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/priv.h>
  31 #include <sys/msg.h>
  32 #include <sys/sysent.h>
  33 #include <sys/sysctl.h>
  34 #include <sys/malloc.h>
  35 #include <sys/jail.h>
  36
  37 static MALLOC_DEFINE(M_MSG, "msg", "SVID compatible message queues");
  38
  39 static void msginit (void *);
  40
  41 #define MSG_DEBUG
  42 #undef MSG_DEBUG_OK
  43
  44 static void msg_freehdr (struct msg *msghdr);
  45
  46 /* XXX casting to (sy_call_t *) is bogus, as usual. */
  47 static sy_call_t *msgcalls[] = {
  48         (sy_call_t *)sys_msgctl, (sy_call_t *)sys_msgget,
  49         (sy_call_t *)sys_msgsnd, (sy_call_t *)sys_msgrcv
  50 };
  51
  52 struct msg {
  53         struct  msg *msg_next;  /* next msg in the chain */
  54         long    msg_type;       /* type of this message */
  55                                 /* >0 -> type of this message */
  56                                 /* 0 -> free header */
  57         u_short msg_ts;         /* size of this message */
  58         short   msg_spot;       /* location of start of msg in buffer */
  59 };
  60
  61
  62 #ifndef MSGSSZ
  63 #define MSGSSZ  8               /* Each segment must be 2^N long */
  64 #endif
  65 #ifndef MSGSEG
  66 #define MSGSEG  2048            /* must be less than 32767 */
  67 #endif
  68 #define MSGMAX  (MSGSSZ*MSGSEG)
  69 #ifndef MSGMNB
  70 #define MSGMNB  2048            /* max # of bytes in a queue */
  71 #endif
  72 #ifndef MSGMNI
  73 #define MSGMNI  40
  74 #endif
  75 #ifndef MSGTQL
  76 #define MSGTQL  40
  77 #endif
  78
  79 /*
  80  * Based on the configuration parameters described in an SVR2 (yes, two)
  81  * config(1m) man page.
  82  *
  83  * Each message is broken up and stored in segments that are msgssz bytes
  84  * long.  For efficiency reasons, this should be a power of two.  Also,
  85  * it doesn't make sense if it is less than 8 or greater than about 256.
  86  * Consequently, msginit in kern/sysv_msg.c checks that msgssz is a power of
  87  * two between 8 and 1024 inclusive (and panic's if it isn't).
  88  */
  89 struct msginfo msginfo = {
  90                 MSGMAX,         /* max chars in a message */
  91                 MSGMNI,         /* # of message queue identifiers */
  92                 MSGMNB,         /* max chars in a queue */
  93                 MSGTQL,         /* max messages in system */
  94                 MSGSSZ,         /* size of a message segment */
  95                                 /* (must be small power of 2 greater than 4) */
  96                 MSGSEG          /* number of message segments */
  97 };
  98
  99 /*
 100  * macros to convert between msqid_ds's and msqid's.
 101  * (specific to this implementation)
 102  */
 103 #define MSQID(ix,ds)    ((ix) & 0xffff | (((ds).msg_perm.seq << 16) & 0xffff0000))
 104 #define MSQID_IX(id)    ((id) & 0xffff)
 105 #define MSQID_SEQ(id)   (((id) >> 16) & 0xffff)
 106
 107 /*
 108  * The rest of this file is specific to this particular implementation.
 109  */
 110
 111 struct msgmap {
 112         short   next;           /* next segment in buffer */
 113                                 /* -1 -> available */
 114                                 /* 0..(MSGSEG-1) -> index of next segment */
 115 };
 116
 117 #define MSG_LOCKED      01000   /* Is this msqid_ds locked? */
 118
 119 static int nfree_msgmaps;       /* # of free map entries */
 120 static short free_msgmaps;      /* head of linked list of free map entries */
 121 static struct msg *free_msghdrs;/* list of free msg headers */
 122 static char *msgpool;           /* MSGMAX byte long msg buffer pool */
 123 static struct msgmap *msgmaps;  /* MSGSEG msgmap structures */
 124 static struct msg *msghdrs;     /* MSGTQL msg headers */
 125 static struct msqid_ds *msqids; /* MSGMNI msqid_ds struct's */
 126
 127 static void
 128 msginit(void *dummy)
 129 {
 130         int i;
 131
 132         msginfo.msgmax = msginfo.msgseg * msginfo.msgssz;
 133         msgpool = kmalloc(msginfo.msgmax, M_MSG, M_WAITOK);
 134         msgmaps = kmalloc(sizeof(struct msgmap) * msginfo.msgseg, M_MSG, M_WAITOK);
 135         msghdrs = kmalloc(sizeof(struct msg) * msginfo.msgtql, M_MSG, M_WAITOK);
 136         msqids = kmalloc(sizeof(struct msqid_ds) * msginfo.msgmni, M_MSG, M_WAITOK);
 137
 138         /*
 139          * msginfo.msgssz should be a power of two for efficiency reasons.
 140          * It is also pretty silly if msginfo.msgssz is less than 8
 141          * or greater than about 256 so ...
