Home | History | Annotate | Line # | Download | only in kern
sysv_msg.c revision 1.23
      1 /*	$NetBSD: sysv_msg.c,v 1.23 1998/10/19 22:19:27 tron Exp $	*/
      2 
      3 /*
      4  * Implementation of SVID messages
      5  *
      6  * Author:  Daniel Boulet
      7  *
      8  * Copyright 1993 Daniel Boulet and RTMX Inc.
      9  *
     10  * This system call was implemented by Daniel Boulet under contract from RTMX.
     11  *
     12  * Redistribution and use in source forms, with and without modification,
     13  * are permitted provided that this entire comment appears intact.
     14  *
     15  * Redistribution in binary form may occur without any restrictions.
     16  * Obviously, it would be nice if you gave credit where credit is due
     17  * but requiring it would be too onerous.
     18  *
     19  * This software is provided ``AS IS'' without any warranties of any kind.
     20  */
     21 
     22 #include "opt_sysv.h"
     23 
     24 #include <sys/param.h>
     25 #include <sys/systm.h>
     26 #include <sys/kernel.h>
     27 #include <sys/proc.h>
     28 #include <sys/msg.h>
     29 #include <sys/malloc.h>
     30 
     31 #include <sys/mount.h>
     32 #include <sys/syscallargs.h>
     33 
     34 #define MSG_DEBUG
     35 #undef MSG_DEBUG_OK
     36 
     37 #ifdef MSG_DEBUG_OK
     38 #define MSG_PRINTF(a)	printf a
     39 #else
     40 #define MSG_PRINTF(a)
     41 #endif
     42 
     43 int nfree_msgmaps;		/* # of free map entries */
     44 short free_msgmaps;		/* head of linked list of free map entries */
     45 struct msg *free_msghdrs;	/* list of free msg headers */
     46 
     47 static void msg_freehdr __P((struct msg *));
     48 
     49 void
     50 msginit()
     51 {
     52 	register int i;
     53 
     54 	/*
     55 	 * msginfo.msgssz should be a power of two for efficiency reasons.
     56 	 * It is also pretty silly if msginfo.msgssz is less than 8
     57 	 * or greater than about 256 so ...
     58 	 */
     59 
     60 	i = 8;
     61 	while (i < 1024 && i != msginfo.msgssz)
     62 		i <<= 1;
     63     	if (i != msginfo.msgssz) {
     64 		MSG_PRINTF(("msginfo.msgssz=%d (0x%x)\n", msginfo.msgssz,
     65 		    msginfo.msgssz));
     66 		panic("msginfo.msgssz not a small power of 2");
     67 	}
     68 
     69 	if (msginfo.msgseg > 32767) {
     70 		MSG_PRINTF(("msginfo.msgseg=%d\n", msginfo.msgseg));
     71 		panic("msginfo.msgseg > 32767");
     72 	}
     73 
     74 	if (msgmaps == NULL)
     75 		panic("msgmaps is NULL");
     76 
     77 	for (i = 0; i < msginfo.msgseg; i++) {
     78 		if (i > 0)
     79 			msgmaps[i-1].next = i;
     80 		msgmaps[i].next = -1;	/* implies entry is available */
     81 	}
     82 	free_msgmaps = 0;
     83 	nfree_msgmaps = msginfo.msgseg;
     84 
     85 	if (msghdrs == NULL)
     86 		panic("msghdrs is NULL");
     87 
     88 	for (i = 0; i < msginfo.msgtql; i++) {
     89 		msghdrs[i].msg_type = 0;
     90 		if (i > 0)
     91 			msghdrs[i-1].msg_next = &msghdrs[i];
     92 		msghdrs[i].msg_next = NULL;
     93     	}
     94 	free_msghdrs = &msghdrs[0];
     95 
     96 	if (msqids == NULL)
     97 		panic("msqids is NULL");
     98 
     99 	for (i = 0; i < msginfo.msgmni; i++) {
    100 		msqids[i].msg_qbytes = 0;	/* implies entry is available */
