sysv_msg.c revision 69644 
1/* $FreeBSD: head/sys/kern/sysv_msg.c 69644 2000-12-05 23:05:45Z alfred $ */
2
3/*
4 * Implementation of SVID messages
5 *
6 * Author:  Daniel Boulet
7 *
8 * Copyright 1993 Daniel Boulet and RTMX Inc.
9 *
10 * This system call was implemented by Daniel Boulet under contract from RTMX.
11 *
12 * Redistribution and use in source forms, with and without modification,
13 * are permitted provided that this entire comment appears intact.
14 *
15 * Redistribution in binary form may occur without any restrictions.
16 * Obviously, it would be nice if you gave credit where credit is due
17 * but requiring it would be too onerous.
18 *
19 * This software is provided ``AS IS'' without any warranties of any kind.
20 */
21
22#include "opt_sysvipc.h"
23
24#include <sys/param.h>
25#include <sys/systm.h>
26#include <sys/sysproto.h>
27#include <sys/kernel.h>
28#include <sys/proc.h>
29#include <sys/msg.h>
30#include <sys/syscall.h>
31#include <sys/sysent.h>
32#include <sys/sysctl.h>
33#include <sys/malloc.h>
34#include <sys/jail.h>
35
36static MALLOC_DEFINE(M_MSG, "msg", "SVID compatible message queues");
37
38static void msginit __P((void));
39static int msgunload __P((void));
40static int sysvmsg_modload __P((struct module *, int, void *));
41
42#define MSG_DEBUG
43#undef MSG_DEBUG_OK
44
45static void msg_freehdr __P((struct msg *msghdr));
46
47/* XXX casting to (sy_call_t *) is bogus, as usual. */
48static sy_call_t *msgcalls[] = {
49	(sy_call_t *)msgctl, (sy_call_t *)msgget,
50	(sy_call_t *)msgsnd, (sy_call_t *)msgrcv
51};
52
53struct msg {
54	struct	msg *msg_next;	/* next msg in the chain */
55	long	msg_type;	/* type of this message */
56    				/* >0 -> type of this message */
57    				/* 0 -> free header */
58	u_short	msg_ts;		/* size of this message */
59	short	msg_spot;	/* location of start of msg in buffer */
60};
61
62
63#ifndef MSGSSZ
64#define MSGSSZ	8		/* Each segment must be 2^N long */
65#endif
66#ifndef MSGSEG
67#define MSGSEG	2048		/* must be less than 32767 */
68#endif
69#define MSGMAX	(MSGSSZ*MSGSEG)
70#ifndef MSGMNB
71#define MSGMNB	2048		/* max # of bytes in a queue */
72#endif
73#ifndef MSGMNI
74#define MSGMNI	40
75#endif
76#ifndef MSGTQL
77#define MSGTQL	40
78#endif
79
80/*
81 * Based on the configuration parameters described in an SVR2 (yes, two)
82 * config(1m) man page.
83 *
84 * Each message is broken up and stored in segments that are msgssz bytes
85 * long.  For efficiency reasons, this should be a power of two.  Also,
86 * it doesn't make sense if it is less than 8 or greater than about 256.
87 * Consequently, msginit in kern/sysv_msg.c checks that msgssz is a power of
88 * two between 8 and 1024 inclusive (and panic's if it isn't).
89 */
90struct msginfo msginfo = {
91                MSGMAX,         /* max chars in a message */
92                MSGMNI,         /* # of message queue identifiers */
93                MSGMNB,         /* max chars in a queue */
94                MSGTQL,         /* max messages in system */
95                MSGSSZ,         /* size of a message segment */
96                		/* (must be small power of 2 greater than 4) */
97                MSGSEG          /* number of message segments */
98};
99
100/*
101 * macros to convert between msqid_ds's and msqid's.
102 * (specific to this implementation)
103 */
104#define MSQID(ix,ds)	((ix) & 0xffff | (((ds).msg_perm.seq << 16) & 0xffff0000))
105#define MSQID_IX(id)	((id) & 0xffff)
106#define MSQID_SEQ(id)	(((id) >> 16) & 0xffff)
107
108/*
109 * The rest of this file is specific to this particular implementation.
