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 1.18 2008/01/06 16:55:51 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         msgmaps = kmalloc(sizeof(struct msgmap) * msginfo.msgseg, M_MSG, M_WAITOK);
 134         msghdrs = kmalloc(sizeof(struct msg) * msginfo.msgtql, M_MSG, M_WAITOK);
 135         msqids = kmalloc(sizeof(struct msqid_ds) * msginfo.msgmni, M_MSG, M_WAITOK);
 136
 137         /*
 138          * msginfo.msgssz should be a power of two for efficiency reasons.
 139          * It is also pretty silly if msginfo.msgssz is less than 8
 140          * or greater than about 256 so ...
 141          */
 142
 143         i = 8;
 144         while (i < 1024 && i != msginfo.msgssz)
 145                 i <<= 1;
 146         if (i != msginfo.msgssz) {
 147                 kprintf("msginfo.msgssz=%d (0x%x)\n", msginfo.msgssz,
 148                     msginfo.msgssz);
 149                 panic("msginfo.msgssz not a small power of 2");
 150         }
 151
 152         if (msginfo.msgseg > 32767) {
 153                 kprintf("msginfo.msgseg=%d\n", msginfo.msgseg);
 154                 panic("msginfo.msgseg > 32767");
 155         }
 156
 157         if (msgmaps == NULL)
 158                 panic("msgmaps is NULL");
 159
 160         for (i = 0; i < msginfo.msgseg; i++) {
 161                 if (i > 0)
 162                         msgmaps[i-1].next = i;
 163                 msgmaps[i].next = -1;   /* implies entry is available */
 164         }
 165         free_msgmaps = 0;
 166         nfree_msgmaps = msginfo.msgseg;
 167
 168         if (msghdrs == NULL)
 169                 panic("msghdrs is NULL");
 170
 171         for (i = 0; i < msginfo.msgtql; i++) {
 172                 msghdrs[i].msg_type = 0;
 173                 if (i > 0)
 174                         msghdrs[i-1].msg_next = &msghdrs[i];
 175                 msghdrs[i].msg_next = NULL;
 176         }
 177         free_msghdrs = &msghdrs[0];
 178
 179         if (msqids == NULL)
 180                 panic("msqids is NULL");
 181
 182         for (i = 0; i < msginfo.msgmni; i++) {
 183                 msqids[i].msg_qbytes = 0;       /* implies entry is available */
 184                 msqids[i].msg_perm.seq = 0;     /* reset to a known value */
 185                 msqids[i].msg_perm.mode = 0;
 186         }
 187 }
 188 SYSINIT(sysv_msg, SI_SUB_SYSV_MSG, SI_ORDER_FIRST, msginit, NULL)
 189
 190 /*
 191  * Entry point for all MSG calls
 192  *
 193  * msgsys_args(int which, int a2, ...) (VARARGS)
 194  */
 195 int
 196 sys_msgsys(struct msgsys_args *uap)
 197 {
 198         struct proc *p = curproc;
 199         unsigned int which = (unsigned int)uap->which;
 200
 201         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 202                 return (ENOSYS);
 203
 204         if (which >= sizeof(msgcalls)/sizeof(msgcalls[0]))
 205                 return (EINVAL);
 206         bcopy(&uap->a2, &uap->which,
 207             sizeof(struct msgsys_args) - offsetof(struct msgsys_args, a2));
 208         return ((*msgcalls[which])(uap));
 209 }
 210
 211 static void
 212 msg_freehdr(struct msg *msghdr)
 213 {
 214         while (msghdr->msg_ts > 0) {
 215                 short next;
 216                 if (msghdr->msg_spot < 0 || msghdr->msg_spot >= msginfo.msgseg)
 217                         panic("msghdr->msg_spot out of range");
 218                 next = msgmaps[msghdr->msg_spot].next;
 219                 msgmaps[msghdr->msg_spot].next = free_msgmaps;
 220                 free_msgmaps = msghdr->msg_spot;
 221                 nfree_msgmaps++;
 222                 msghdr->msg_spot = next;
 223                 if (msghdr->msg_ts >= msginfo.msgssz)
 224                         msghdr->msg_ts -= msginfo.msgssz;
 225                 else
 226                         msghdr->msg_ts = 0;
 227         }
 228         if (msghdr->msg_spot != -1)
 229                 panic("msghdr->msg_spot != -1");
 230         msghdr->msg_next = free_msghdrs;
 231         free_msghdrs = msghdr;
 232 }
 233
 234 int
 235 sys_msgctl(struct msgctl_args *uap)
 236 {
 237         struct thread *td = curthread;
 238         struct proc *p = td->td_proc;
 239         int msqid = uap->msqid;
 240         int cmd = uap->cmd;
 241         struct msqid_ds *user_msqptr = uap->buf;
 242         int rval, eval;
 243         struct msqid_ds msqbuf;
 244         struct msqid_ds *msqptr;
 245
 246 #ifdef MSG_DEBUG_OK
 247         kprintf("call to msgctl(%d, %d, 0x%x)\n", msqid, cmd, user_msqptr);
 248 #endif
 249
 250         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 251                 return (ENOSYS);
 252
 253         msqid = IPCID_TO_IX(msqid);
