
 611                         }
 612
 613                         /*
 614                          * Make sure that the msq queue still exists
 615                          */
 616
 617                         if (msqptr->msg_qbytes == 0) {
 618 #ifdef MSG_DEBUG_OK
 619                                 kprintf("msqid deleted\n");
 620 #endif
 621                                 return(EIDRM);
 622                         }
 623
 624                 } else {
 625 #ifdef MSG_DEBUG_OK
 626                         kprintf("got all the resources that we need\n");
 627 #endif
 628                         break;
 629                 }
 630         }
 631
 632         /*
 633          * We have the resources that we need.
 634          * Make sure!
 635          */
 636
 637         if (msqptr->msg_perm.mode & MSG_LOCKED)
 638                 panic("msg_perm.mode & MSG_LOCKED");
 639         if (segs_needed > nfree_msgmaps)
 640                 panic("segs_needed > nfree_msgmaps");
 641         if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes)
 642                 panic("msgsz + msg_cbytes > msg_qbytes");
 643         if (free_msghdrs == NULL)
 644                 panic("no more msghdrs");
 645
 646         /*
 647          * Re-lock the msqid_ds in case we page-fault when copying in the
 648          * message
 649          */
 650
 651         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0)
 652                 panic("msqid_ds is already locked");
 653         msqptr->msg_perm.mode |= MSG_LOCKED;
 654
 655         /*
 656          * Allocate a message header
 657          */
 658
 659         msghdr = free_msghdrs;
 660         free_msghdrs = msghdr->msg_next;
 661         msghdr->msg_spot = -1;
 662         msghdr->msg_ts = msgsz;
 663
 664         /*
 665          * Allocate space for the message
 666          */
 667
 668         while (segs_needed > 0) {
 669                 if (nfree_msgmaps <= 0)
 670                         panic("not enough msgmaps");
 671                 if (free_msgmaps == -1)
 672                         panic("nil free_msgmaps");
 673                 next = free_msgmaps;
 674                 if (next <= -1)
 675                         panic("next too low #1");
 676                 if (next >= msginfo.msgseg)
 677                         panic("next out of range #1");
 678 #ifdef MSG_DEBUG_OK
 679                 kprintf("allocating segment %d to message\n", next);
 680 #endif
 681                 free_msgmaps = msgmaps[next].next;
 682                 nfree_msgmaps--;
 683                 msgmaps[next].next = msghdr->msg_spot;
 684                 msghdr->msg_spot = next;
 685                 segs_needed--;
 686         }
 687
 688         /*
 689          * Copy in the message type
 690          */
 691
 692         if ((eval = copyin(user_msgp, &msghdr->msg_type,
 693             sizeof(msghdr->msg_type))) != 0) {
 694 #ifdef MSG_DEBUG_OK
 695                 kprintf("error %d copying the message type\n", eval);
 696 #endif
 697                 msg_freehdr(msghdr);
 698                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 699                 wakeup((caddr_t)msqptr);
 700                 return(eval);
 701         }
 702         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
 703
 704         /*
 705          * Validate the message type
 706          */
 707
 708         if (msghdr->msg_type < 1) {
 709                 msg_freehdr(msghdr);
 710                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 711                 wakeup((caddr_t)msqptr);
 712 #ifdef MSG_DEBUG_OK
 713                 kprintf("mtype (%d) < 1\n", msghdr->msg_type);
 714 #endif
 715                 return(EINVAL);
 716         }
 717
 718         /*
 719          * Copy in the message body
 720          */
 721
 722         next = msghdr->msg_spot;
 723         while (msgsz > 0) {
 724                 size_t tlen;
 725                 if (msgsz > msginfo.msgssz)
 726                         tlen = msginfo.msgssz;
 727                 else
 728                         tlen = msgsz;
 729                 if (next <= -1)
 730                         panic("next too low #2");
 731                 if (next >= msginfo.msgseg)
 732                         panic("next out of range #2");
 733                 if ((eval = copyin(user_msgp, &msgpool[next * msginfo.msgssz],
 734                     tlen)) != 0) {
 735 #ifdef MSG_DEBUG_OK
 736                         kprintf("error %d copying in message segment\n", eval);
 737 #endif
 738                         msg_freehdr(msghdr);
 739                         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 740                         wakeup((caddr_t)msqptr);
 741                         return(eval);
 742                 }
 743                 msgsz -= tlen;
 744                 user_msgp = (char *)user_msgp + tlen;
 745                 next = msgmaps[next].next;
 746         }
 747         if (next != -1)
 748                 panic("didn't use all the msg segments");
 749
 750         /*
 751          * We've got the message.  Unlock the msqid_ds.