110 */
111
112struct msgmap {
113	short	next;		/* next segment in buffer */
114    				/* -1 -> available */
115    				/* 0..(MSGSEG-1) -> index of next segment */
116};
117
118#define MSG_LOCKED	01000	/* Is this msqid_ds locked? */
119
120static int nfree_msgmaps;	/* # of free map entries */
121static short free_msgmaps;	/* head of linked list of free map entries */
122static struct msg *free_msghdrs;/* list of free msg headers */
123static char *msgpool;		/* MSGMAX byte long msg buffer pool */
124static struct msgmap *msgmaps;	/* MSGSEG msgmap structures */
125static struct msg *msghdrs;	/* MSGTQL msg headers */
126static struct msqid_ds *msqids;	/* MSGMNI msqid_ds struct's */
127
128static void
129msginit()
130{
131	register int i;
132
133	msgpool = malloc(msginfo.msgmax, M_MSG, M_WAITOK);
134	if (msgpool == NULL)
135		panic("msgpool is NULL");
136	msgmaps = malloc(sizeof(struct msgmap) * msginfo.msgseg, M_MSG, M_WAITOK);
137	if (msgmaps == NULL)
138		panic("msgmaps is NULL");
139	msghdrs = malloc(sizeof(struct msg) * msginfo.msgtql, M_MSG, M_WAITOK);
140	if (msghdrs == NULL)
141		panic("msghdrs is NULL");
142	msqids = malloc(sizeof(struct msqid_ds) * msginfo.msgmni, M_MSG, M_WAITOK);
143	if (msqids == NULL)
144		panic("msqids is NULL");
145
146	/*
147	 * msginfo.msgssz should be a power of two for efficiency reasons.
148	 * It is also pretty silly if msginfo.msgssz is less than 8
149	 * or greater than about 256 so ...
150	 */
151
152	i = 8;
153	while (i < 1024 && i != msginfo.msgssz)
154		i <<= 1;
155    	if (i != msginfo.msgssz) {
156		printf("msginfo.msgssz=%d (0x%x)\n", msginfo.msgssz,
157		    msginfo.msgssz);
158		panic("msginfo.msgssz not a small power of 2");
159	}
160
161	if (msginfo.msgseg > 32767) {
162		printf("msginfo.msgseg=%d\n", msginfo.msgseg);
163		panic("msginfo.msgseg > 32767");
164	}
165
166	if (msgmaps == NULL)
167		panic("msgmaps is NULL");
168
169	for (i = 0; i < msginfo.msgseg; i++) {
170		if (i > 0)
171			msgmaps[i-1].next = i;
172		msgmaps[i].next = -1;	/* implies entry is available */
173	}
174	free_msgmaps = 0;
175	nfree_msgmaps = msginfo.msgseg;
176
177	if (msghdrs == NULL)
178		panic("msghdrs is NULL");
179
180	for (i = 0; i < msginfo.msgtql; i++) {
181		msghdrs[i].msg_type = 0;
182		if (i > 0)
183			msghdrs[i-1].msg_next = &msghdrs[i];
184		msghdrs[i].msg_next = NULL;
185    	}
186	free_msghdrs = &msghdrs[0];
187
188	if (msqids == NULL)
189		panic("msqids is NULL");
190
191	for (i = 0; i < msginfo.msgmni; i++) {
192		msqids[i].msg_qbytes = 0;	/* implies entry is available */
193		msqids[i].msg_perm.seq = 0;	/* reset to a known value */
194		msqids[i].msg_perm.mode = 0;
195	}
196}
197
198static int
199msgunload()
200{
201	struct msqid_ds *msqptr;
202	int msqid;
203
204	for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
205		/*
206		 * Look for an unallocated and unlocked msqid_ds.
207		 * msqid_ds's can be locked by msgsnd or msgrcv while
208		 * they are copying the message in/out.  We can't
209		 * re-use the entry until they release it.
210		 */
211		msqptr = &msqids[msqid];
212		if (msqptr->msg_qbytes != 0 ||
213		    (msqptr->msg_perm.mode & MSG_LOCKED) != 0)
214			break;
215	}
216	if (msqid != msginfo.msgmni)
217		return (EBUSY);
218
219	free(msgpool, M_MSG);
220	free(msgmaps, M_MSG);
221	free(msghdrs, M_MSG);
222	free(msqids, M_MSG);
223	return (0);
224}
225
226
227static int
228sysvmsg_modload(struct module *module, int cmd, void *arg)
229{
230	int error = 0;
231
232	switch (cmd) {
233	case MOD_LOAD:
234		msginit();
235		break;
236	case MOD_UNLOAD:
237		error = msgunload();
238		break;
239	case MOD_SHUTDOWN:
240		break;
241	default:
242		error = EINVAL;
243		break;
244	}
245	return (error);
246}
247
248static moduledata_t sysvmsg_moduledata = {
249	"sysvmsg_mod",
250	&sysvmsg_modload,
251	NULL
252};
253
254SYSCALL_MODULE_HELPER(msgsys, 6);
255SYSCALL_MODULE_HELPER(msgctl, 3);
256SYSCALL_MODULE_HELPER(msgget, 2);