 254
 255         if (msqid < 0 || msqid >= msginfo.msgmni) {
 256 #ifdef MSG_DEBUG_OK
 257                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 258                     msginfo.msgmni);
 259 #endif
 260                 return(EINVAL);
 261         }
 262
 263         msqptr = &msqids[msqid];
 264
 265         if (msqptr->msg_qbytes == 0) {
 266 #ifdef MSG_DEBUG_OK
 267                 kprintf("no such msqid\n");
 268 #endif
 269                 return(EINVAL);
 270         }
 271         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 272 #ifdef MSG_DEBUG_OK
 273                 kprintf("wrong sequence number\n");
 274 #endif
 275                 return(EINVAL);
 276         }
 277
 278         eval = 0;
 279         rval = 0;
 280
 281         switch (cmd) {
 282
 283         case IPC_RMID:
 284         {
 285                 struct msg *msghdr;
 286                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)))
 287                         return(eval);
 288                 /* Free the message headers */
 289                 msghdr = msqptr->msg_first;
 290                 while (msghdr != NULL) {
 291                         struct msg *msghdr_tmp;
 292
 293                         /* Free the segments of each message */
 294                         msqptr->msg_cbytes -= msghdr->msg_ts;
 295                         msqptr->msg_qnum--;
 296                         msghdr_tmp = msghdr;
 297                         msghdr = msghdr->msg_next;
 298                         msg_freehdr(msghdr_tmp);
 299                 }
 300
 301                 if (msqptr->msg_cbytes != 0)
 302                         panic("msg_cbytes is screwed up");
 303                 if (msqptr->msg_qnum != 0)
 304                         panic("msg_qnum is screwed up");
 305
 306                 msqptr->msg_qbytes = 0; /* Mark it as free */
 307
 308                 wakeup((caddr_t)msqptr);
 309         }
 310
 311                 break;
 312
 313         case IPC_SET:
 314                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)))
 315                         return(eval);
 316                 if ((eval = copyin(user_msqptr, &msqbuf, sizeof(msqbuf))) != 0)
 317                         return(eval);
 318                 if (msqbuf.msg_qbytes > msqptr->msg_qbytes) {
 319                         eval = suser(td);
 320                         if (eval)
 321                                 return(eval);
 322                 }
 323                 if (msqbuf.msg_qbytes > msginfo.msgmnb) {
 324 #ifdef MSG_DEBUG_OK
 325                         kprintf("can't increase msg_qbytes beyond %d (truncating)\n",
 326                             msginfo.msgmnb);
 327 #endif
 328                         msqbuf.msg_qbytes = msginfo.msgmnb;     /* silently restrict qbytes to system limit */
 329                 }
 330                 if (msqbuf.msg_qbytes == 0) {
 331 #ifdef MSG_DEBUG_OK
 332                         kprintf("can't reduce msg_qbytes to 0\n");
 333 #endif
 334                         return(EINVAL);         /* non-standard errno! */
 335                 }
 336                 msqptr->msg_perm.uid = msqbuf.msg_perm.uid;     /* change the owner */
 337                 msqptr->msg_perm.gid = msqbuf.msg_perm.gid;     /* change the owner */
 338                 msqptr->msg_perm.mode = (msqptr->msg_perm.mode & ~0777) |
 339                     (msqbuf.msg_perm.mode & 0777);
 340                 msqptr->msg_qbytes = msqbuf.msg_qbytes;
 341                 msqptr->msg_ctime = time_second;
 342                 break;
 343
 344         case IPC_STAT:
 345                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
 346 #ifdef MSG_DEBUG_OK
 347                         kprintf("requester doesn't have read access\n");
 348 #endif
 349                         return(eval);
 350                 }
 351                 eval = copyout((caddr_t)msqptr, user_msqptr,
 352                     sizeof(struct msqid_ds));
 353                 break;
 354
 355         default:
 356 #ifdef MSG_DEBUG_OK
 357                 kprintf("invalid command %d\n", cmd);
 358 #endif
 359                 return(EINVAL);
 360         }
 361
 362         if (eval == 0)
 363                 uap->sysmsg_result = rval;
 364         return(eval);
 365 }
 366
 367 int
 368 sys_msgget(struct msgget_args *uap)
 369 {
 370         struct proc *p = curproc;
 371         int msqid, eval;
 372         int key = uap->key;
 373         int msgflg = uap->msgflg;
 374         struct ucred *cred = p->p_ucred;
 375         struct msqid_ds *msqptr = NULL;
 376
 377 #ifdef MSG_DEBUG_OK
 378         kprintf("msgget(0x%x, 0%o)\n", key, msgflg);