 752          */
 753
 754         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 755
 756         /*
 757          * Make sure that the msqid_ds is still allocated.
 758          */
 759
 760         if (msqptr->msg_qbytes == 0) {
 761                 msg_freehdr(msghdr);
 762                 wakeup((caddr_t)msqptr);
 763                 return(EIDRM);
 764         }
 765
 766         /*
 767          * Put the message into the queue
 768          */
 769
 770         if (msqptr->msg_first == NULL) {
 771                 msqptr->msg_first = msghdr;
 772                 msqptr->msg_last = msghdr;
 773         } else {
 774                 msqptr->msg_last->msg_next = msghdr;
 775                 msqptr->msg_last = msghdr;
 776         }
 777         msqptr->msg_last->msg_next = NULL;
 778
 779         msqptr->msg_cbytes += msghdr->msg_ts;
 780         msqptr->msg_qnum++;
 781         msqptr->msg_lspid = p->p_pid;
 782         msqptr->msg_stime = time_second;
 783
 784         wakeup((caddr_t)msqptr);
 785         uap->sysmsg_result = 0;
 786         return(0);
 787 }
 788
 789 int
 790 sys_msgrcv(struct msgrcv_args *uap)
 791 {
 792         struct proc *p = curproc;
 793         int msqid = uap->msqid;
 794         void *user_msgp = uap->msgp;
 795         size_t msgsz = uap->msgsz;
 796         long msgtyp = uap->msgtyp;
 797         int msgflg = uap->msgflg;
 798         size_t len;
 799         struct msqid_ds *msqptr;
 800         struct msg *msghdr;
 801         int eval;
 802         short next;
 803
 804 #ifdef MSG_DEBUG_OK
 805         kprintf("call to msgrcv(%d, 0x%x, %d, %ld, %d)\n", msqid, user_msgp,
 806             msgsz, msgtyp, msgflg);
 807 #endif
 808
 809         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 810                 return (ENOSYS);
 811
 812         msqid = IPCID_TO_IX(msqid);
 813
 814         if (msqid < 0 || msqid >= msginfo.msgmni) {
 815 #ifdef MSG_DEBUG_OK
 816                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 817                     msginfo.msgmni);
 818 #endif
 819                 return(EINVAL);
 820         }
 821
 822         msqptr = &msqids[msqid];
 823         if (msqptr->msg_qbytes == 0) {
 824 #ifdef MSG_DEBUG_OK
 825                 kprintf("no such message queue id\n");
 826 #endif
 827                 return(EINVAL);
 828         }
 829         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 830 #ifdef MSG_DEBUG_OK
 831                 kprintf("wrong sequence number\n");
 832 #endif
 833                 return(EINVAL);
 834         }
 835
 836         if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
 837 #ifdef MSG_DEBUG_OK
 838                 kprintf("requester doesn't have read access\n");
 839 #endif
 840                 return(eval);
 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                                         return(E2BIG);
 855                                 }
 856                                 if (msqptr->msg_first == msqptr->msg_last) {
 857                                         msqptr->msg_first = NULL;
 858                                         msqptr->msg_last = NULL;
 859                                 } else {
 860                                         msqptr->msg_first = msghdr->msg_next;
 861                                         if (msqptr->msg_first == NULL)
 862                                                 panic("msg_first/last screwed up #1");
 863                                 }
 864                         }
 865                 } else {
 866                         struct msg *previous;
 867                         struct msg **prev;
 868
 869                         previous = NULL;
 870                         prev = &(msqptr->msg_first);
 871                         while ((msghdr = *prev) != NULL) {
 872                                 /*
 873                                  * Is this message's type an exact match or is
 874                                  * this message's type less than or equal to
 875                                  * the absolute value of a negative msgtyp?