257SYSCALL_MODULE_HELPER(msgsnd, 4);
258SYSCALL_MODULE_HELPER(msgrcv, 5);
259
260DECLARE_MODULE(sysvmsg_mod, sysvmsg_moduledata,
261	SI_SUB_SYSV_MSG, SI_ORDER_FIRST);
262
263/*
264 * Entry point for all MSG calls
265 */
266int
267msgsys(p, uap)
268	struct proc *p;
269	/* XXX actually varargs. */
270	struct msgsys_args /* {
271		u_int	which;
272		int	a2;
273		int	a3;
274		int	a4;
275		int	a5;
276		int	a6;
277	} */ *uap;
278{
279
280	if (!jail_sysvipc_allowed && p->p_prison != NULL)
281		return (ENOSYS);
282
283	if (uap->which >= sizeof(msgcalls)/sizeof(msgcalls[0]))
284		return (EINVAL);
285	return ((*msgcalls[uap->which])(p, &uap->a2));
286}
287
288static void
289msg_freehdr(msghdr)
290	struct msg *msghdr;
291{
292	while (msghdr->msg_ts > 0) {
293		short next;
294		if (msghdr->msg_spot < 0 || msghdr->msg_spot >= msginfo.msgseg)
295			panic("msghdr->msg_spot out of range");
296		next = msgmaps[msghdr->msg_spot].next;
297		msgmaps[msghdr->msg_spot].next = free_msgmaps;
298		free_msgmaps = msghdr->msg_spot;
299		nfree_msgmaps++;
300		msghdr->msg_spot = next;
301		if (msghdr->msg_ts >= msginfo.msgssz)
302			msghdr->msg_ts -= msginfo.msgssz;
303		else
304			msghdr->msg_ts = 0;
305	}
306	if (msghdr->msg_spot != -1)
307		panic("msghdr->msg_spot != -1");
308	msghdr->msg_next = free_msghdrs;
309	free_msghdrs = msghdr;
310}
311
312#ifndef _SYS_SYSPROTO_H_
313struct msgctl_args {
314	int	msqid;
315	int	cmd;
316	struct	msqid_ds *buf;
317};
318#endif
319
320int
321msgctl(p, uap)
322	struct proc *p;
323	register struct msgctl_args *uap;
324{
325	int msqid = uap->msqid;
326	int cmd = uap->cmd;
327	struct msqid_ds *user_msqptr = uap->buf;
328	int rval, eval;
329	struct msqid_ds msqbuf;
330	register struct msqid_ds *msqptr;
331
332#ifdef MSG_DEBUG_OK
333	printf("call to msgctl(%d, %d, 0x%x)\n", msqid, cmd, user_msqptr);
334#endif
335
336	if (!jail_sysvipc_allowed && p->p_prison != NULL)
337		return (ENOSYS);
338
339	msqid = IPCID_TO_IX(msqid);
340
341	if (msqid < 0 || msqid >= msginfo.msgmni) {
342#ifdef MSG_DEBUG_OK
343		printf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
344		    msginfo.msgmni);
345#endif
346		return(EINVAL);
347	}
348
349	msqptr = &msqids[msqid];
350
351	if (msqptr->msg_qbytes == 0) {
352#ifdef MSG_DEBUG_OK
353		printf("no such msqid\n");
354#endif
355		return(EINVAL);
356	}
357	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
358#ifdef MSG_DEBUG_OK
359		printf("wrong sequence number\n");
360#endif
361		return(EINVAL);
362	}
363
364	eval = 0;
365	rval = 0;
366
367	switch (cmd) {
368
369	case IPC_RMID:
370	{
371		struct msg *msghdr;
372		if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)))
373			return(eval);
374		/* Free the message headers */
375		msghdr = msqptr->msg_first;
376		while (msghdr != NULL) {
377			struct msg *msghdr_tmp;
378
379			/* Free the segments of each message */
380			msqptr->msg_cbytes -= msghdr->msg_ts;
381			msqptr->msg_qnum--;
382			msghdr_tmp = msghdr;
383			msghdr = msghdr->msg_next;
384			msg_freehdr(msghdr_tmp);
385		}
386
387		if (msqptr->msg_cbytes != 0)
388			panic("msg_cbytes is screwed up");
389		if (msqptr->msg_qnum != 0)
390			panic("msg_qnum is screwed up");
391
392		msqptr->msg_qbytes = 0;	/* Mark it as free */
393
394		wakeup((caddr_t)msqptr);
395	}
396
397		break;
398
399	case IPC_SET:
400		if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_M)))
401			return(eval);
402		if ((eval = copyin(user_msqptr, &msqbuf, sizeof(msqbuf))) != 0)
403			return(eval);
404		if (msqbuf.msg_qbytes > msqptr->msg_qbytes) {
405			eval = suser(p);
406			if (eval)
407				return(eval);
408		}
409		if (msqbuf.msg_qbytes > msginfo.msgmnb) {
410#ifdef MSG_DEBUG_OK
411			printf("can't increase msg_qbytes beyond %d (truncating)\n",
412			    msginfo.msgmnb);
413#endif
414			msqbuf.msg_qbytes = msginfo.msgmnb;	/* silently restrict qbytes to system limit */
415		}
416		if (msqbuf.msg_qbytes == 0) {
417#ifdef MSG_DEBUG_OK
418			printf("can't reduce msg_qbytes to 0\n");
