sysv_msg.c revision 88633 
1/* $FreeBSD: head/sys/kern/sysv_msg.c 88633 2001-12-29 07:13:47Z 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/lock.h>
30#include <sys/mutex.h>
31#include <sys/msg.h>
32#include <sys/syscall.h>
33#include <sys/sysent.h>
34#include <sys/sysctl.h>
35#include <sys/malloc.h>
36#include <sys/jail.h>
37
38static MALLOC_DEFINE(M_MSG, "msg", "SVID compatible message queues");
39
40static void msginit __P((void));
41static int msgunload __P((void));
42static int sysvmsg_modload __P((struct module *, int, void *));
43
44#define MSG_DEBUG
45#undef MSG_DEBUG_OK
46
47static void msg_freehdr __P((struct msg *msghdr));
48
49/* XXX casting to (sy_call_t *) is bogus, as usual. */
50static sy_call_t *msgcalls[] = {
51	(sy_call_t *)msgctl, (sy_call_t *)msgget,
52	(sy_call_t *)msgsnd, (sy_call_t *)msgrcv
53};
54
55struct msg {
56	struct	msg *msg_next;	/* next msg in the chain */
57	long	msg_type;	/* type of this message */
58    				/* >0 -> type of this message */
59    				/* 0 -> free header */
60	u_short	msg_ts;		/* size of this message */
61	short	msg_spot;	/* location of start of msg in buffer */
62};
63
64
65#ifndef MSGSSZ
66#define MSGSSZ	8		/* Each segment must be 2^N long */
67#endif
68#ifndef MSGSEG
69#define MSGSEG	2048		/* must be less than 32767 */
70#endif
71#define MSGMAX	(MSGSSZ*MSGSEG)
72#ifndef MSGMNB
73#define MSGMNB	2048		/* max # of bytes in a queue */
74#endif
75#ifndef MSGMNI
76#define MSGMNI	40
77#endif
78#ifndef MSGTQL
79#define MSGTQL	40
80#endif
81
82/*
83 * Based on the configuration parameters described in an SVR2 (yes, two)
84 * config(1m) man page.
85 *
86 * Each message is broken up and stored in segments that are msgssz bytes
87 * long.  For efficiency reasons, this should be a power of two.  Also,
88 * it doesn't make sense if it is less than 8 or greater than about 256.
89 * Consequently, msginit in kern/sysv_msg.c checks that msgssz is a power of
90 * two between 8 and 1024 inclusive (and panic's if it isn't).
91 */
92struct msginfo msginfo = {
93                MSGMAX,         /* max chars in a message */
94                MSGMNI,         /* # of message queue identifiers */
95                MSGMNB,         /* max chars in a queue */
96                MSGTQL,         /* max messages in system */
97                MSGSSZ,         /* size of a message segment */
98                		/* (must be small power of 2 greater than 4) */
99                MSGSEG          /* number of message segments */
100};
101
102/*
103 * macros to convert between msqid_ds's and msqid's.
104 * (specific to this implementation)
105 */
106#define MSQID(ix,ds)	((ix) & 0xffff | (((ds).msg_perm.seq << 16) & 0xffff0000))
107#define MSQID_IX(id)	((id) & 0xffff)
108#define MSQID_SEQ(id)	(((id) >> 16) & 0xffff)
109
110/*
111 * The rest of this file is specific to this particular implementation.