 876                                  * Note that the second half of this test can
 877                                  * NEVER be true if msgtyp is positive since
 878                                  * msg_type is always positive!
 879                                  */
 880
 881                                 if (msgtyp == msghdr->msg_type ||
 882                                     msghdr->msg_type <= -msgtyp) {
 883 #ifdef MSG_DEBUG_OK
 884                                         kprintf("found message type %d, requested %d\n",
 885                                             msghdr->msg_type, msgtyp);
 886 #endif
 887                                         if (msgsz < msghdr->msg_ts &&
 888                                             (msgflg & MSG_NOERROR) == 0) {
 889 #ifdef MSG_DEBUG_OK
 890                                                 kprintf("requested message on the queue is too big (want %d, got %d)\n",
 891                                                     msgsz, msghdr->msg_ts);
 892 #endif
 893                                                 return(E2BIG);
 894                                         }
 895                                         *prev = msghdr->msg_next;
 896                                         if (msghdr == msqptr->msg_last) {
 897                                                 if (previous == NULL) {
 898                                                         if (prev !=
 899                                                             &msqptr->msg_first)
 900                                                                 panic("msg_first/last screwed up #2");
 901                                                         msqptr->msg_first =
 902                                                             NULL;
 903                                                         msqptr->msg_last =
 904                                                             NULL;
 905                                                 } else {
 906                                                         if (prev ==
 907                                                             &msqptr->msg_first)
 908                                                                 panic("msg_first/last screwed up #3");
 909                                                         msqptr->msg_last =
 910                                                             previous;
 911                                                 }
 912                                         }
 913                                         break;
 914                                 }
 915                                 previous = msghdr;
 916                                 prev = &(msghdr->msg_next);
 917                         }
 918                 }
 919
 920                 /*
 921                  * We've either extracted the msghdr for the appropriate
 922                  * message or there isn't one.
 923                  * If there is one then bail out of this loop.
 924                  */
 925
 926                 if (msghdr != NULL)
 927                         break;
 928
 929                 /*
 930                  * Hmph!  No message found.  Does the user want to wait?
 931                  */
 932
 933                 if ((msgflg & IPC_NOWAIT) != 0) {
 934 #ifdef MSG_DEBUG_OK
 935                         kprintf("no appropriate message found (msgtyp=%d)\n",
 936                             msgtyp);
 937 #endif
 938                         /* The SVID says to return ENOMSG. */
 939 #ifdef ENOMSG
 940                         return(ENOMSG);
 941 #else
 942                         /* Unfortunately, BSD doesn't define that code yet! */
 943                         return(EAGAIN);
 944 #endif
 945                 }
 946
 947                 /*
 948                  * Wait for something to happen
 949                  */
 950
 951 #ifdef MSG_DEBUG_OK
 952                 kprintf("msgrcv:  goodnight\n");
 953 #endif
 954                 eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 955 #ifdef MSG_DEBUG_OK
 956                 kprintf("msgrcv:  good morning (eval=%d)\n", eval);
 957 #endif
 958
 959                 if (eval != 0) {
 960 #ifdef MSG_DEBUG_OK
 961                         kprintf("msgsnd:  interrupted system call\n");
 962 #endif
 963                         return(EINTR);
 964                 }
 965
 966                 /*
 967                  * Make sure that the msq queue still exists
 968                  */
 969
 970                 if (msqptr->msg_qbytes == 0 ||
 971                     msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 972 #ifdef MSG_DEBUG_OK
 973                         kprintf("msqid deleted\n");
 974 #endif
 975                         return(EIDRM);
 976                 }
 977         }
 978
 979         /*
 980          * Return the message to the user.
 981          *
 982          * First, do the bookkeeping (before we risk being interrupted).