419#endif
420			return(EINVAL);		/* non-standard errno! */
421		}
422		msqptr->msg_perm.uid = msqbuf.msg_perm.uid;	/* change the owner */
423		msqptr->msg_perm.gid = msqbuf.msg_perm.gid;	/* change the owner */
424		msqptr->msg_perm.mode = (msqptr->msg_perm.mode & ~0777) |
425		    (msqbuf.msg_perm.mode & 0777);
426		msqptr->msg_qbytes = msqbuf.msg_qbytes;
427		msqptr->msg_ctime = time_second;
428		break;
429
430	case IPC_STAT:
431		if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
432#ifdef MSG_DEBUG_OK
433			printf("requester doesn't have read access\n");
434#endif
435			return(eval);
436		}
437		eval = copyout((caddr_t)msqptr, user_msqptr,
438		    sizeof(struct msqid_ds));
439		break;
440
441	default:
442#ifdef MSG_DEBUG_OK
443		printf("invalid command %d\n", cmd);
444#endif
445		return(EINVAL);
446	}
447
448	if (eval == 0)
449		p->p_retval[0] = rval;
450	return(eval);
451}
452
453#ifndef _SYS_SYSPROTO_H_
454struct msgget_args {
455	key_t	key;
456	int	msgflg;
457};
458#endif
459
460int
461msgget(p, uap)
462	struct proc *p;
463	register struct msgget_args *uap;
464{
465	int msqid, eval;
466	int key = uap->key;
467	int msgflg = uap->msgflg;
468	struct ucred *cred = p->p_ucred;
469	register struct msqid_ds *msqptr = NULL;
470
471#ifdef MSG_DEBUG_OK
472	printf("msgget(0x%x, 0%o)\n", key, msgflg);
473#endif
474
475	if (!jail_sysvipc_allowed && p->p_prison != NULL)
476		return (ENOSYS);
477
478	if (key != IPC_PRIVATE) {
479		for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
480			msqptr = &msqids[msqid];
481			if (msqptr->msg_qbytes != 0 &&
482			    msqptr->msg_perm.key == key)
483				break;
484		}
485		if (msqid < msginfo.msgmni) {
486#ifdef MSG_DEBUG_OK
487			printf("found public key\n");
488#endif
489			if ((msgflg & IPC_CREAT) && (msgflg & IPC_EXCL)) {
490#ifdef MSG_DEBUG_OK
491				printf("not exclusive\n");
492#endif
493				return(EEXIST);
494			}
495			if ((eval = ipcperm(p, &msqptr->msg_perm, msgflg & 0700 ))) {
496#ifdef MSG_DEBUG_OK
497				printf("requester doesn't have 0%o access\n",
498				    msgflg & 0700);
499#endif
500				return(eval);
501			}
502			goto found;
503		}
504	}
505
506#ifdef MSG_DEBUG_OK
507	printf("need to allocate the msqid_ds\n");
508#endif
509	if (key == IPC_PRIVATE || (msgflg & IPC_CREAT)) {
510		for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
511			/*
512			 * Look for an unallocated and unlocked msqid_ds.
513			 * msqid_ds's can be locked by msgsnd or msgrcv while
514			 * they are copying the message in/out.  We can't
515			 * re-use the entry until they release it.
516			 */
517			msqptr = &msqids[msqid];
518			if (msqptr->msg_qbytes == 0 &&
519			    (msqptr->msg_perm.mode & MSG_LOCKED) == 0)
520				break;
521		}
522		if (msqid == msginfo.msgmni) {
523#ifdef MSG_DEBUG_OK
524			printf("no more msqid_ds's available\n");
525#endif
526			return(ENOSPC);
527		}
528#ifdef MSG_DEBUG_OK
529		printf("msqid %d is available\n", msqid);
530#endif
531		msqptr->msg_perm.key = key;
532		msqptr->msg_perm.cuid = cred->cr_uid;
533		msqptr->msg_perm.uid = cred->cr_uid;
534		msqptr->msg_perm.cgid = cred->cr_gid;
535		msqptr->msg_perm.gid = cred->cr_gid;
536		msqptr->msg_perm.mode = (msgflg & 0777);
537		/* Make sure that the returned msqid is unique */
538		msqptr->msg_perm.seq++;
539		msqptr->msg_first = NULL;
540		msqptr->msg_last = NULL;
541		msqptr->msg_cbytes = 0;
542		msqptr->msg_qnum = 0;
543		msqptr->msg_qbytes = msginfo.msgmnb;
544		msqptr->msg_lspid = 0;
545		msqptr->msg_lrpid = 0;
546		msqptr->msg_stime = 0;
547		msqptr->msg_rtime = 0;
548		msqptr->msg_ctime = time_second;
549	} else {
550#ifdef MSG_DEBUG_OK
551		printf("didn't find it and wasn't asked to create it\n");
552#endif
553		return(ENOENT);
554	}
555
556found:
557	/* Construct the unique msqid */
558	p->p_retval[0] = IXSEQ_TO_IPCID(msqid, msqptr->msg_perm);
559	return(0);
560}
561
562#ifndef _SYS_SYSPROTO_H_
563struct msgsnd_args {
564	int	msqid;
565	void	*msgp;