 142          */
 143
 144         i = 8;
 145         while (i < 1024 && i != msginfo.msgssz)
 146                 i <<= 1;
 147         if (i != msginfo.msgssz) {
 148                 kprintf("msginfo.msgssz=%d (0x%x)\n", msginfo.msgssz,
 149                     msginfo.msgssz);
 150                 panic("msginfo.msgssz not a small power of 2");
 151         }
 152
 153         if (msginfo.msgseg > 32767) {
 154                 kprintf("msginfo.msgseg=%d\n", msginfo.msgseg);
 155                 panic("msginfo.msgseg > 32767");
 156         }
 157
 158         if (msgmaps == NULL)
 159                 panic("msgmaps is NULL");
 160
 161         for (i = 0; i < msginfo.msgseg; i++) {
 162                 if (i > 0)
 163                         msgmaps[i-1].next = i;
 164                 msgmaps[i].next = -1;   /* implies entry is available */
 165         }
 166         free_msgmaps = 0;
 167         nfree_msgmaps = msginfo.msgseg;
 168
 169         if (msghdrs == NULL)
 170                 panic("msghdrs is NULL");
 171
 172         for (i = 0; i < msginfo.msgtql; i++) {
 173                 msghdrs[i].msg_type = 0;
 174                 if (i > 0)
 175                         msghdrs[i-1].msg_next = &msghdrs[i];
 176                 msghdrs[i].msg_next = NULL;
 177         }
 178         free_msghdrs = &msghdrs[0];
 179
 180         if (msqids == NULL)
 181                 panic("msqids is NULL");
 182
 183         for (i = 0; i < msginfo.msgmni; i++) {
 184                 msqids[i].msg_qbytes = 0;       /* implies entry is available */
 185                 msqids[i].msg_perm.seq = 0;     /* reset to a known value */
 186                 msqids[i].msg_perm.mode = 0;
 187         }
 188 }
 189 SYSINIT(sysv_msg, SI_SUB_SYSV_MSG, SI_ORDER_FIRST, msginit, NULL)
 190
 191 /*
 192  * Entry point for all MSG calls
 193  *
 194  * msgsys_args(int which, int a2, ...) (VARARGS)
 195  *
 196  * MPALMOSTSAFE
 197  */
 198 int
 199 sys_msgsys(struct msgsys_args *uap)
 200 {
 201         struct proc *p = curproc;
 202         unsigned int which = (unsigned int)uap->which;
 203         int error;
 204
 205         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 206                 return (ENOSYS);
 207
 208         if (which >= sizeof(msgcalls)/sizeof(msgcalls[0]))
 209                 return (EINVAL);
 210         bcopy(&uap->a2, &uap->which,
 211               sizeof(struct msgsys_args) - offsetof(struct msgsys_args, a2));
 212         get_mplock();
 213         error = (*msgcalls[which])(uap);
 214         rel_mplock();
 215         return (error);
 216 }
 217
 218 static void
 219 msg_freehdr(struct msg *msghdr)
 220 {
 221         while (msghdr->msg_ts > 0) {
 222                 short next;
 223                 if (msghdr->msg_spot < 0 || msghdr->msg_spot >= msginfo.msgseg)
 224                         panic("msghdr->msg_spot out of range");
 225                 next = msgmaps[msghdr->msg_spot].next;
 226                 msgmaps[msghdr->msg_spot].next = free_msgmaps;
 227                 free_msgmaps = msghdr->msg_spot;
 228                 nfree_msgmaps++;
 229                 msghdr->msg_spot = next;
 230                 if (msghdr->msg_ts >= msginfo.msgssz)
 231                         msghdr->msg_ts -= msginfo.msgssz;
 232                 else
 233                         msghdr->msg_ts = 0;
 234         }
 235         if (msghdr->msg_spot != -1)
 236                 panic("msghdr->msg_spot != -1");
 237         msghdr->msg_next = free_msghdrs;
 238         free_msghdrs = msghdr;
 239 }
 240
 241 /*
 242  * MPALMOSTSAFE
 243  */
 244 int
 245 sys_msgctl(struct msgctl_args *uap)
 246 {
 247         struct thread *td = curthread;
 248         struct proc *p = td->td_proc;
 249         int msqid = uap->msqid;
 250         int cmd = uap->cmd;
 251         struct msqid_ds *user_msqptr = uap->buf;
 252         int rval, eval;
 253         struct msqid_ds msqbuf;
 254         struct msqid_ds *msqptr;
 255
 256 #ifdef MSG_DEBUG_OK
 257         kprintf("call to msgctl(%d, %d, 0x%x)\n", msqid, cmd, user_msqptr);
 258 #endif
 259
 260         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 261                 return (ENOSYS);
 262
 263         get_mplock();
 264         msqid = IPCID_TO_IX(msqid);
 265
 266         if (msqid < 0 || msqid >= msginfo.msgmni) {
 267 #ifdef MSG_DEBUG_OK
 268                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 269                     msginfo.msgmni);
 270 #endif
 271                 eval = EINVAL;
 272                 goto done;
 273         }
 274
 275         msqptr = &msqids[msqid];
 276
 277         if (msqptr->msg_qbytes == 0) {
 278 #ifdef MSG_DEBUG_OK
 279                 kprintf("no such msqid\n");
 280 #endif
 281                 eval = EINVAL;
 282                 goto done;
 283         }
 284         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 285 #ifdef MSG_DEBUG_OK
 286                 kprintf("wrong sequence number\n");
 287 #endif
 288                 eval = EINVAL;
 289                 goto done;
 290         }
 291
 292         rval = 0;
 293
 294         switch (cmd) {
 295         case IPC_RMID:
 296         {
 297                 struct msg *msghdr;
 298                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)) != 0)
 299                         break;
 300                 /* Free the message headers */
 301                 msghdr = msqptr->msg_first;
 302                 while (msghdr != NULL) {