    101 		msqids[i].msg_perm.seq = 0;	/* reset to a known value */
    102 	}
    103 }
    104 
    105 static void
    106 msg_freehdr(msghdr)
    107 	struct msg *msghdr;
    108 {
    109 	while (msghdr->msg_ts > 0) {
    110 		short next;
    111 		if (msghdr->msg_spot < 0 || msghdr->msg_spot >= msginfo.msgseg)
    112 			panic("msghdr->msg_spot out of range");
    113 		next = msgmaps[msghdr->msg_spot].next;
    114 		msgmaps[msghdr->msg_spot].next = free_msgmaps;
    115 		free_msgmaps = msghdr->msg_spot;
    116 		nfree_msgmaps++;
    117 		msghdr->msg_spot = next;
    118 		if (msghdr->msg_ts >= msginfo.msgssz)
    119 			msghdr->msg_ts -= msginfo.msgssz;
    120 		else
    121 			msghdr->msg_ts = 0;
    122 	}
    123 	if (msghdr->msg_spot != -1)
    124 		panic("msghdr->msg_spot != -1");
    125 	msghdr->msg_next = free_msghdrs;
    126 	free_msghdrs = msghdr;
    127 }
    128 
    129 int
    130 sys_msgctl(p, v, retval)
    131 	struct proc *p;
    132 	void *v;
    133 	register_t *retval;
    134 {
    135 	register struct sys_msgctl_args /* {
    136 		syscallarg(int) msqid;
    137 		syscallarg(int) cmd;
    138 		syscallarg(struct msqid_ds *) buf;
    139 	} */ *uap = v;
    140 	int msqid = SCARG(uap, msqid);
    141 	int cmd = SCARG(uap, cmd);
    142 	struct msqid_ds *user_msqptr = SCARG(uap, buf);
    143 	struct ucred *cred = p->p_ucred;
    144 	int rval, eval;
    145 	struct msqid_ds msqbuf;
    146 	register struct msqid_ds *msqptr;
    147 
    148 	MSG_PRINTF(("call to msgctl(%d, %d, %p)\n", msqid, cmd, user_msqptr));
    149 
    150 	msqid = IPCID_TO_IX(msqid);
    151 
    152 	if (msqid < 0 || msqid >= msginfo.msgmni) {
    153 		MSG_PRINTF(("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
    154 		    msginfo.msgmni));
    155 		return(EINVAL);
    156 	}
    157 
    158 	msqptr = &msqids[msqid];
    159 
    160 	if (msqptr->msg_qbytes == 0) {
    161 		MSG_PRINTF(("no such msqid\n"));
    162 		return(EINVAL);
    163 	}
    164 	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(SCARG(uap, msqid))) {
    165 		MSG_PRINTF(("wrong sequence number\n"));
    166 		return(EINVAL);
    167 	}
    168 
    169 	eval = 0;
    170 	rval = 0;
    171 
    172 	switch (cmd) {
    173 
    174 	case IPC_RMID:
    175 	{
    176 		struct msg *msghdr;
    177 		if ((eval = ipcperm(cred, &msqptr->msg_perm, IPC_M)) != 0)
    178 			return(eval);
    179 		/* Free the message headers */
    180 		msghdr = msqptr->msg_first;
    181 		while (msghdr != NULL) {
    182 			struct msg *msghdr_tmp;
    183 
    184 			/* Free the segments of each message */
    185 			msqptr->msg_cbytes -= msghdr->msg_ts;
    186 			msqptr->msg_qnum--;
    187 			msghdr_tmp = msghdr;
    188 			msghdr = msghdr->msg_next;
    189 			msg_freehdr(msghdr_tmp);
    190 		}
    191 
    192 		if (msqptr->msg_cbytes != 0)
    193 			panic("msg_cbytes is screwed up");
    194 		if (msqptr->msg_qnum != 0)
    195 			panic("msg_qnum is screwed up");
    196 
    197 		msqptr->msg_qbytes = 0;	/* Mark it as free */
    198 
    199 		wakeup((caddr_t)msqptr);
    200 	}
    201 
    202 		break;
    203 
    204 	case IPC_SET:
    205 		if ((eval = ipcperm(cred, &msqptr->msg_perm, IPC_M)))