566	size_t	msgsz;
567	int	msgflg;
568};
569#endif
570
571int
572msgsnd(p, uap)
573	struct proc *p;
574	register struct msgsnd_args *uap;
575{
576	int msqid = uap->msqid;
577	void *user_msgp = uap->msgp;
578	size_t msgsz = uap->msgsz;
579	int msgflg = uap->msgflg;
580	int segs_needed, eval;
581	register struct msqid_ds *msqptr;
582	register struct msg *msghdr;
583	short next;
584
585#ifdef MSG_DEBUG_OK
586	printf("call to msgsnd(%d, 0x%x, %d, %d)\n", msqid, user_msgp, msgsz,
587	    msgflg);
588#endif
589
590	if (!jail_sysvipc_allowed && p->p_prison != NULL)
591		return (ENOSYS);
592
593	msqid = IPCID_TO_IX(msqid);
594
595	if (msqid < 0 || msqid >= msginfo.msgmni) {
596#ifdef MSG_DEBUG_OK
597		printf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
598		    msginfo.msgmni);
599#endif
600		return(EINVAL);
601	}
602
603	msqptr = &msqids[msqid];
604	if (msqptr->msg_qbytes == 0) {
605#ifdef MSG_DEBUG_OK
606		printf("no such message queue id\n");
607#endif
608		return(EINVAL);
609	}
610	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
611#ifdef MSG_DEBUG_OK
612		printf("wrong sequence number\n");
613#endif
614		return(EINVAL);
615	}
616
617	if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_W))) {
618#ifdef MSG_DEBUG_OK
619		printf("requester doesn't have write access\n");
620#endif
621		return(eval);
622	}
623
624	segs_needed = (msgsz + msginfo.msgssz - 1) / msginfo.msgssz;
625#ifdef MSG_DEBUG_OK
626	printf("msgsz=%d, msgssz=%d, segs_needed=%d\n", msgsz, msginfo.msgssz,
627	    segs_needed);
628#endif
629	for (;;) {
630		int need_more_resources = 0;
631
632		/*
633		 * check msgsz
634		 * (inside this loop in case msg_qbytes changes while we sleep)
635		 */
636
637		if (msgsz > msqptr->msg_qbytes) {
638#ifdef MSG_DEBUG_OK
639			printf("msgsz > msqptr->msg_qbytes\n");
640#endif
641			return(EINVAL);
642		}
643
644		if (msqptr->msg_perm.mode & MSG_LOCKED) {
645#ifdef MSG_DEBUG_OK
646			printf("msqid is locked\n");
647#endif
648			need_more_resources = 1;
649		}
650		if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes) {
651#ifdef MSG_DEBUG_OK
652			printf("msgsz + msg_cbytes > msg_qbytes\n");
653#endif
654			need_more_resources = 1;
655		}
656		if (segs_needed > nfree_msgmaps) {
657#ifdef MSG_DEBUG_OK
658			printf("segs_needed > nfree_msgmaps\n");
659#endif
660			need_more_resources = 1;
661		}
662		if (free_msghdrs == NULL) {
663#ifdef MSG_DEBUG_OK
664			printf("no more msghdrs\n");
665#endif
666			need_more_resources = 1;
667		}
668
669		if (need_more_resources) {
670			int we_own_it;
671
672			if ((msgflg & IPC_NOWAIT) != 0) {
673#ifdef MSG_DEBUG_OK
674				printf("need more resources but caller doesn't want to wait\n");
675#endif
676				return(EAGAIN);
677			}
678
679			if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0) {
680#ifdef MSG_DEBUG_OK
681				printf("we don't own the msqid_ds\n");
682#endif
683				we_own_it = 0;
684			} else {
685				/* Force later arrivals to wait for our
686				   request */
687#ifdef MSG_DEBUG_OK
688				printf("we own the msqid_ds\n");
689#endif
690				msqptr->msg_perm.mode |= MSG_LOCKED;
691				we_own_it = 1;
692			}
693#ifdef MSG_DEBUG_OK
694			printf("goodnight\n");
695#endif
696			eval = tsleep((caddr_t)msqptr, (PZERO - 4) | PCATCH,
697			    "msgwait", 0);
698#ifdef MSG_DEBUG_OK
699			printf("good morning, eval=%d\n", eval);
700#endif
701			if (we_own_it)
702				msqptr->msg_perm.mode &= ~MSG_LOCKED;
703			if (eval != 0) {
704#ifdef MSG_DEBUG_OK
705				printf("msgsnd:  interrupted system call\n");
706#endif
707				return(EINTR);
708			}
709
710			/*
711			 * Make sure that the msq queue still exists
712			 */
713
714			if (msqptr->msg_qbytes == 0) {
715#ifdef MSG_DEBUG_OK
716				printf("msqid deleted\n");
717#endif
718				return(EIDRM);
719			}
720
721		} else {
722#ifdef MSG_DEBUG_OK
723			printf("got all the resources that we need\n");
724#endif
725			break;
726		}
727	}
728
729	/*
730	 * We have the resources that we need.