 379 #endif
 380
 381         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 382                 return (ENOSYS);
 383
 384         if (key != IPC_PRIVATE) {
 385                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 386                         msqptr = &msqids[msqid];
 387                         if (msqptr->msg_qbytes != 0 &&
 388                             msqptr->msg_perm.key == key)
 389                                 break;
 390                 }
 391                 if (msqid < msginfo.msgmni) {
 392 #ifdef MSG_DEBUG_OK
 393                         kprintf("found public key\n");
 394 #endif
 395                         if ((msgflg & IPC_CREAT) && (msgflg & IPC_EXCL)) {
 396 #ifdef MSG_DEBUG_OK
 397                                 kprintf("not exclusive\n");
 398 #endif
 399                                 return(EEXIST);
 400                         }
 401                         if ((eval = ipcperm(p, &msqptr->msg_perm, msgflg & 0700 ))) {
 402 #ifdef MSG_DEBUG_OK
 403                                 kprintf("requester doesn't have 0%o access\n",
 404                                     msgflg & 0700);
 405 #endif
 406                                 return(eval);
 407                         }
 408                         goto found;
 409                 }
 410         }
 411
 412 #ifdef MSG_DEBUG_OK
 413         kprintf("need to allocate the msqid_ds\n");
 414 #endif
 415         if (key == IPC_PRIVATE || (msgflg & IPC_CREAT)) {
 416                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 417                         /*
 418                          * Look for an unallocated and unlocked msqid_ds.
 419                          * msqid_ds's can be locked by msgsnd or msgrcv while
 420                          * they are copying the message in/out.  We can't
 421                          * re-use the entry until they release it.
 422                          */
 423                         msqptr = &msqids[msqid];
 424                         if (msqptr->msg_qbytes == 0 &&
 425                             (msqptr->msg_perm.mode & MSG_LOCKED) == 0)
 426                                 break;
 427                 }
 428                 if (msqid == msginfo.msgmni) {
 429 #ifdef MSG_DEBUG_OK
 430                         kprintf("no more msqid_ds's available\n");
 431 #endif
 432                         return(ENOSPC);
 433                 }
 434 #ifdef MSG_DEBUG_OK
 435                 kprintf("msqid %d is available\n", msqid);
 436 #endif
 437                 msqptr->msg_perm.key = key;
 438                 msqptr->msg_perm.cuid = cred->cr_uid;
 439                 msqptr->msg_perm.uid = cred->cr_uid;
 440                 msqptr->msg_perm.cgid = cred->cr_gid;
 441                 msqptr->msg_perm.gid = cred->cr_gid;
 442                 msqptr->msg_perm.mode = (msgflg & 0777);
 443                 /* Make sure that the returned msqid is unique */
 444                 msqptr->msg_perm.seq = (msqptr->msg_perm.seq + 1) & 0x7fff;
 445                 msqptr->msg_first = NULL;
 446                 msqptr->msg_last = NULL;
 447                 msqptr->msg_cbytes = 0;
 448                 msqptr->msg_qnum = 0;
 449                 msqptr->msg_qbytes = msginfo.msgmnb;
 450                 msqptr->msg_lspid = 0;
 451                 msqptr->msg_lrpid = 0;
 452                 msqptr->msg_stime = 0;
 453                 msqptr->msg_rtime = 0;
 454                 msqptr->msg_ctime = time_second;
 455         } else {
 456 #ifdef MSG_DEBUG_OK
 457                 kprintf("didn't find it and wasn't asked to create it\n");
 458 #endif
 459                 return(ENOENT);
 460         }
 461
 462 found:
 463         /* Construct the unique msqid */
 464         uap->sysmsg_result = IXSEQ_TO_IPCID(msqid, msqptr->msg_perm);
 465         return(0);
 466 }
 467
 468 int
 469 sys_msgsnd(struct msgsnd_args *uap)
 470 {
 471         struct proc *p = curproc;
 472         int msqid = uap->msqid;
 473         void *user_msgp = uap->msgp;
 474         size_t msgsz = uap->msgsz;
 475         int msgflg = uap->msgflg;
 476         int segs_needed, eval;
 477         struct msqid_ds *msqptr;
 478         struct msg *msghdr;
 479         short next;
 480
 481 #ifdef MSG_DEBUG_OK
 482         kprintf("call to msgsnd(%d, 0x%x, %d, %d)\n", msqid, user_msgp, msgsz,
 483             msgflg);
 484 #endif
 485
 486         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 487                 return (ENOSYS);
 488
 489         msqid = IPCID_TO_IX(msqid);
 490
 491         if (msqid < 0 || msqid >= msginfo.msgmni) {
 492 #ifdef MSG_DEBUG_OK