    206 			return(eval);
    207 		if ((eval = copyin(user_msqptr, &msqbuf, sizeof(msqbuf))) != 0)
    208 			return(eval);
    209 		if (msqbuf.msg_qbytes > msqptr->msg_qbytes && cred->cr_uid != 0)
    210 			return(EPERM);
    211 		if (msqbuf.msg_qbytes > msginfo.msgmnb) {
    212 			MSG_PRINTF(("can't increase msg_qbytes beyond %d (truncating)\n",
    213 			    msginfo.msgmnb));
    214 			msqbuf.msg_qbytes = msginfo.msgmnb;	/* silently restrict qbytes to system limit */
    215 		}
    216 		if (msqbuf.msg_qbytes == 0) {
    217 			MSG_PRINTF(("can't reduce msg_qbytes to 0\n"));
    218 			return(EINVAL);		/* non-standard errno! */
    219 		}
    220 		msqptr->msg_perm.uid = msqbuf.msg_perm.uid;	/* change the owner */
    221 		msqptr->msg_perm.gid = msqbuf.msg_perm.gid;	/* change the owner */
    222 		msqptr->msg_perm.mode = (msqptr->msg_perm.mode & ~0777) |
    223 		    (msqbuf.msg_perm.mode & 0777);
    224 		msqptr->msg_qbytes = msqbuf.msg_qbytes;
    225 		msqptr->msg_ctime = time.tv_sec;
    226 		break;
    227 
    228 	case IPC_STAT:
    229 		if ((eval = ipcperm(cred, &msqptr->msg_perm, IPC_R))) {
    230 			MSG_PRINTF(("requester doesn't have read access\n"));
    231 			return(eval);
    232 		}
    233 		eval = copyout((caddr_t)msqptr, user_msqptr,
    234 		    sizeof(struct msqid_ds));
    235 		break;
    236 
    237 	default:
    238 		MSG_PRINTF(("invalid command %d\n", cmd));
    239 		return(EINVAL);
    240 	}
    241 
    242 	if (eval == 0)
    243 		*retval = rval;
    244 	return(eval);
    245 }
    246 
    247 int
    248 sys_msgget(p, v, retval)
    249 	struct proc *p;
    250 	void *v;
    251 	register_t *retval;
    252 {
    253 	register struct sys_msgget_args /* {
    254 		syscallarg(key_t) key;
    255 		syscallarg(int) msgflg;
    256 	} */ *uap = v;
    257 	int msqid, eval;
    258 	int key = SCARG(uap, key);
    259 	int msgflg = SCARG(uap, msgflg);
    260 	struct ucred *cred = p->p_ucred;
    261 	register struct msqid_ds *msqptr = NULL;
    262 
    263 	MSG_PRINTF(("msgget(0x%x, 0%o)\n", key, msgflg));
    264 
    265 	if (key != IPC_PRIVATE) {
    266 		for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
    267 			msqptr = &msqids[msqid];
    268 			if (msqptr->msg_qbytes != 0 &&
    269 			    msqptr->msg_perm.key == key)
    270 				break;
    271 		}
    272 		if (msqid < msginfo.msgmni) {
    273 			MSG_PRINTF(("found public key\n"));
    274 			if ((msgflg & IPC_CREAT) && (msgflg & IPC_EXCL)) {
    275 				MSG_PRINTF(("not exclusive\n"));
    276 				return(EEXIST);
    277 			}
    278 			if ((eval = ipcperm(cred, &msqptr->msg_perm, msgflg & 0700 ))) {
    279 				MSG_PRINTF(("requester doesn't have 0%o access\n",
    280 				    msgflg & 0700));
    281 				return(eval);
    282 			}
    283 			goto found;
    284 		}
    285 	}
    286 
    287 	MSG_PRINTF(("need to allocate the msqid_ds\n"));
    288 	if (key == IPC_PRIVATE || (msgflg & IPC_CREAT)) {
    289 		for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
    290 			/*
    291 			 * Look for an unallocated and unlocked msqid_ds.
    292 			 * msqid_ds's can be locked by msgsnd or msgrcv while
    293 			 * they are copying the message in/out.  We can't
    294 			 * re-use the entry until they release it.