 983          */
 984
 985         msqptr->msg_cbytes -= msghdr->msg_ts;
 986         msqptr->msg_qnum--;
 987         msqptr->msg_lrpid = p->p_pid;
 988         msqptr->msg_rtime = time_second;
 989
 990         /*
 991          * Make msgsz the actual amount that we'll be returning.
 992          * Note that this effectively truncates the message if it is too long
 993          * (since msgsz is never increased).
 994          */
 995
 996 #ifdef MSG_DEBUG_OK
 997         kprintf("found a message, msgsz=%d, msg_ts=%d\n", msgsz,
 998             msghdr->msg_ts);
 999 #endif
1000         if (msgsz > msghdr->msg_ts)
1001                 msgsz = msghdr->msg_ts;
1002
1003         /*
1004          * Return the type to the user.
1005          */
1006
1007         eval = copyout((caddr_t)&(msghdr->msg_type), user_msgp,
1008             sizeof(msghdr->msg_type));
1009         if (eval != 0) {
1010 #ifdef MSG_DEBUG_OK
1011                 kprintf("error (%d) copying out message type\n", eval);
1012 #endif
1013                 msg_freehdr(msghdr);
1014                 wakeup((caddr_t)msqptr);
1015                 return(eval);
1016         }
1017         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
1018
1019         /*
1020          * Return the segments to the user
1021          */
1022
1023         next = msghdr->msg_spot;
1024         for (len = 0; len < msgsz; len += msginfo.msgssz) {
1025                 size_t tlen;
1026
1027                 if (msgsz - len > msginfo.msgssz)
1028                         tlen = msginfo.msgssz;
1029                 else
1030                         tlen = msgsz - len;
1031                 if (next <= -1)
1032                         panic("next too low #3");
1033                 if (next >= msginfo.msgseg)
1034                         panic("next out of range #3");
1035                 eval = copyout((caddr_t)&msgpool[next * msginfo.msgssz],
1036                     user_msgp, tlen);
1037                 if (eval != 0) {
1038 #ifdef MSG_DEBUG_OK
1039                         kprintf("error (%d) copying out message segment\n",
1040                             eval);
1041 #endif
1042                         msg_freehdr(msghdr);
1043                         wakeup((caddr_t)msqptr);
1044                         return(eval);
1045                 }
1046                 user_msgp = (char *)user_msgp + tlen;
1047                 next = msgmaps[next].next;
1048         }
1049
1050         /*
1051          * Done, return the actual number of bytes copied out.
1052          */
1053
1054         msg_freehdr(msghdr);
1055         wakeup((caddr_t)msqptr);
1056         uap->sysmsg_result = msgsz;
1057         return(0);
1058 }
1059
1060 static int
1061 sysctl_msqids(SYSCTL_HANDLER_ARGS)
1062 {
1063
1064         return (SYSCTL_OUT(req, msqids,
1065             sizeof(struct msqid_ds) * msginfo.msgmni));
1066 }
1067
1068 TUNABLE_INT("kern.ipc.msgseg", &msginfo.msgseg);
1069 TUNABLE_INT("kern.ipc.msgssz", &msginfo.msgssz);
1070 TUNABLE_INT("kern.ipc.msgmni", &msginfo.msgmni);
1071
1072 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmax, CTLFLAG_RD, &msginfo.msgmax, 0, "");
1073 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmni, CTLFLAG_RD, &msginfo.msgmni, 0, "");
1074 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmnb, CTLFLAG_RD, &msginfo.msgmnb, 0, "");
1075 SYSCTL_INT(_kern_ipc, OID_AUTO, msgtql, CTLFLAG_RD, &msginfo.msgtql, 0, "");
1076 SYSCTL_INT(_kern_ipc, OID_AUTO, msgssz, CTLFLAG_RD, &msginfo.msgssz, 0, "");
1077 SYSCTL_INT(_kern_ipc, OID_AUTO, msgseg, CTLFLAG_RD, &msginfo.msgseg, 0, "");
1078 SYSCTL_PROC(_kern_ipc, OID_AUTO, msqids, CTLFLAG_RD,
1079     NULL, 0, sysctl_msqids, "", "Message queue IDs");