731	 * Make sure!
732	 */
733
734	if (msqptr->msg_perm.mode & MSG_LOCKED)
735		panic("msg_perm.mode & MSG_LOCKED");
736	if (segs_needed > nfree_msgmaps)
737		panic("segs_needed > nfree_msgmaps");
738	if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes)
739		panic("msgsz + msg_cbytes > msg_qbytes");
740	if (free_msghdrs == NULL)
741		panic("no more msghdrs");
742
743	/*
744	 * Re-lock the msqid_ds in case we page-fault when copying in the
745	 * message
746	 */
747
748	if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0)
749		panic("msqid_ds is already locked");
750	msqptr->msg_perm.mode |= MSG_LOCKED;
751
752	/*
753	 * Allocate a message header
754	 */
755
756	msghdr = free_msghdrs;
757	free_msghdrs = msghdr->msg_next;
758	msghdr->msg_spot = -1;
759	msghdr->msg_ts = msgsz;
760
761	/*
762	 * Allocate space for the message
763	 */
764
765	while (segs_needed > 0) {
766		if (nfree_msgmaps <= 0)
767			panic("not enough msgmaps");
768		if (free_msgmaps == -1)
769			panic("nil free_msgmaps");
770		next = free_msgmaps;
771		if (next <= -1)
772			panic("next too low #1");
773		if (next >= msginfo.msgseg)
774			panic("next out of range #1");
775#ifdef MSG_DEBUG_OK
776		printf("allocating segment %d to message\n", next);
777#endif
778		free_msgmaps = msgmaps[next].next;
779		nfree_msgmaps--;
780		msgmaps[next].next = msghdr->msg_spot;
781		msghdr->msg_spot = next;
782		segs_needed--;
783	}
784
785	/*
786	 * Copy in the message type
787	 */
788
789	if ((eval = copyin(user_msgp, &msghdr->msg_type,
790	    sizeof(msghdr->msg_type))) != 0) {
791#ifdef MSG_DEBUG_OK
792		printf("error %d copying the message type\n", eval);
793#endif
794		msg_freehdr(msghdr);
795		msqptr->msg_perm.mode &= ~MSG_LOCKED;
796		wakeup((caddr_t)msqptr);
797		return(eval);
798	}
799	user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
800
801	/*
802	 * Validate the message type
803	 */
804
805	if (msghdr->msg_type < 1) {
806		msg_freehdr(msghdr);
807		msqptr->msg_perm.mode &= ~MSG_LOCKED;
808		wakeup((caddr_t)msqptr);
809#ifdef MSG_DEBUG_OK
810		printf("mtype (%d) < 1\n", msghdr->msg_type);
811#endif
812		return(EINVAL);
813	}
814
815	/*
816	 * Copy in the message body
817	 */
818
819	next = msghdr->msg_spot;
820	while (msgsz > 0) {
821		size_t tlen;
822		if (msgsz > msginfo.msgssz)
823			tlen = msginfo.msgssz;
824		else
825			tlen = msgsz;
826		if (next <= -1)
827			panic("next too low #2");
828		if (next >= msginfo.msgseg)
829			panic("next out of range #2");
830		if ((eval = copyin(user_msgp, &msgpool[next * msginfo.msgssz],
831		    tlen)) != 0) {
832#ifdef MSG_DEBUG_OK
833			printf("error %d copying in message segment\n", eval);
834#endif
835			msg_freehdr(msghdr);
836			msqptr->msg_perm.mode &= ~MSG_LOCKED;
837			wakeup((caddr_t)msqptr);
838			return(eval);
839		}
840		msgsz -= tlen;
841		user_msgp = (char *)user_msgp + tlen;
842		next = msgmaps[next].next;
843	}
844	if (next != -1)
845		panic("didn't use all the msg segments");
846
847	/*
848	 * We've got the message.  Unlock the msqid_ds.
849	 */
850
851	msqptr->msg_perm.mode &= ~MSG_LOCKED;
852
853	/*
854	 * Make sure that the msqid_ds is still allocated.