 303                         struct msg *msghdr_tmp;
 304
 305                         /* Free the segments of each message */
 306                         msqptr->msg_cbytes -= msghdr->msg_ts;
 307                         msqptr->msg_qnum--;
 308                         msghdr_tmp = msghdr;
 309                         msghdr = msghdr->msg_next;
 310                         msg_freehdr(msghdr_tmp);
 311                 }
 312
 313                 if (msqptr->msg_cbytes != 0)
 314                         panic("msg_cbytes is screwed up");
 315                 if (msqptr->msg_qnum != 0)
 316                         panic("msg_qnum is screwed up");
 317
 318                 msqptr->msg_qbytes = 0; /* Mark it as free */
 319
 320                 wakeup((caddr_t)msqptr);
 321         }
 322
 323                 break;
 324
 325         case IPC_SET:
 326                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)) != 0)
 327                         break;
 328                 if ((eval = copyin(user_msqptr, &msqbuf, sizeof(msqbuf))) != 0)
 329                         break;
 330                 if (msqbuf.msg_qbytes > msqptr->msg_qbytes) {
 331                         eval = priv_check(td, PRIV_ROOT);
 332                         if (eval)
 333                                 break;
 334                 }
 335                 if (msqbuf.msg_qbytes > msginfo.msgmnb) {
 336 #ifdef MSG_DEBUG_OK
 337                         kprintf("can't increase msg_qbytes beyond %d (truncating)\n",
 338                             msginfo.msgmnb);
 339 #endif
 340                         msqbuf.msg_qbytes = msginfo.msgmnb;     /* silently restrict qbytes to system limit */
 341                 }
 342                 if (msqbuf.msg_qbytes == 0) {
 343 #ifdef MSG_DEBUG_OK
 344                         kprintf("can't reduce msg_qbytes to 0\n");
 345 #endif
 346                         eval = EINVAL;          /* non-standard errno! */
 347                         break;
 348                 }
 349                 msqptr->msg_perm.uid = msqbuf.msg_perm.uid;     /* change the owner */
 350                 msqptr->msg_perm.gid = msqbuf.msg_perm.gid;     /* change the owner */
 351                 msqptr->msg_perm.mode = (msqptr->msg_perm.mode & ~0777) |
 352                                         (msqbuf.msg_perm.mode & 0777);
 353                 msqptr->msg_qbytes = msqbuf.msg_qbytes;
 354                 msqptr->msg_ctime = time_second;
 355                 break;
 356
 357         case IPC_STAT:
 358                 if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
 359 #ifdef MSG_DEBUG_OK
 360                         kprintf("requester doesn't have read access\n");
 361 #endif
 362                         eval = EINVAL;
 363                         break;
 364                 }
 365                 eval = copyout(msqptr, user_msqptr, sizeof(struct msqid_ds));
 366                 break;
 367
 368         default:
 369 #ifdef MSG_DEBUG_OK
 370                 kprintf("invalid command %d\n", cmd);
 371 #endif
 372                 eval = EINVAL;
 373                 break;
 374         }
 375 done:
 376         rel_mplock();
 377         if (eval == 0)
 378                 uap->sysmsg_result = rval;
 379         return(eval);
 380 }
 381
 382 /*
 383  * MPALMOSTSAFE
 384  */
 385 int
 386 sys_msgget(struct msgget_args *uap)
 387 {
 388         struct proc *p = curproc;
 389         int msqid, eval;
 390         int key = uap->key;
 391         int msgflg = uap->msgflg;
 392         struct ucred *cred = p->p_ucred;
 393         struct msqid_ds *msqptr = NULL;
 394
 395 #ifdef MSG_DEBUG_OK
 396         kprintf("msgget(0x%x, 0%o)\n", key, msgflg);
 397 #endif
 398         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 399                 return (ENOSYS);
 400
 401         eval = 0;
 402         get_mplock();
 403
 404         if (key != IPC_PRIVATE) {
 405                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 406                         msqptr = &msqids[msqid];
 407                         if (msqptr->msg_qbytes != 0 &&
 408                             msqptr->msg_perm.key == key)
 409                                 break;
 410                 }
 411                 if (msqid < msginfo.msgmni) {
 412 #ifdef MSG_DEBUG_OK
 413                         kprintf("found public key\n");
 414 #endif
 415                         if ((msgflg & IPC_CREAT) && (msgflg & IPC_EXCL)) {
 416 #ifdef MSG_DEBUG_OK
 417                                 kprintf("not exclusive\n");
 418 #endif
 419                                 eval = EEXIST;
 420                                 goto done;
 421                         }
 422                         if ((eval = ipcperm(p, &msqptr->msg_perm, msgflg & 0700 ))) {
 423 #ifdef MSG_DEBUG_OK
 424                                 kprintf("requester doesn't have 0%o access\n",
 425                                     msgflg & 0700);
 426 #endif
 427                                 goto done;
 428                         }
 429                         goto done;
 430                 }
 431         }
 432
 433 #ifdef MSG_DEBUG_OK
 434         kprintf("need to allocate the msqid_ds\n");
 435 #endif
 436         if (key == IPC_PRIVATE || (msgflg & IPC_CREAT)) {
 437                 for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
 438                         /*
 439                          * Look for an unallocated and unlocked msqid_ds.