    295 			 */
    296 			msqptr = &msqids[msqid];
    297 			if (msqptr->msg_qbytes == 0 &&
    298 			    (msqptr->msg_perm.mode & MSG_LOCKED) == 0)
    299 				break;
    300 		}
    301 		if (msqid == msginfo.msgmni) {
    302 			MSG_PRINTF(("no more msqid_ds's available\n"));
    303 			return(ENOSPC);
    304 		}
    305 		MSG_PRINTF(("msqid %d is available\n", msqid));
    306 		msqptr->msg_perm.key = key;
    307 		msqptr->msg_perm.cuid = cred->cr_uid;
    308 		msqptr->msg_perm.uid = cred->cr_uid;
    309 		msqptr->msg_perm.cgid = cred->cr_gid;
    310 		msqptr->msg_perm.gid = cred->cr_gid;
    311 		msqptr->msg_perm.mode = (msgflg & 0777);
    312 		/* Make sure that the returned msqid is unique */
    313 		msqptr->msg_perm.seq++;
    314 		msqptr->msg_first = NULL;
    315 		msqptr->msg_last = NULL;
    316 		msqptr->msg_cbytes = 0;
    317 		msqptr->msg_qnum = 0;
    318 		msqptr->msg_qbytes = msginfo.msgmnb;
    319 		msqptr->msg_lspid = 0;
    320 		msqptr->msg_lrpid = 0;
    321 		msqptr->msg_stime = 0;
    322 		msqptr->msg_rtime = 0;
    323 		msqptr->msg_ctime = time.tv_sec;
    324 	} else {
    325 		MSG_PRINTF(("didn't find it and wasn't asked to create it\n"));
    326 		return(ENOENT);
    327 	}
    328 
    329 found:
    330 	/* Construct the unique msqid */
    331 	*retval = IXSEQ_TO_IPCID(msqid, msqptr->msg_perm);
    332 	return(0);
    333 }
    334 
    335 int
    336 sys_msgsnd(p, v, retval)
    337 	struct proc *p;
    338 	void *v;
    339 	register_t *retval;
    340 {
    341 	register struct sys_msgsnd_args /* {
    342 		syscallarg(int) msqid;
    343 		syscallarg(const void *) msgp;
    344 		syscallarg(size_t) msgsz;
    345 		syscallarg(int) msgflg;
    346 	} */ *uap = v;
    347 	int msqid = SCARG(uap, msqid);
    348 	const char *user_msgp = SCARG(uap, msgp);
    349 	size_t msgsz = SCARG(uap, msgsz);
    350 	int msgflg = SCARG(uap, msgflg);
    351 	int segs_needed, eval;
    352 	struct ucred *cred = p->p_ucred;
    353 	register struct msqid_ds *msqptr;
    354 	register struct msg *msghdr;
    355 	short next;
    356 
    357 	MSG_PRINTF(("call to msgsnd(%d, %p, %d, %d)\n", msqid, user_msgp, msgsz,
    358 	    msgflg));
    359 
    360 	msqid = IPCID_TO_IX(msqid);
    361 
    362 	if (msqid < 0 || msqid >= msginfo.msgmni) {
    363 		MSG_PRINTF(("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
    364 		    msginfo.msgmni));
    365 		return(EINVAL);
    366 	}
    367 
    368 	msqptr = &msqids[msqid];
    369 	if (msqptr->msg_qbytes == 0) {
    370 		MSG_PRINTF(("no such message queue id\n"));
    371 		return(EINVAL);
    372 	}
    373 	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(SCARG(uap, msqid))) {
    374 		MSG_PRINTF(("wrong sequence number\n"));
    375 		return(EINVAL);
    376 	}
    377 
    378 	if ((eval = ipcperm(cred, &msqptr->msg_perm, IPC_W))) {
    379 		MSG_PRINTF(("requester doesn't have write access\n"));
    380 		return(eval);
    381 	}
    382 
    383 	segs_needed = (msgsz + msginfo.msgssz - 1) / msginfo.msgssz;
    384 	MSG_PRINTF(("msgsz=%d, msgssz=%d, segs_needed=%d\n", msgsz, msginfo.msgssz,
    385 	    segs_needed));
    386 	for (;;) {
    387 		int need_more_resources = 0;
    388 
    389 		/*
    390 		 * check msgsz [cannot be negative since it is unsigned]
    391 		 * (inside this loop in case msg_qbytes changes while we sleep)
    392 		 */
    393 
    394 		if (msgsz > msqptr->msg_qbytes) {
    395 			MSG_PRINTF(("msgsz > msqptr->msg_qbytes\n"));
    396 			return(EINVAL);
    397 		}
    398 
    399 		if (msqptr->msg_perm.mode & MSG_LOCKED) {
    400 			MSG_PRINTF(("msqid is locked\n"));
    401 			need_more_resources = 1;
    402 		}
    403 		if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes) {