 493                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 494                     msginfo.msgmni);
 495 #endif
 496                 return(EINVAL);
 497         }
 498
 499         msqptr = &msqids[msqid];
 500         if (msqptr->msg_qbytes == 0) {
 501 #ifdef MSG_DEBUG_OK
 502                 kprintf("no such message queue id\n");
 503 #endif
 504                 return(EINVAL);
 505         }
 506         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 507 #ifdef MSG_DEBUG_OK
 508                 kprintf("wrong sequence number\n");
 509 #endif
 510                 return(EINVAL);
 511         }
 512
 513         if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_W))) {
 514 #ifdef MSG_DEBUG_OK
 515                 kprintf("requester doesn't have write access\n");
 516 #endif
 517                 return(eval);
 518         }
 519
 520         segs_needed = (msgsz + msginfo.msgssz - 1) / msginfo.msgssz;
 521 #ifdef MSG_DEBUG_OK
 522         kprintf("msgsz=%d, msgssz=%d, segs_needed=%d\n", msgsz, msginfo.msgssz,
 523             segs_needed);
 524 #endif
 525         for (;;) {
 526                 int need_more_resources = 0;
 527
 528                 /*
 529                  * check msgsz
 530                  * (inside this loop in case msg_qbytes changes while we sleep)
 531                  */
 532
 533                 if (msgsz > msqptr->msg_qbytes) {
 534 #ifdef MSG_DEBUG_OK
 535                         kprintf("msgsz > msqptr->msg_qbytes\n");
 536 #endif
 537                         return(EINVAL);
 538                 }
 539
 540                 if (msqptr->msg_perm.mode & MSG_LOCKED) {
 541 #ifdef MSG_DEBUG_OK
 542                         kprintf("msqid is locked\n");
 543 #endif
 544                         need_more_resources = 1;
 545                 }
 546                 if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes) {
 547 #ifdef MSG_DEBUG_OK
 548                         kprintf("msgsz + msg_cbytes > msg_qbytes\n");
 549 #endif
 550                         need_more_resources = 1;
 551                 }
 552                 if (segs_needed > nfree_msgmaps) {
 553 #ifdef MSG_DEBUG_OK
 554                         kprintf("segs_needed > nfree_msgmaps\n");
 555 #endif
 556                         need_more_resources = 1;
 557                 }
 558                 if (free_msghdrs == NULL) {
 559 #ifdef MSG_DEBUG_OK
 560                         kprintf("no more msghdrs\n");
 561 #endif
 562                         need_more_resources = 1;
 563                 }
 564
 565                 if (need_more_resources) {
 566                         int we_own_it;
 567
 568                         if ((msgflg & IPC_NOWAIT) != 0) {
 569 #ifdef MSG_DEBUG_OK
 570                                 kprintf("need more resources but caller doesn't want to wait\n");
 571 #endif
 572                                 return(EAGAIN);
 573                         }
 574
 575                         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0) {
 576 #ifdef MSG_DEBUG_OK
 577                                 kprintf("we don't own the msqid_ds\n");
 578 #endif
 579                                 we_own_it = 0;
 580                         } else {
 581                                 /* Force later arrivals to wait for our
 582                                    request */
 583 #ifdef MSG_DEBUG_OK
 584                                 kprintf("we own the msqid_ds\n");
 585 #endif
 586                                 msqptr->msg_perm.mode |= MSG_LOCKED;
 587                                 we_own_it = 1;
 588                         }
 589 #ifdef MSG_DEBUG_OK
 590                         kprintf("goodnight\n");
 591 #endif
 592                         eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 593 #ifdef MSG_DEBUG_OK
 594                         kprintf("good morning, eval=%d\n", eval);
 595 #endif
 596                         if (we_own_it)
 597                                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 598                         if (eval != 0) {
 599 #ifdef MSG_DEBUG_OK
 600                                 kprintf("msgsnd:  interrupted system call\n");
 601 #endif
 602                                 return(EINTR);
 603                         }
 604
 605                         /*
 606                          * Make sure that the msq queue still exists
 607                          */
 608
 609                         if (msqptr->msg_qbytes == 0) {
 610 #ifdef MSG_DEBUG_OK
 611                                 kprintf("msqid deleted\n");
 612 #endif
 613                                 return(EIDRM);
 614                         }
 615
 616                 } else {