 440                          * msqid_ds's can be locked by msgsnd or msgrcv while
 441                          * they are copying the message in/out.  We can't
 442                          * re-use the entry until they release it.
 443                          */
 444                         msqptr = &msqids[msqid];
 445                         if (msqptr->msg_qbytes == 0 &&
 446                             (msqptr->msg_perm.mode & MSG_LOCKED) == 0)
 447                                 break;
 448                 }
 449                 if (msqid == msginfo.msgmni) {
 450 #ifdef MSG_DEBUG_OK
 451                         kprintf("no more msqid_ds's available\n");
 452 #endif
 453                         eval = ENOSPC;
 454                         goto done;
 455                 }
 456 #ifdef MSG_DEBUG_OK
 457                 kprintf("msqid %d is available\n", msqid);
 458 #endif
 459                 msqptr->msg_perm.key = key;
 460                 msqptr->msg_perm.cuid = cred->cr_uid;
 461                 msqptr->msg_perm.uid = cred->cr_uid;
 462                 msqptr->msg_perm.cgid = cred->cr_gid;
 463                 msqptr->msg_perm.gid = cred->cr_gid;
 464                 msqptr->msg_perm.mode = (msgflg & 0777);
 465                 /* Make sure that the returned msqid is unique */
 466                 msqptr->msg_perm.seq = (msqptr->msg_perm.seq + 1) & 0x7fff;
 467                 msqptr->msg_first = NULL;
 468                 msqptr->msg_last = NULL;
 469                 msqptr->msg_cbytes = 0;
 470                 msqptr->msg_qnum = 0;
 471                 msqptr->msg_qbytes = msginfo.msgmnb;
 472                 msqptr->msg_lspid = 0;
 473                 msqptr->msg_lrpid = 0;
 474                 msqptr->msg_stime = 0;
 475                 msqptr->msg_rtime = 0;
 476                 msqptr->msg_ctime = time_second;
 477         } else {
 478 #ifdef MSG_DEBUG_OK
 479                 kprintf("didn't find it and wasn't asked to create it\n");
 480 #endif
 481                 eval = ENOENT;
 482         }
 483
 484 done:
 485         rel_mplock();
 486         /* Construct the unique msqid */
 487         if (eval == 0)
 488                 uap->sysmsg_result = IXSEQ_TO_IPCID(msqid, msqptr->msg_perm);
 489         return(eval);
 490 }
 491
 492 /*
 493  * MPALMOSTSAFE
 494  */
 495 int
 496 sys_msgsnd(struct msgsnd_args *uap)
 497 {
 498         struct proc *p = curproc;
 499         int msqid = uap->msqid;
 500         void *user_msgp = uap->msgp;
 501         size_t msgsz = uap->msgsz;
 502         int msgflg = uap->msgflg;
 503         int segs_needed, eval;
 504         struct msqid_ds *msqptr;
 505         struct msg *msghdr;
 506         short next;
 507
 508 #ifdef MSG_DEBUG_OK
 509         kprintf("call to msgsnd(%d, 0x%x, %d, %d)\n", msqid, user_msgp, msgsz,
 510             msgflg);
 511 #endif
 512
 513         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 514                 return (ENOSYS);
 515
 516         get_mplock();
 517         msqid = IPCID_TO_IX(msqid);
 518
 519         if (msqid < 0 || msqid >= msginfo.msgmni) {
 520 #ifdef MSG_DEBUG_OK
 521                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 522                     msginfo.msgmni);
 523 #endif
 524                 eval = EINVAL;
 525                 goto done;
 526         }
 527
 528         msqptr = &msqids[msqid];
 529         if (msqptr->msg_qbytes == 0) {
 530 #ifdef MSG_DEBUG_OK
 531                 kprintf("no such message queue id\n");
 532 #endif
 533                 eval = EINVAL;
 534                 goto done;
 535         }
 536         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 537 #ifdef MSG_DEBUG_OK
 538                 kprintf("wrong sequence number\n");
 539 #endif
 540                 eval = EINVAL;
 541                 goto done;
 542         }
 543
 544         if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_W))) {
 545 #ifdef MSG_DEBUG_OK
 546                 kprintf("requester doesn't have write access\n");
 547 #endif
 548                 eval = eval;
 549                 goto done;
 550         }
 551
 552         segs_needed = (msgsz + msginfo.msgssz - 1) / msginfo.msgssz;
 553 #ifdef MSG_DEBUG_OK