855	 */
856
857	if (msqptr->msg_qbytes == 0) {
858		msg_freehdr(msghdr);
859		wakeup((caddr_t)msqptr);
860		return(EIDRM);
861	}
862
863	/*
864	 * Put the message into the queue
865	 */
866
867	if (msqptr->msg_first == NULL) {
868		msqptr->msg_first = msghdr;
869		msqptr->msg_last = msghdr;
870	} else {
871		msqptr->msg_last->msg_next = msghdr;
872		msqptr->msg_last = msghdr;
873	}
874	msqptr->msg_last->msg_next = NULL;
875
876	msqptr->msg_cbytes += msghdr->msg_ts;
877	msqptr->msg_qnum++;
878	msqptr->msg_lspid = p->p_pid;
879	msqptr->msg_stime = time_second;
880
881	wakeup((caddr_t)msqptr);
882	p->p_retval[0] = 0;
883	return(0);
884}
885
886#ifndef _SYS_SYSPROTO_H_
887struct msgrcv_args {
888	int	msqid;
889	void	*msgp;
890	size_t	msgsz;
891	long	msgtyp;
892	int	msgflg;
893};
894#endif
895
896int
897msgrcv(p, uap)
898	struct proc *p;
899	register struct msgrcv_args *uap;
900{
901	int msqid = uap->msqid;
902	void *user_msgp = uap->msgp;
903	size_t msgsz = uap->msgsz;
904	long msgtyp = uap->msgtyp;
905	int msgflg = uap->msgflg;
906	size_t len;
907	register struct msqid_ds *msqptr;
908	register struct msg *msghdr;
909	int eval;
910	short next;
911
912#ifdef MSG_DEBUG_OK
913	printf("call to msgrcv(%d, 0x%x, %d, %ld, %d)\n", msqid, user_msgp,
914	    msgsz, msgtyp, msgflg);
915#endif
916
917	if (!jail_sysvipc_allowed && p->p_prison != NULL)
918		return (ENOSYS);
919
920	msqid = IPCID_TO_IX(msqid);
921
922	if (msqid < 0 || msqid >= msginfo.msgmni) {
923#ifdef MSG_DEBUG_OK
924		printf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
925		    msginfo.msgmni);
926#endif
927		return(EINVAL);
928	}
929
930	msqptr = &msqids[msqid];
931	if (msqptr->msg_qbytes == 0) {
932#ifdef MSG_DEBUG_OK
933		printf("no such message queue id\n");
934#endif
935		return(EINVAL);
936	}
937	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
938#ifdef MSG_DEBUG_OK
939		printf("wrong sequence number\n");
940#endif
941		return(EINVAL);
942	}
943
944	if ((eval = ipcperm(p, &msqptr->msg_perm, IPC_R))) {
945#ifdef MSG_DEBUG_OK
946		printf("requester doesn't have read access\n");
947#endif
948		return(eval);
949	}
950
951	msghdr = NULL;
952	while (msghdr == NULL) {
953		if (msgtyp == 0) {
954			msghdr = msqptr->msg_first;
955			if (msghdr != NULL) {
956				if (msgsz < msghdr->msg_ts &&
957				    (msgflg & MSG_NOERROR) == 0) {
958#ifdef MSG_DEBUG_OK
959					printf("first message on the queue is too big (want %d, got %d)\n",
960					    msgsz, msghdr->msg_ts);
961#endif
962					return(E2BIG);
963				}
964				if (msqptr->msg_first == msqptr->msg_last) {
965					msqptr->msg_first = NULL;
966					msqptr->msg_last = NULL;
967				} else {
968					msqptr->msg_first = msghdr->msg_next;
969					if (msqptr->msg_first == NULL)
970						panic("msg_first/last screwed up #1");
971				}
972			}
973		} else {
974			struct msg *previous;
975			struct msg **prev;
976
977			previous = NULL;
978			prev = &(msqptr->msg_first);
979			while ((msghdr = *prev) != NULL) {
980				/*
981				 * Is this message's type an exact match or is
982				 * this message's type less than or equal to
983				 * the absolute value of a negative msgtyp?
984				 * Note that the second half of this test can
985				 * NEVER be true if msgtyp is positive since
986				 * msg_type is always positive!
987				 */
988
989				if (msgtyp == msghdr->msg_type ||
990				    msghdr->msg_type <= -msgtyp) {
991#ifdef MSG_DEBUG_OK
992					printf("found message type %d, requested %d\n",
993					    msghdr->msg_type, msgtyp);
994#endif
995					if (msgsz < msghdr->msg_ts &&
996					    (msgflg & MSG_NOERROR) == 0) {
997#ifdef MSG_DEBUG_OK
998						printf("requested message on the queue is too big (want %d, got %d)\n",
999						    msgsz, msghdr->msg_ts);
1000#endif
1001						return(E2BIG);
1002					}
1003					*prev = msghdr->msg_next;
1004					if (msghdr == msqptr->msg_last) {
1005						if (previous == NULL) {
1006							if (prev !=
1007							    &msqptr->msg_first)
1008								panic("msg_first/last screwed up #2");
1009							msqptr->msg_first =
1010							    NULL;
1011							msqptr->msg_last =
1012							    NULL;
1013						} else {
1014							if (prev ==
1015							    &msqptr->msg_first)
1016								panic("msg_first/last screwed up #3");
1017							msqptr->msg_last =
1018							    previous;
1019						}
1020					}
1021					break;
1022				}
1023				previous = msghdr;
1024				prev = &(msghdr->msg_next);
1025			}
1026		}
1027
1028		/*
1029		 * We've either extracted the msghdr for the appropriate
1030		 * message or there isn't one.