    404 			MSG_PRINTF(("msgsz + msg_cbytes > msg_qbytes\n"));
    405 			need_more_resources = 1;
    406 		}
    407 		if (segs_needed > nfree_msgmaps) {
    408 			MSG_PRINTF(("segs_needed > nfree_msgmaps\n"));
    409 			need_more_resources = 1;
    410 		}
    411 		if (free_msghdrs == NULL) {
    412 			MSG_PRINTF(("no more msghdrs\n"));
    413 			need_more_resources = 1;
    414 		}
    415 
    416 		if (need_more_resources) {
    417 			int we_own_it;
    418 
    419 			if ((msgflg & IPC_NOWAIT) != 0) {
    420 				MSG_PRINTF(("need more resources but caller doesn't want to wait\n"));
    421 				return(EAGAIN);
    422 			}
    423 
    424 			if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0) {
    425 				MSG_PRINTF(("we don't own the msqid_ds\n"));
    426 				we_own_it = 0;
    427 			} else {
    428 				/* Force later arrivals to wait for our
    429 				   request */
    430 				MSG_PRINTF(("we own the msqid_ds\n"));
    431 				msqptr->msg_perm.mode |= MSG_LOCKED;
    432 				we_own_it = 1;
    433 			}
    434 			MSG_PRINTF(("goodnight\n"));
    435 			eval = tsleep((caddr_t)msqptr, (PZERO - 4) | PCATCH,
    436 			    "msgwait", 0);
    437 			MSG_PRINTF(("good morning, eval=%d\n", eval));
    438 			if (we_own_it)
    439 				msqptr->msg_perm.mode &= ~MSG_LOCKED;
    440 			if (eval != 0) {
    441 				MSG_PRINTF(("msgsnd:  interrupted system call\n"));
    442 				return(EINTR);
    443 			}
    444 
    445 			/*
    446 			 * Make sure that the msq queue still exists
    447 			 */
    448 
    449 			if (msqptr->msg_qbytes == 0) {
    450 				MSG_PRINTF(("msqid deleted\n"));
    451 				/* The SVID says to return EIDRM. */
    452 #ifdef EIDRM
    453 				return(EIDRM);
    454 #else
    455 				/* Unfortunately, BSD doesn't define that code
    456 				   yet! */
    457 				return(EINVAL);
    458 #endif
    459 			}
    460 
    461 		} else {
    462 			MSG_PRINTF(("got all the resources that we need\n"));
    463 			break;
    464 		}
    465 	}
    466 
    467 	/*
    468 	 * We have the resources that we need.
    469 	 * Make sure!
    470 	 */
    471 
    472 	if (msqptr->msg_perm.mode & MSG_LOCKED)
    473 		panic("msg_perm.mode & MSG_LOCKED");
    474 	if (segs_needed > nfree_msgmaps)
    475 		panic("segs_needed > nfree_msgmaps");
    476 	if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes)
    477 		panic("msgsz + msg_cbytes > msg_qbytes");
    478 	if (free_msghdrs == NULL)
    479 		panic("no more msghdrs");
    480 
    481 	/*
    482 	 * Re-lock the msqid_ds in case we page-fault when copying in the
    483 	 * message
    484 	 */
    485 
    486 	if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0)
    487 		panic("msqid_ds is already locked");
    488 	msqptr->msg_perm.mode |= MSG_LOCKED;
    489 
    490 	/*
    491 	 * Allocate a message header
    492 	 */
    493 
    494 	msghdr = free_msghdrs;
    495 	free_msghdrs = msghdr->msg_next;
    496 	msghdr->msg_spot = -1;
    497 	msghdr->msg_ts = msgsz;
    498 
    499 	/*
    500 	 * Allocate space for the message
    501 	 */
    502 
    503 	while (segs_needed > 0) {
    504 		if (nfree_msgmaps <= 0)
    505 			panic("not enough msgmaps");
    506 		if (free_msgmaps == -1)
    507 			panic("nil free_msgmaps");
    508 		next = free_msgmaps;
    509 		if (next <= -1)
    510 			panic("next too low #1");
    511 		if (next >= msginfo.msgseg)
    512 			panic("next out of range #1");
    513 		MSG_PRINTF(("allocating segment %d to message\n", next));
    514 		free_msgmaps = msgmaps[next].next;
    515 		nfree_msgmaps--;
    516 		msgmaps[next].next = msghdr->msg_spot;
    517 		msghdr->msg_spot = next;
    518 		segs_needed--;
    519 	}
    520 
    521 	/*
    522 	 * Copy in the message type