 554         kprintf("msgsz=%d, msgssz=%d, segs_needed=%d\n", msgsz, msginfo.msgssz,
 555             segs_needed);
 556 #endif
 557         for (;;) {
 558                 int need_more_resources = 0;
 559
 560                 /*
 561                  * check msgsz
 562                  * (inside this loop in case msg_qbytes changes while we sleep)
 563                  */
 564
 565                 if (msgsz > msqptr->msg_qbytes) {
 566 #ifdef MSG_DEBUG_OK
 567                         kprintf("msgsz > msqptr->msg_qbytes\n");
 568 #endif
 569                         eval = EINVAL;
 570                         goto done;
 571                 }
 572
 573                 if (msqptr->msg_perm.mode & MSG_LOCKED) {
 574 #ifdef MSG_DEBUG_OK
 575                         kprintf("msqid is locked\n");
 576 #endif
 577                         need_more_resources = 1;
 578                 }
 579                 if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes) {
 580 #ifdef MSG_DEBUG_OK
 581                         kprintf("msgsz + msg_cbytes > msg_qbytes\n");
 582 #endif
 583                         need_more_resources = 1;
 584                 }
 585                 if (segs_needed > nfree_msgmaps) {
 586 #ifdef MSG_DEBUG_OK
 587                         kprintf("segs_needed > nfree_msgmaps\n");
 588 #endif
 589                         need_more_resources = 1;
 590                 }
 591                 if (free_msghdrs == NULL) {
 592 #ifdef MSG_DEBUG_OK
 593                         kprintf("no more msghdrs\n");
 594 #endif
 595                         need_more_resources = 1;
 596                 }
 597
 598                 if (need_more_resources) {
 599                         int we_own_it;
 600
 601                         if ((msgflg & IPC_NOWAIT) != 0) {
 602 #ifdef MSG_DEBUG_OK
 603                                 kprintf("need more resources but caller doesn't want to wait\n");
 604 #endif
 605                                 eval = EAGAIN;
 606                                 goto done;
 607                         }
 608
 609                         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0) {
 610 #ifdef MSG_DEBUG_OK
 611                                 kprintf("we don't own the msqid_ds\n");
 612 #endif
 613                                 we_own_it = 0;
 614                         } else {
 615                                 /* Force later arrivals to wait for our
 616                                    request */
 617 #ifdef MSG_DEBUG_OK
 618                                 kprintf("we own the msqid_ds\n");
 619 #endif
 620                                 msqptr->msg_perm.mode |= MSG_LOCKED;
 621                                 we_own_it = 1;
 622                         }
 623 #ifdef MSG_DEBUG_OK
 624                         kprintf("goodnight\n");
 625 #endif
 626                         eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 627 #ifdef MSG_DEBUG_OK
 628                         kprintf("good morning, eval=%d\n", eval);
 629 #endif
 630                         if (we_own_it)
 631                                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 632                         if (eval != 0) {
 633 #ifdef MSG_DEBUG_OK
 634                                 kprintf("msgsnd:  interrupted system call\n");
 635 #endif
 636                                 eval = EINTR;
 637                                 goto done;
 638                         }
 639
 640                         /*
 641                          * Make sure that the msq queue still exists
 642                          */
 643
 644                         if (msqptr->msg_qbytes == 0) {
 645 #ifdef MSG_DEBUG_OK
 646                                 kprintf("msqid deleted\n");
 647 #endif
 648                                 eval = EIDRM;
 649                                 goto done;
 650                         }
 651
 652                 } else {
 653 #ifdef MSG_DEBUG_OK
 654                         kprintf("got all the resources that we need\n");
 655 #endif
 656                         break;
 657                 }
 658         }
 659
 660         /*
 661          * We have the resources that we need.
 662          * Make sure!