1031		 * If there is one then bail out of this loop.
1032		 */
1033
1034		if (msghdr != NULL)
1035			break;
1036
1037		/*
1038		 * Hmph!  No message found.  Does the user want to wait?
1039		 */
1040
1041		if ((msgflg & IPC_NOWAIT) != 0) {
1042#ifdef MSG_DEBUG_OK
1043			printf("no appropriate message found (msgtyp=%d)\n",
1044			    msgtyp);
1045#endif
1046			/* The SVID says to return ENOMSG. */
1047#ifdef ENOMSG
1048			return(ENOMSG);
1049#else
1050			/* Unfortunately, BSD doesn't define that code yet! */
1051			return(EAGAIN);
1052#endif
1053		}
1054
1055		/*
1056		 * Wait for something to happen
1057		 */
1058
1059#ifdef MSG_DEBUG_OK
1060		printf("msgrcv:  goodnight\n");
1061#endif
1062		eval = tsleep((caddr_t)msqptr, (PZERO - 4) | PCATCH, "msgwait",
1063		    0);
1064#ifdef MSG_DEBUG_OK
1065		printf("msgrcv:  good morning (eval=%d)\n", eval);
1066#endif
1067
1068		if (eval != 0) {
1069#ifdef MSG_DEBUG_OK
1070			printf("msgsnd:  interrupted system call\n");
1071#endif
1072			return(EINTR);
1073		}
1074
1075		/*
1076		 * Make sure that the msq queue still exists
1077		 */
1078
1079		if (msqptr->msg_qbytes == 0 ||
1080		    msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
1081#ifdef MSG_DEBUG_OK
1082			printf("msqid deleted\n");
1083#endif
1084			return(EIDRM);
1085		}
1086	}
1087
1088	/*
1089	 * Return the message to the user.
1090	 *
1091	 * First, do the bookkeeping (before we risk being interrupted).
1092	 */
1093
1094	msqptr->msg_cbytes -= msghdr->msg_ts;
1095	msqptr->msg_qnum--;
1096	msqptr->msg_lrpid = p->p_pid;
1097	msqptr->msg_rtime = time_second;
1098
1099	/*
1100	 * Make msgsz the actual amount that we'll be returning.
1101	 * Note that this effectively truncates the message if it is too long
1102	 * (since msgsz is never increased).
1103	 */
1104
1105#ifdef MSG_DEBUG_OK
1106	printf("found a message, msgsz=%d, msg_ts=%d\n", msgsz,
1107	    msghdr->msg_ts);
1108#endif
1109	if (msgsz > msghdr->msg_ts)
1110		msgsz = msghdr->msg_ts;
1111
1112	/*
1113	 * Return the type to the user.
1114	 */
1115
1116	eval = copyout((caddr_t)&(msghdr->msg_type), user_msgp,
1117	    sizeof(msghdr->msg_type));
1118	if (eval != 0) {
1119#ifdef MSG_DEBUG_OK
1120		printf("error (%d) copying out message type\n", eval);
1121#endif
1122		msg_freehdr(msghdr);
1123		wakeup((caddr_t)msqptr);
1124		return(eval);
1125	}
1126	user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
1127
1128	/*
1129	 * Return the segments to the user
1130	 */
1131
1132	next = msghdr->msg_spot;
1133	for (len = 0; len < msgsz; len += msginfo.msgssz) {
1134		size_t tlen;
1135
1136		if (msgsz - len > msginfo.msgssz)
1137			tlen = msginfo.msgssz;
1138		else
1139			tlen = msgsz - len;
1140		if (next <= -1)
1141			panic("next too low #3");
1142		if (next >= msginfo.msgseg)
1143			panic("next out of range #3");
1144		eval = copyout((caddr_t)&msgpool[next * msginfo.msgssz],
1145		    user_msgp, tlen);
1146		if (eval != 0) {
1147#ifdef MSG_DEBUG_OK
1148			printf("error (%d) copying out message segment\n",
1149			    eval);
1150#endif
1151			msg_freehdr(msghdr);
1152			wakeup((caddr_t)msqptr);
1153			return(eval);
1154		}
1155		user_msgp = (char *)user_msgp + tlen;
1156		next = msgmaps[next].next;
1157	}
1158
1159	/*
1160	 * Done, return the actual number of bytes copied out.
1161	 */
1162
1163	msg_freehdr(msghdr);
1164	wakeup((caddr_t)msqptr);
1165	p->p_retval[0] = msgsz;
1166	return(0);
1167}
1168