 617 #ifdef MSG_DEBUG_OK
 618                         kprintf("got all the resources that we need\n");
 619 #endif
 620                         break;
 621                 }
 622         }
 623
 624         /*
 625          * We have the resources that we need.
 626          * Make sure!
 627          */
 628
 629         if (msqptr->msg_perm.mode & MSG_LOCKED)
 630                 panic("msg_perm.mode & MSG_LOCKED");
 631         if (segs_needed > nfree_msgmaps)
 632                 panic("segs_needed > nfree_msgmaps");
 633         if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes)
 634                 panic("msgsz + msg_cbytes > msg_qbytes");
 635         if (free_msghdrs == NULL)
 636                 panic("no more msghdrs");
 637
 638         /*
 639          * Re-lock the msqid_ds in case we page-fault when copying in the
 640          * message
 641          */
 642
 643         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0)
 644                 panic("msqid_ds is already locked");
 645         msqptr->msg_perm.mode |= MSG_LOCKED;
 646
 647         /*
 648          * Allocate a message header
 649          */
 650
 651         msghdr = free_msghdrs;
 652         free_msghdrs = msghdr->msg_next;
 653         msghdr->msg_spot = -1;
 654         msghdr->msg_ts = msgsz;
 655
 656         /*
 657          * Allocate space for the message
 658          */
 659
 660         while (segs_needed > 0) {
 661                 if (nfree_msgmaps <= 0)
 662                         panic("not enough msgmaps");
 663                 if (free_msgmaps == -1)
 664                         panic("nil free_msgmaps");
 665                 next = free_msgmaps;
 666                 if (next <= -1)
 667                         panic("next too low #1");
 668                 if (next >= msginfo.msgseg)
 669                         panic("next out of range #1");
 670 #ifdef MSG_DEBUG_OK
 671                 kprintf("allocating segment %d to message\n", next);
 672 #endif
 673                 free_msgmaps = msgmaps[next].next;
 674                 nfree_msgmaps--;
 675                 msgmaps[next].next = msghdr->msg_spot;
 676                 msghdr->msg_spot = next;
 677                 segs_needed--;
 678         }
 679
 680         /*
 681          * Copy in the message type
 682          */
 683
 684         if ((eval = copyin(user_msgp, &msghdr->msg_type,
 685             sizeof(msghdr->msg_type))) != 0) {
 686 #ifdef MSG_DEBUG_OK
 687                 kprintf("error %d copying the message type\n", eval);
 688 #endif
 689                 msg_freehdr(msghdr);
 690                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 691                 wakeup((caddr_t)msqptr);
 692                 return(eval);
 693         }
 694         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
 695
 696         /*
 697          * Validate the message type
 698          */
 699
 700         if (msghdr->msg_type < 1) {
 701                 msg_freehdr(msghdr);
 702                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 703                 wakeup((caddr_t)msqptr);
 704 #ifdef MSG_DEBUG_OK
 705                 kprintf("mtype (%d) < 1\n", msghdr->msg_type);
 706 #endif
 707                 return(EINVAL);
 708         }
 709
 710         /*
 711          * Copy in the message body
 712          */
 713
 714         next = msghdr->msg_spot;
 715         while (msgsz > 0) {
 716                 size_t tlen;
 717                 if (msgsz > msginfo.msgssz)
 718                         tlen = msginfo.msgssz;
 719                 else
 720                         tlen = msgsz;
 721                 if (next <= -1)
 722                         panic("next too low #2");
 723                 if (next >= msginfo.msgseg)
 724                         panic("next out of range #2");
 725                 if ((eval = copyin(user_msgp, &msgpool[next * msginfo.msgssz],
 726                     tlen)) != 0) {
 727 #ifdef MSG_DEBUG_OK
 728                         kprintf("error %d copying in message segment\n", eval);
 729 #endif
 730                         msg_freehdr(msghdr);
 731                         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 732                         wakeup((caddr_t)msqptr);
 733                         return(eval);
 734                 }
 735                 msgsz -= tlen;
 736                 user_msgp = (char *)user_msgp + tlen;
 737                 next = msgmaps[next].next;
 738         }
 739         if (next != -1)
 740                 panic("didn't use all the msg segments");
 741
 742         /*
 743          * We've got the message.  Unlock the msqid_ds.