 663          */
 664
 665         if (msqptr->msg_perm.mode & MSG_LOCKED)
 666                 panic("msg_perm.mode & MSG_LOCKED");
 667         if (segs_needed > nfree_msgmaps)
 668                 panic("segs_needed > nfree_msgmaps");
 669         if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes)
 670                 panic("msgsz + msg_cbytes > msg_qbytes");
 671         if (free_msghdrs == NULL)
 672                 panic("no more msghdrs");
 673
 674         /*
 675          * Re-lock the msqid_ds in case we page-fault when copying in the
 676          * message
 677          */
 678
 679         if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0)
 680                 panic("msqid_ds is already locked");
 681         msqptr->msg_perm.mode |= MSG_LOCKED;
 682
 683         /*
 684          * Allocate a message header
 685          */
 686
 687         msghdr = free_msghdrs;
 688         free_msghdrs = msghdr->msg_next;
 689         msghdr->msg_spot = -1;
 690         msghdr->msg_ts = msgsz;
 691
 692         /*
 693          * Allocate space for the message
 694          */
 695
 696         while (segs_needed > 0) {
 697                 if (nfree_msgmaps <= 0)
 698                         panic("not enough msgmaps");
 699                 if (free_msgmaps == -1)
 700                         panic("nil free_msgmaps");
 701                 next = free_msgmaps;
 702                 if (next <= -1)
 703                         panic("next too low #1");
 704                 if (next >= msginfo.msgseg)
 705                         panic("next out of range #1");
 706 #ifdef MSG_DEBUG_OK
 707                 kprintf("allocating segment %d to message\n", next);
 708 #endif
 709                 free_msgmaps = msgmaps[next].next;
 710                 nfree_msgmaps--;
 711                 msgmaps[next].next = msghdr->msg_spot;
 712                 msghdr->msg_spot = next;
 713                 segs_needed--;
 714         }
 715
 716         /*
 717          * Copy in the message type
 718          */
 719
 720         if ((eval = copyin(user_msgp, &msghdr->msg_type,
 721             sizeof(msghdr->msg_type))) != 0) {
 722 #ifdef MSG_DEBUG_OK
 723                 kprintf("error %d copying the message type\n", eval);
 724 #endif
 725                 msg_freehdr(msghdr);
 726                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 727                 wakeup((caddr_t)msqptr);
 728                 goto done;
 729         }
 730         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
 731
 732         /*
 733          * Validate the message type
 734          */
 735
 736         if (msghdr->msg_type < 1) {
 737                 msg_freehdr(msghdr);
 738                 msqptr->msg_perm.mode &= ~MSG_LOCKED;
 739                 wakeup((caddr_t)msqptr);
 740 #ifdef MSG_DEBUG_OK
 741                 kprintf("mtype (%d) < 1\n", msghdr->msg_type);
 742 #endif
 743                 eval = EINVAL;
 744                 goto done;
 745         }
 746
 747         /*
 748          * Copy in the message body
 749          */
 750
 751         next = msghdr->msg_spot;
 752         while (msgsz > 0) {
 753                 size_t tlen;
 754                 if (msgsz > msginfo.msgssz)
 755                         tlen = msginfo.msgssz;
 756                 else
 757                         tlen = msgsz;
 758                 if (next <= -1)
 759                         panic("next too low #2");
 760                 if (next >= msginfo.msgseg)
 761                         panic("next out of range #2");
 762                 if ((eval = copyin(user_msgp, &msgpool[next * msginfo.msgssz],
 763                     tlen)) != 0) {
 764 #ifdef MSG_DEBUG_OK
 765                         kprintf("error %d copying in message segment\n", eval);
 766 #endif
 767                         msg_freehdr(msghdr);
 768                         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 769                         wakeup((caddr_t)msqptr);
 770                         goto done;
 771                 }
 772                 msgsz -= tlen;
 773                 user_msgp = (char *)user_msgp + tlen;
 774                 next = msgmaps[next].next;
 775         }
 776         if (next != -1)
 777                 panic("didn't use all the msg segments");
 778
 779         /*
 780          * We've got the message.  Unlock the msqid_ds.
 781          */
 782
 783         msqptr->msg_perm.mode &= ~MSG_LOCKED;
 784
 785         /*
 786          * Make sure that the msqid_ds is still allocated.
 787          */
 788
 789         if (msqptr->msg_qbytes == 0) {
 790                 msg_freehdr(msghdr);
 791                 wakeup((caddr_t)msqptr);
 792                 eval = EIDRM;
 793                 goto done;
 794         }
 795
 796         /*
 797          * Put the message into the queue
 798          */
 799
 800         if (msqptr->msg_first == NULL) {
 801                 msqptr->msg_first = msghdr;
 802                 msqptr->msg_last = msghdr;
 803         } else {
 804                 msqptr->msg_last->msg_next = msghdr;
 805                 msqptr->msg_last = msghdr;
 806         }
 807         msqptr->msg_last->msg_next = NULL;
 808
 809         msqptr->msg_cbytes += msghdr->msg_ts;
 810         msqptr->msg_qnum++;
 811         msqptr->msg_lspid = p->p_pid;
 812         msqptr->msg_stime = time_second;
 813
 814         wakeup((caddr_t)msqptr);
 815         eval = 0;
 816 done:
 817         rel_mplock();
 818         if (eval == 0)
 819                 uap->sysmsg_result = 0;
 820         return (eval);
 821 }
 822
 823 /*
 824  * MPALMOSTSAFE
 825  */
 826 int
 827 sys_msgrcv(struct msgrcv_args *uap)
 828 {
 829         struct proc *p = curproc;
 830         int msqid = uap->msqid;
 831         void *user_msgp = uap->msgp;
 832         size_t msgsz = uap->msgsz;
 833         long msgtyp = uap->msgtyp;
 834         int msgflg = uap->msgflg;
 835         size_t len;
 836         struct msqid_ds *msqptr;
 837         struct msg *msghdr;
 838         int eval;
 839         short next;
 840
 841 #ifdef MSG_DEBUG_OK
 842         kprintf("call to msgrcv(%d, 0x%x, %d, %ld, %d)\n", msqid, user_msgp,
 843             msgsz, msgtyp, msgflg);
 844 #endif
 845
 846         if (!jail_sysvipc_allowed && p->p_ucred->cr_prison != NULL)
 847                 return (ENOSYS);
 848
 849         get_mplock();
 850         msqid = IPCID_TO_IX(msqid);
 851
 852         if (msqid < 0 || msqid >= msginfo.msgmni) {
 853 #ifdef MSG_DEBUG_OK
 854                 kprintf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
 855                     msginfo.msgmni);
 856 #endif
 857                 eval = EINVAL;
 858                 goto done;
 859         }
 860
 861         msqptr = &msqids[msqid];
 862         if (msqptr->msg_qbytes == 0) {
 863 #ifdef MSG_DEBUG_OK
 864                 kprintf("no such message queue id\n");
 865 #endif
 866                 eval = EINVAL;
 867                 goto done;
 868         }
 869         if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
 870 #ifdef MSG_DEBUG_OK
 871                 kprintf("wrong sequence number\n");
 872 #endif
 873                 eval = EINVAL;
 874                 goto done;