    523 	 */
    524 
    525 	if ((eval = copyin(user_msgp, &msghdr->msg_type,
    526 	    sizeof(msghdr->msg_type))) != 0) {
    527 		MSG_PRINTF(("error %d copying the message type\n", eval));
    528 		msg_freehdr(msghdr);
    529 		msqptr->msg_perm.mode &= ~MSG_LOCKED;
    530 		wakeup((caddr_t)msqptr);
    531 		return(eval);
    532 	}
    533 	user_msgp += sizeof(msghdr->msg_type);
    534 
    535 	/*
    536 	 * Validate the message type
    537 	 */
    538 
    539 	if (msghdr->msg_type < 1) {
    540 		msg_freehdr(msghdr);
    541 		msqptr->msg_perm.mode &= ~MSG_LOCKED;
    542 		wakeup((caddr_t)msqptr);
    543 		MSG_PRINTF(("mtype (%d) < 1\n", msghdr->msg_type));
    544 		return(EINVAL);
    545 	}
    546 
    547 	/*
    548 	 * Copy in the message body
    549 	 */
    550 
    551 	next = msghdr->msg_spot;
    552 	while (msgsz > 0) {
    553 		size_t tlen;
    554 		if (msgsz > msginfo.msgssz)
    555 			tlen = msginfo.msgssz;
    556 		else
    557 			tlen = msgsz;
    558 		if (next <= -1)
    559 			panic("next too low #2");
    560 		if (next >= msginfo.msgseg)
    561 			panic("next out of range #2");
    562 		if ((eval = copyin(user_msgp, &msgpool[next * msginfo.msgssz],
    563 		    tlen)) != 0) {
    564 			MSG_PRINTF(("error %d copying in message segment\n", eval));
    565 			msg_freehdr(msghdr);
    566 			msqptr->msg_perm.mode &= ~MSG_LOCKED;
    567 			wakeup((caddr_t)msqptr);
    568 			return(eval);
    569 		}
    570 		msgsz -= tlen;
    571 		user_msgp += tlen;
    572 		next = msgmaps[next].next;
    573 	}
    574 	if (next != -1)
    575 		panic("didn't use all the msg segments");
    576 
    577 	/*
    578 	 * We've got the message.  Unlock the msqid_ds.
    579 	 */
    580 
    581 	msqptr->msg_perm.mode &= ~MSG_LOCKED;
    582 
    583 	/*
    584 	 * Make sure that the msqid_ds is still allocated.
    585 	 */
    586 
    587 	if (msqptr->msg_qbytes == 0) {
    588 		msg_freehdr(msghdr);
    589 		wakeup((caddr_t)msqptr);
    590 		/* The SVID says to return EIDRM. */
    591 #ifdef EIDRM
    592 		return(EIDRM);
    593 #else
    594 		/* Unfortunately, BSD doesn't define that code yet! */
    595 		return(EINVAL);
    596 #endif
    597 	}
    598 
    599 	/*
    600 	 * Put the message into the queue
    601 	 */
    602 
    603 	if (msqptr->msg_first == NULL) {
    604 		msqptr->msg_first = msghdr;
    605 		msqptr->msg_last = msghdr;
    606 	} else {
    607 		msqptr->msg_last->msg_next = msghdr;
    608 		msqptr->msg_last = msghdr;
    609 	}
    610 	msqptr->msg_last->msg_next = NULL;
    611 
    612 	msqptr->msg_cbytes += msghdr->msg_ts;
    613 	msqptr->msg_qnum++;
    614 	msqptr->msg_lspid = p->p_pid;
    615 	msqptr->msg_stime = time.tv_sec;
    616 
    617 	wakeup((caddr_t)msqptr);
    618 	*retval = 0;
    619 	return(0);
    620 }
    621 
    622 int
    623 sys_msgrcv(p, v, retval)
    624 	struct proc *p;
    625 	void *v;
    626 	register_t *retval;
    627 {
    628 	register struct sys_msgrcv_args /* {
    629 		syscallarg(int) msqid;
    630 		syscallarg(void *) msgp;
    631 		syscallarg(size_t) msgsz;
    632 		syscallarg(long) msgtyp;
    633 		syscallarg(int) msgflg;
    634 	} */ *uap = v;
    635 	int msqid = SCARG(uap, msqid);
    636 	char *user_msgp = SCARG(uap, msgp);
    637 	size_t msgsz = SCARG(uap, msgsz);
    638 	long msgtyp = SCARG(uap, msgtyp);
    639 	int msgflg = SCARG(uap, msgflg);
    640 	size_t len;
    641 	struct ucred *cred = p->p_ucred;
    642 	register struct msqid_ds *msqptr;
    643 	register struct msg *msghdr;
    644 	int eval;
    645 	short next;
    646 
    647 	MSG_PRINTF(("call to msgrcv(%d, %p, %d, %ld, %d)\n", msqid, user_msgp,
    648 	    msgsz, msgtyp, msgflg));
    649 
    650 	msqid = IPCID_TO_IX(msqid);