 744          */
 745
 746         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 747
 748         /*
 749          * Make sure that the msqid_ds is still allocated.
 750          */
 751
 752         if (msqptr->msg_qbytes == 0) {
 753                 msg_freehdr(msghdr);
 754                 wakeup((caddr_t)msqptr);
 755                 return(EIDRM);
 756         }
 757
 758         /*
 759          * Put the message into the queue
 760          */
 761
 762         if (msqptr->msg_first == NULL) {
 763                 msqptr->msg_first = msghdr;
 764                 msqptr->msg_last = msghdr;
 765         } else {
 766                 msqptr->msg_last->msg_next = msghdr;
 767                 msqptr->msg_last = msghdr;
 768         }
 769         msqptr->msg_last->msg_next = NULL;
 770
 771         msqptr->msg_cbytes += msghdr->msg_ts;
 772         msqptr->msg_qnum++;
 773         msqptr->msg_lspid = p->p_pid;
 774         msqptr->msg_stime = time_second;
 775
 776         wakeup((caddr_t)msqptr);
 777         uap->sysmsg_result = 0;
 778         return(0);
 779 }
 780
 781 int
 782 sys_msgrcv(struct msgrcv_args *uap)
 783 {
 784         struct proc *p = curproc;
 785         int msqid = uap->msqid;
 786         void *user_msgp = uap->msgp;
 787         size_t msgsz = uap->msgsz;
 788         long msgtyp = uap->msgtyp;
 789         int msgflg = uap->msgflg;
 790         size_t len;
 791         struct msqid_ds *msqptr;
 792         struct msg *msghdr;
 793         int eval;
 794         short next;
 795
 796 #ifdef MSG_DEBUG_OK
 797         kprintf("call to msgrcv(%d, 0x%x, %d, %ld, %d)\n", msqid, user_msgp,
 798             msgsz, msgtyp, msgflg);
 799 #endif
 800
 801         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 802                 return (ENOSYS);
 803
 804         msqid = IPCID_TO_IX(msqid);
 805
 806         if (msqid < 0 || msqid >= msginfo.msgmni) {
 807 #ifdef MSG_DEBUG_OK
 808                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 809                     msginfo.msgmni);
 810 #endif
 811                 return(EINVAL);
 812         }
 813
 814         msqptr = &msqids[msqid];
 815         if (msqptr->msg_qbytes == 0) {
 816 #ifdef MSG_DEBUG_OK
 817                 kprintf("no such message queue id\n");
 818 #endif
 819                 return(EINVAL);
 820         }
 821         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 822 #ifdef MSG_DEBUG_OK
 823                 kprintf("wrong sequence number\n");
 824 #endif
 825                 return(EINVAL);
 826         }
 827
 828         if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
 829 #ifdef MSG_DEBUG_OK
 830                 kprintf("requester doesn't have read access\n");
 831 #endif
 832                 return(eval);
 833         }
 834
 835         msghdr = NULL;
 836         while (msghdr == NULL) {
 837                 if (msgtyp == 0) {
 838                         msghdr = msqptr->msg_first;
 839                         if (msghdr != NULL) {
 840                                 if (msgsz < msghdr->msg_ts &&
 841                                     (msgflg & MSG_NOERROR) == 0) {
 842 #ifdef MSG_DEBUG_OK
 843                                         kprintf("first message on the queue is too big (want %d, got %d)\n",
 844                                             msgsz, msghdr->msg_ts);
 845 #endif
 846                                         return(E2BIG);
 847                                 }
 848                                 if (msqptr->msg_first == msqptr->msg_last) {
 849                                         msqptr->msg_first = NULL;
 850                                         msqptr->msg_last = NULL;
 851                                 } else {
 852                                         msqptr->msg_first = msghdr->msg_next;
 853                                         if (msqptr->msg_first == NULL)
 854                                                 panic("msg_first/last screwed up #1");
 855                                 }
 856                         }
 857                 } else {
 858                         struct msg *previous;
 859                         struct msg **prev;
 860
 861                         previous = NULL;
 862                         prev = &(msqptr->msg_first);
 863                         while ((msghdr = *prev) != NULL) {
 864                                 /*
 865                                  * Is this message's type an exact match or is
 866                                  * this message's type less than or equal to
 867                                  * the absolute value of a negative msgtyp?