 875         }
 876
 877         if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
 878 #ifdef MSG_DEBUG_OK
 879                 kprintf("requester doesn't have read access\n");
 880 #endif
 881                 goto done;
 882         }
 883
 884         msghdr = NULL;
 885         while (msghdr == NULL) {
 886                 if (msgtyp == 0) {
 887                         msghdr = msqptr->msg_first;
 888                         if (msghdr != NULL) {
 889                                 if (msgsz < msghdr->msg_ts &&
 890                                     (msgflg & MSG_NOERROR) == 0) {
 891 #ifdef MSG_DEBUG_OK
 892                                         kprintf("first message on the queue is too big (want %d, got %d)\n",
 893                                             msgsz, msghdr->msg_ts);
 894 #endif
 895                                         eval = E2BIG;
 896                                         goto done;
 897                                 }
 898                                 if (msqptr->msg_first == msqptr->msg_last) {
 899                                         msqptr->msg_first = NULL;
 900                                         msqptr->msg_last = NULL;
 901                                 } else {
 902                                         msqptr->msg_first = msghdr->msg_next;
 903                                         if (msqptr->msg_first == NULL)
 904                                                 panic("msg_first/last screwed up #1");
 905                                 }
 906                         }
 907                 } else {
 908                         struct msg *previous;
 909                         struct msg **prev;
 910
 911                         previous = NULL;
 912                         prev = &(msqptr->msg_first);
 913                         while ((msghdr = *prev) != NULL) {
 914                                 /*
 915                                  * Is this message's type an exact match or is
 916                                  * this message's type less than or equal to
 917                                  * the absolute value of a negative msgtyp?
 918                                  * Note that the second half of this test can
 919                                  * NEVER be true if msgtyp is positive since
 920                                  * msg_type is always positive!
 921                                  */
 922
 923                                 if (msgtyp == msghdr->msg_type ||
 924                                     msghdr->msg_type <= -msgtyp) {
 925 #ifdef MSG_DEBUG_OK
 926                                         kprintf("found message type %d, requested %d\n",
 927                                             msghdr->msg_type, msgtyp);
 928 #endif
 929                                         if (msgsz < msghdr->msg_ts &&
 930                                             (msgflg & MSG_NOERROR) == 0) {
 931 #ifdef MSG_DEBUG_OK
 932                                                 kprintf("requested message on the queue is too big (want %d, got %d)\n",
 933                                                     msgsz, msghdr->msg_ts);
 934 #endif
 935                                                 eval = E2BIG;
 936                                                 goto done;
 937                                         }
 938                                         *prev = msghdr->msg_next;
 939                                         if (msghdr == msqptr->msg_last) {
 940                                                 if (previous == NULL) {
 941                                                         if (prev !=
 942                                                             &msqptr->msg_first)
 943                                                                 panic("msg_first/last screwed up #2");
 944                                                         msqptr->msg_first =
 945                                                             NULL;
 946                                                         msqptr->msg_last =
 947                                                             NULL;
 948                                                 } else {
 949                                                         if (prev ==
 950                                                             &msqptr->msg_first)
 951                                                                 panic("msg_first/last screwed up #3");
 952                                                         msqptr->msg_last =
 953                                                             previous;
 954                                                 }
 955                                         }
 956                                         break;
 957                                 }
 958                                 previous = msghdr;
 959                                 prev = &(msghdr->msg_next);
 960                         }
 961                 }
 962
 963                 /*
 964                  * We've either extracted the msghdr for the appropriate
 965                  * message or there isn't one.
 966                  * If there is one then bail out of this loop.
 967                  */
 968
 969                 if (msghdr != NULL)
 970                         break;
 971
 972                 /*
 973                  * Hmph!  No message found.  Does the user want to wait?
 974                  */
 975
 976                 if ((msgflg & IPC_NOWAIT) != 0) {
 977 #ifdef MSG_DEBUG_OK
 978                         kprintf("no appropriate message found (msgtyp=%d)\n",
 979                             msgtyp);
 980 #endif
 981                         /* The SVID says to return ENOMSG. */
 982 #ifdef ENOMSG
 983                         eval = ENOMSG;
 984 #else
 985                         /* Unfortunately, BSD doesn't define that code yet! */
 986                         eval = EAGAIN;
 987 #endif
 988                         goto done;
 989                 }
 990
 991                 /*
 992                  * Wait for something to happen
 993                  */
 994
 995 #ifdef MSG_DEBUG_OK
 996                 kprintf("msgrcv:  goodnight\n");
 997 #endif
 998                 eval = tsleep((caddr_t)msqptr, PCATCH, "msgwait", 0);
 999 #ifdef MSG_DEBUG_OK
1000                 kprintf("msgrcv:  good morning (eval=%d)\n", eval);
1001 #endif
1002
1003                 if (eval != 0) {
1004 #ifdef MSG_DEBUG_OK
1005                         kprintf("msgsnd:  interrupted system call\n");
1006 #endif
1007                         eval = EINTR;
1008                         goto done;
1009                 }
1010
1011                 /*
1012                  * Make sure that the msq queue still exists
1013                  */
1014
1015                 if (msqptr->msg_qbytes == 0 ||
1016                     msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
1017 #ifdef MSG_DEBUG_OK
1018                         kprintf("msqid deleted\n");
1019 #endif
1020                         eval = EIDRM;
1021                         goto done;
1022                 }
1023         }
1024
1025         /*
1026          * Return the message to the user.