    651 
    652 	if (msqid < 0 || msqid >= msginfo.msgmni) {
    653 		MSG_PRINTF(("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
    654 		    msginfo.msgmni));
    655 		return(EINVAL);
    656 	}
    657 
    658 	msqptr = &msqids[msqid];
    659 	if (msqptr->msg_qbytes == 0) {
    660 		MSG_PRINTF(("no such message queue id\n"));
    661 		return(EINVAL);
    662 	}
    663 	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(SCARG(uap, msqid))) {
    664 		MSG_PRINTF(("wrong sequence number\n"));
    665 		return(EINVAL);
    666 	}
    667 
    668 	if ((eval = ipcperm(cred, &msqptr->msg_perm, IPC_R))) {
    669 		MSG_PRINTF(("requester doesn't have read access\n"));
    670 		return(eval);
    671 	}
    672 
    673 #if 0
    674 	/* cannot happen, msgsz is unsigned */
    675 	if (msgsz < 0) {
    676 		MSG_PRINTF(("msgsz < 0\n"));
    677 		return(EINVAL);
    678 	}
    679 #endif
    680 
    681 	msghdr = NULL;
    682 	while (msghdr == NULL) {
    683 		if (msgtyp == 0) {
    684 			msghdr = msqptr->msg_first;
    685 			if (msghdr != NULL) {
    686 				if (msgsz < msghdr->msg_ts &&
    687 				    (msgflg & MSG_NOERROR) == 0) {
    688 					MSG_PRINTF(("first message on the queue is too big (want %d, got %d)\n",
    689 					    msgsz, msghdr->msg_ts));
    690 					return(E2BIG);
    691 				}
    692 				if (msqptr->msg_first == msqptr->msg_last) {
    693 					msqptr->msg_first = NULL;
    694 					msqptr->msg_last = NULL;
    695 				} else {
    696 					msqptr->msg_first = msghdr->msg_next;
    697 					if (msqptr->msg_first == NULL)
    698 						panic("msg_first/last screwed up #1");
    699 				}
    700 			}
    701 		} else {
    702 			struct msg *previous;
    703 			struct msg **prev;
    704 
    705 			for (previous = NULL, prev = &msqptr->msg_first;
    706 			     (msghdr = *prev) != NULL;
    707 			     previous = msghdr, prev = &msghdr->msg_next) {
    708 				/*
    709 				 * Is this message's type an exact match or is
    710 				 * this message's type less than or equal to
    711 				 * the absolute value of a negative msgtyp?
    712 				 * Note that the second half of this test can
    713 				 * NEVER be true if msgtyp is positive since
    714 				 * msg_type is always positive!
    715 				 */
    716 
    717 				if (msgtyp == msghdr->msg_type ||
    718 				    msghdr->msg_type <= -msgtyp) {
    719 					MSG_PRINTF(("found message type %d, requested %d\n",
    720 					    msghdr->msg_type, msgtyp));
    721 					if (msgsz < msghdr->msg_ts &&
    722 					    (msgflg & MSG_NOERROR) == 0) {
    723 						MSG_PRINTF(("requested message on the queue is too big (want %d, got %d)\n",
    724 						    msgsz, msghdr->msg_ts));
    725 						return(E2BIG);
    726 					}
    727 					*prev = msghdr->msg_next;
    728 					if (msghdr == msqptr->msg_last) {
    729 						if (previous == NULL) {
    730 							if (prev !=
    731 							    &msqptr->msg_first)
    732 								panic("msg_first/last screwed up #2");
    733 							msqptr->msg_first =
    734 							    NULL;
    735 							msqptr->msg_last =
    736 							    NULL;
    737 						} else {
    738 							if (prev ==
    739 							    &msqptr->msg_first)
    740 								panic("msg_first/last screwed up #3");
    741 							msqptr->msg_last =
    742 							    previous;
    743 						}
    744 					}
    745 					break;
    746 				}
    747 			}
    748 		}
    749 
    750 		/*
    751 		 * We've either extracted the msghdr for the appropriate
    752 		 * message or there isn't one.
    753 		 * If there is one then bail out of this loop.
    754 		 */
    755 
    756 		if (msghdr != NULL)
    757 			break;
    758 
    759 		/*
    760 		 * Hmph!  No message found.  Does the user want to wait?