 868                                  * Note that the second half of this test can
 869                                  * NEVER be true if msgtyp is positive since
 870                                  * msg_type is always positive!
 871                                  */
 872
 873                                 if (msgtyp == msghdr->msg_type ||
 874                                     msghdr->msg_type <= -msgtyp) {
 875 #ifdef MSG_DEBUG_OK
 876                                         kprintf("found message type %d, requested %d\n",
 877                                             msghdr->msg_type, msgtyp);
 878 #endif
 879                                         if (msgsz < msghdr->msg_ts &&
 880                                             (msgflg & MSG_NOERROR) == 0) {
 881 #ifdef MSG_DEBUG_OK
 882                                                 kprintf("requested message on the queue is too big (want %d, got %d)\n",
 883                                                     msgsz, msghdr->msg_ts);
 884 #endif
 885                                                 return(E2BIG);
 886                                         }
 887                                         *prev = msghdr->msg_next;
 888                                         if (msghdr == msqptr->msg_last) {
 889                                                 if (previous == NULL) {
 890                                                         if (prev !=
 891                                                             &msqptr->msg_first)
 892                                                                 panic("msg_first/last screwed up #2");
 893                                                         msqptr->msg_first =
 894                                                             NULL;
 895                                                         msqptr->msg_last =
 896                                                             NULL;
 897                                                 } else {
 898                                                         if (prev ==
 899                                                             &msqptr->msg_first)
 900                                                                 panic("msg_first/last screwed up #3");
 901                                                         msqptr->msg_last =
 902                                                             previous;
 903                                                 }
 904                                         }
 905                                         break;
 906                                 }
 907                                 previous = msghdr;
 908                                 prev = &(msghdr->msg_next);
 909                         }
 910                 }
 911
 912                 /*
 913                  * We've either extracted the msghdr for the appropriate
 914                  * message or there isn't one.
 915                  * If there is one then bail out of this loop.
 916                  */
 917
 918                 if (msghdr != NULL)
 919                         break;
 920
 921                 /*
 922                  * Hmph!  No message found.  Does the user want to wait?
 923                  */
 924
 925                 if ((msgflg & IPC_NOWAIT) != 0) {
 926 #ifdef MSG_DEBUG_OK
 927                         kprintf("no appropriate message found (msgtyp=%d)\n",
 928                             msgtyp);
 929 #endif
 930                         /* The SVID says to return ENOMSG. */
 931 #ifdef ENOMSG
 932                         return(ENOMSG);
 933 #else
 934                         /* Unfortunately, BSD doesn't define that code yet! */
 935                         return(EAGAIN);
 936 #endif
 937                 }
 938
 939                 /*
 940                  * Wait for something to happen
 941                  */
 942
 943 #ifdef MSG_DEBUG_OK
 944                 kprintf("msgrcv:  goodnight\n");
 945 #endif
 946                 eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 947 #ifdef MSG_DEBUG_OK
 948                 kprintf("msgrcv:  good morning (eval=%d)\n", eval);
 949 #endif
 950
 951                 if (eval != 0) {
 952 #ifdef MSG_DEBUG_OK
 953                         kprintf("msgsnd:  interrupted system call\n");
 954 #endif
 955                         return(EINTR);
 956                 }
 957
 958                 /*
 959                  * Make sure that the msq queue still exists
 960                  */
 961
 962                 if (msqptr->msg_qbytes == 0 ||
 963                     msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 964 #ifdef MSG_DEBUG_OK
 965                         kprintf("msqid deleted\n");
 966 #endif
 967                         return(EIDRM);
 968                 }
 969         }
 970
 971         /*
 972          * Return the message to the user.