1027          *
1028          * First, do the bookkeeping (before we risk being interrupted).
1029          */
1030
1031         msqptr->msg_cbytes -= msghdr->msg_ts;
1032         msqptr->msg_qnum--;
1033         msqptr->msg_lrpid = p->p_pid;
1034         msqptr->msg_rtime = time_second;
1035
1036         /*
1037          * Make msgsz the actual amount that we'll be returning.
1038          * Note that this effectively truncates the message if it is too long
1039          * (since msgsz is never increased).
1040          */
1041
1042 #ifdef MSG_DEBUG_OK
1043         kprintf("found a message, msgsz=%d, msg_ts=%d\n", msgsz,
1044             msghdr->msg_ts);
1045 #endif
1046         if (msgsz > msghdr->msg_ts)
1047                 msgsz = msghdr->msg_ts;
1048
1049         /*
1050          * Return the type to the user.
1051          */
1052
1053         eval = copyout((caddr_t)&(msghdr->msg_type), user_msgp,
1054             sizeof(msghdr->msg_type));
1055         if (eval != 0) {
1056 #ifdef MSG_DEBUG_OK
1057                 kprintf("error (%d) copying out message type\n", eval);
1058 #endif
1059                 msg_freehdr(msghdr);
1060                 wakeup((caddr_t)msqptr);
1061                 goto done;
1062         }
1063         user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
1064
1065         /*
1066          * Return the segments to the user
1067          */
1068
1069         next = msghdr->msg_spot;
1070         for (len = 0; len < msgsz; len += msginfo.msgssz) {
1071                 size_t tlen;
1072
1073                 if (msgsz - len > msginfo.msgssz)
1074                         tlen = msginfo.msgssz;
1075                 else
1076                         tlen = msgsz - len;
1077                 if (next <= -1)
1078                         panic("next too low #3");
1079                 if (next >= msginfo.msgseg)
1080                         panic("next out of range #3");
1081                 eval = copyout((caddr_t)&msgpool[next * msginfo.msgssz],
1082                     user_msgp, tlen);
1083                 if (eval != 0) {
1084 #ifdef MSG_DEBUG_OK
1085                         kprintf("error (%d) copying out message segment\n",
1086                             eval);
1087 #endif
1088                         msg_freehdr(msghdr);
1089                         wakeup((caddr_t)msqptr);
1090                         goto done;
1091                 }
1092                 user_msgp = (char *)user_msgp + tlen;
1093                 next = msgmaps[next].next;
1094         }
1095
1096         /*
1097          * Done, return the actual number of bytes copied out.
1098          */
1099
1100         msg_freehdr(msghdr);
1101         wakeup((caddr_t)msqptr);
1102         eval = 0;
1103 done:
1104         rel_mplock();
1105         if (eval == 0)
1106                 uap->sysmsg_result = msgsz;
1107         return(eval);
1108 }
1109
1110 static int
1111 sysctl_msqids(SYSCTL_HANDLER_ARGS)
1112 {
1113
1114         return (SYSCTL_OUT(req, msqids,
1115             sizeof(struct msqid_ds) * msginfo.msgmni));
1116 }
1117
1118 TUNABLE_INT("kern.ipc.msgseg", &msginfo.msgseg);
1119 TUNABLE_INT("kern.ipc.msgssz", &msginfo.msgssz);
1120 TUNABLE_INT("kern.ipc.msgmni", &msginfo.msgmni);
1121
1122 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmax, CTLFLAG_RD, &msginfo.msgmax, 0, "");
1123 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmni, CTLFLAG_RD, &msginfo.msgmni, 0, "");
1124 SYSCTL_INT(_kern_ipc, OID_AUTO, msgmnb, CTLFLAG_RD, &msginfo.msgmnb, 0, "");
1125 SYSCTL_INT(_kern_ipc, OID_AUTO, msgtql, CTLFLAG_RD, &msginfo.msgtql, 0, "");
1126 SYSCTL_INT(_kern_ipc, OID_AUTO, msgssz, CTLFLAG_RD, &msginfo.msgssz, 0, "");
1127 SYSCTL_INT(_kern_ipc, OID_AUTO, msgseg, CTLFLAG_RD, &msginfo.msgseg, 0, "");
1128 SYSCTL_PROC(_kern_ipc, OID_AUTO, msqids, CTLFLAG_RD,
1129     NULL, 0, sysctl_msqids, "", "Message queue IDs");