    761 		 */
    762 
    763 		if ((msgflg & IPC_NOWAIT) != 0) {
    764 			MSG_PRINTF(("no appropriate message found (msgtyp=%d)\n",
    765 			    msgtyp));
    766 			/* The SVID says to return ENOMSG. */
    767 #ifdef ENOMSG
    768 			return(ENOMSG);
    769 #else
    770 			/* Unfortunately, BSD doesn't define that code yet! */
    771 			return(EAGAIN);
    772 #endif
    773 		}
    774 
    775 		/*
    776 		 * Wait for something to happen
    777 		 */
    778 
    779 		MSG_PRINTF(("msgrcv:  goodnight\n"));
    780 		eval = tsleep((caddr_t)msqptr, (PZERO - 4) | PCATCH, "msgwait",
    781 		    0);
    782 		MSG_PRINTF(("msgrcv:  good morning (eval=%d)\n", eval));
    783 
    784 		if (eval != 0) {
    785 			MSG_PRINTF(("msgsnd:  interrupted system call\n"));
    786 			return(EINTR);
    787 		}
    788 
    789 		/*
    790 		 * Make sure that the msq queue still exists
    791 		 */
    792 
    793 		if (msqptr->msg_qbytes == 0 ||
    794 		    msqptr->msg_perm.seq != IPCID_TO_SEQ(SCARG(uap, msqid))) {
    795 			MSG_PRINTF(("msqid deleted\n"));
    796 			/* The SVID says to return EIDRM. */
    797 #ifdef EIDRM
    798 			return(EIDRM);
    799 #else
    800 			/* Unfortunately, BSD doesn't define that code yet! */
    801 			return(EINVAL);
    802 #endif
    803 		}
    804 	}
    805 
    806 	/*
    807 	 * Return the message to the user.
    808 	 *
    809 	 * First, do the bookkeeping (before we risk being interrupted).
    810 	 */
    811 
    812 	msqptr->msg_cbytes -= msghdr->msg_ts;
    813 	msqptr->msg_qnum--;
    814 	msqptr->msg_lrpid = p->p_pid;
    815 	msqptr->msg_rtime = time.tv_sec;
    816 
    817 	/*
    818 	 * Make msgsz the actual amount that we'll be returning.
    819 	 * Note that this effectively truncates the message if it is too long
    820 	 * (since msgsz is never increased).
    821 	 */
    822 
    823 	MSG_PRINTF(("found a message, msgsz=%d, msg_ts=%d\n", msgsz,
    824 	    msghdr->msg_ts));
    825 	if (msgsz > msghdr->msg_ts)
    826 		msgsz = msghdr->msg_ts;
    827 
    828 	/*
    829 	 * Return the type to the user.
    830 	 */
    831 
    832 	eval = copyout((caddr_t)&msghdr->msg_type, user_msgp,
    833 	    sizeof(msghdr->msg_type));
    834 	if (eval != 0) {
    835 		MSG_PRINTF(("error (%d) copying out message type\n", eval));
    836 		msg_freehdr(msghdr);
    837 		wakeup((caddr_t)msqptr);
    838 		return(eval);
    839 	}
    840 	user_msgp += sizeof(msghdr->msg_type);
    841 
    842 	/*
    843 	 * Return the segments to the user
    844 	 */
    845 
    846 	next = msghdr->msg_spot;
    847 	for (len = 0; len < msgsz; len += msginfo.msgssz) {
    848 		size_t tlen;
    849 
    850 		if (msgsz > msginfo.msgssz)
    851 			tlen = msginfo.msgssz;
    852 		else
    853 			tlen = msgsz;
    854 		if (next <= -1)
    855 			panic("next too low #3");
    856 		if (next >= msginfo.msgseg)
    857 			panic("next out of range #3");
    858 		eval = copyout((caddr_t)&msgpool[next * msginfo.msgssz],
    859 		    user_msgp, tlen);
    860 		if (eval != 0) {
    861 			MSG_PRINTF(("error (%d) copying out message segment\n",
    862 			    eval));
    863 			msg_freehdr(msghdr);
    864 			wakeup((caddr_t)msqptr);
    865 			return(eval);
    866 		}
    867 		user_msgp += tlen;
    868 		next = msgmaps[next].next;
    869 	}
    870 
    871 	/*
    872 	 * Done, return the actual number of bytes copied out.
    873 	 */
    874 
    875 	msg_freehdr(msghdr);
    876 	wakeup((caddr_t)msqptr);
    877 	*retval = msgsz;
    878 	return(0);
    879 }
    880