 973          *
 974          * First, do the bookkeeping (before we risk being interrupted).
 975          */
 976
 977         msqptr->msg_cbytes -= msghdr->msg_ts;
 978         msqptr->msg_qnum--;
 979         msqptr->msg_lrpid = p->p_pid;
 980         msqptr->msg_rtime = time_second;
 981
 982         /*
 983          * Make msgsz the actual amount that we'll be returning.
 984          * Note that this effectively truncates the message if it is too long
 985          * (since msgsz is never increased).
 986          */
 987
 988 #ifdef MSG_DEBUG_OK
 989         kprintf("found a message, msgsz=%d, msg_ts=%d\n", msgsz,
 990             msghdr->msg_ts);
 991 #endif
 992         if (msgsz > msghdr->msg_ts)
 993                 msgsz = msghdr->msg_ts;
 994
 995         /*
 996          * Return the type to the user.
 997          */
 998
 999         eval = copyout((caddr_t)&(msghdr->msg_type), user_msgp,
1000             sizeof(msghdr->msg_type));
1001         if (eval != 0) {
1002 #ifdef MSG_DEBUG_OK
1003                 kprintf("error (%d) copying out message type\n", eval);
1004 #endif
1005                 msg_freehdr(msghdr);
1006                 wakeup((caddr_t)msqptr);
1007                 return(eval);
1008         }
1009         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
1010
1011         /*
1012          * Return the segments to the user
1013          */
1014
1015         next = msghdr->msg_spot;
1016         for (len = 0; len < msgsz; len += msginfo.msgssz) {
1017                 size_t tlen;
1018
1019                 if (msgsz - len > msginfo.msgssz)
1020                         tlen = msginfo.msgssz;
1021                 else
1022                         tlen = msgsz - len;
1023                 if (next <= -1)
1024                         panic("next too low #3");
1025                 if (next >= msginfo.msgseg)
1026                         panic("next out of range #3");
1027                 eval = copyout((caddr_t)&msgpool[next * msginfo.msgssz],
1028                     user_msgp, tlen);
1029                 if (eval != 0) {
1030 #ifdef MSG_DEBUG_OK
1031                         kprintf("error (%d) copying out message segment\n",
1032                             eval);
1033 #endif
1034                         msg_freehdr(msghdr);
1035                         wakeup((caddr_t)msqptr);
1036                         return(eval);
1037                 }
1038                 user_msgp = (char *)user_msgp + tlen;
1039                 next = msgmaps[next].next;
1040         }
1041
1042         /*
1043          * Done, return the actual number of bytes copied out.
1044          */
1045
1046         msg_freehdr(msghdr);
1047         wakeup((caddr_t)msqptr);
1048         uap->sysmsg_result = msgsz;
1049         return(0);
1050 }
1051
1052 static int
1053 sysctl_msqids(SYSCTL_HANDLER_ARGS)
1054 {
1055
1056         return (SYSCTL_OUT(req, msqids,
1057             sizeof(struct msqid_ds) * msginfo.msgmni));
1058 }
1059
1060 TUNABLE_INT("kern.ipc.msgseg", &msginfo.msgseg);
1061 TUNABLE_INT("kern.ipc.msgssz", &msginfo.msgssz);
1062 TUNABLE_INT("kern.ipc.msgmni", &msginfo.msgmni);
1063
1064 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmax, CTLFLAG_RD, &msginfo.msgmax, 0, "");
1065 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmni, CTLFLAG_RD, &msginfo.msgmni, 0, "");
1066 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmnb, CTLFLAG_RD, &msginfo.msgmnb, 0, "");
1067 SYSCTL_INT(_kern_ipc, OID_AUTO, msgtql, CTLFLAG_RD, &msginfo.msgtql, 0, "");
1068 SYSCTL_INT(_kern_ipc, OID_AUTO, msgssz, CTLFLAG_RD, &msginfo.msgssz, 0, "");
1069 SYSCTL_INT(_kern_ipc, OID_AUTO, msgseg, CTLFLAG_RD, &msginfo.msgseg, 0, "");
1070 SYSCTL_PROC(_kern_ipc, OID_AUTO, msqids, CTLFLAG_RD,
1071     NULL, 0, sysctl_msqids, "", "Message queue IDs");