112 */
113
114struct msgmap {
115	short	next;		/* next segment in buffer */
116    				/* -1 -> available */
117    				/* 0..(MSGSEG-1) -> index of next segment */
118};
119
120#define MSG_LOCKED	01000	/* Is this msqid_ds locked? */
121
122static int nfree_msgmaps;	/* # of free map entries */
123static short free_msgmaps;	/* head of linked list of free map entries */
124static struct msg *free_msghdrs;/* list of free msg headers */
125static char *msgpool;		/* MSGMAX byte long msg buffer pool */
126static struct msgmap *msgmaps;	/* MSGSEG msgmap structures */
127static struct msg *msghdrs;	/* MSGTQL msg headers */
128static struct msqid_ds *msqids;	/* MSGMNI msqid_ds struct's */
129
130static void
131msginit()
132{
133	register int i;
134
135	TUNABLE_INT_FETCH("kern.ipc.msgseg", &msginfo.msgseg);
136	TUNABLE_INT_FETCH("kern.ipc.msgssz", &msginfo.msgssz);
137	msginfo.msgmax = msginfo.msgseg * msginfo.msgssz;
138	TUNABLE_INT_FETCH("kern.ipc.msgmni", &msginfo.msgmni);
139
140	msgpool = malloc(msginfo.msgmax, M_MSG, M_WAITOK);
141	if (msgpool == NULL)
142		panic("msgpool is NULL");
143	msgmaps = malloc(sizeof(struct msgmap) * msginfo.msgseg, M_MSG, M_WAITOK);
144	if (msgmaps == NULL)
145		panic("msgmaps is NULL");
146	msghdrs = malloc(sizeof(struct msg) * msginfo.msgtql, M_MSG, M_WAITOK);
147	if (msghdrs == NULL)
148		panic("msghdrs is NULL");
149	msqids = malloc(sizeof(struct msqid_ds) * msginfo.msgmni, M_MSG, M_WAITOK);
150	if (msqids == NULL)
151		panic("msqids is NULL");
152
153	/*
154	 * msginfo.msgssz should be a power of two for efficiency reasons.
155	 * It is also pretty silly if msginfo.msgssz is less than 8
156	 * or greater than about 256 so ...
157	 */
158
159	i = 8;
160	while (i < 1024 && i != msginfo.msgssz)
161		i <<= 1;
162    	if (i != msginfo.msgssz) {
163		printf("msginfo.msgssz=%d (0x%x)\n", msginfo.msgssz,
164		    msginfo.msgssz);
165		panic("msginfo.msgssz not a small power of 2");
166	}
167
168	if (msginfo.msgseg > 32767) {
169		printf("msginfo.msgseg=%d\n", msginfo.msgseg);
170		panic("msginfo.msgseg > 32767");
171	}
172
173	if (msgmaps == NULL)
174		panic("msgmaps is NULL");
175
176	for (i = 0; i < msginfo.msgseg; i++) {
177		if (i > 0)
178			msgmaps[i-1].next = i;
179		msgmaps[i].next = -1;	/* implies entry is available */
180	}
181	free_msgmaps = 0;
182	nfree_msgmaps = msginfo.msgseg;
183
184	if (msghdrs == NULL)
185		panic("msghdrs is NULL");
186
187	for (i = 0; i < msginfo.msgtql; i++) {
188		msghdrs[i].msg_type = 0;
189		if (i > 0)
190			msghdrs[i-1].msg_next = &msghdrs[i];
191		msghdrs[i].msg_next = NULL;
192    	}
193	free_msghdrs = &msghdrs[0];
194
195	if (msqids == NULL)
196		panic("msqids is NULL");
197
198	for (i = 0; i < msginfo.msgmni; i++) {
199		msqids[i].msg_qbytes = 0;	/* implies entry is available */
200		msqids[i].msg_perm.seq = 0;	/* reset to a known value */
201		msqids[i].msg_perm.mode = 0;
202	}
203}
204
205static int
206msgunload()
207{
208	struct msqid_ds *msqptr;
209	int msqid;
210
211	for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
212		/*
213		 * Look for an unallocated and unlocked msqid_ds.
214		 * msqid_ds's can be locked by msgsnd or msgrcv while
215		 * they are copying the message in/out.  We can't
216		 * re-use the entry until they release it.
217		 */
218		msqptr = &msqids[msqid];
219		if (msqptr->msg_qbytes != 0 ||
220		    (msqptr->msg_perm.mode & MSG_LOCKED) != 0)
221			break;
222	}
223	if (msqid != msginfo.msgmni)
224		return (EBUSY);
225
226	free(msgpool, M_MSG);
227	free(msgmaps, M_MSG);
228	free(msghdrs, M_MSG);
229	free(msqids, M_MSG);
230	return (0);
231}
232
233
234static int
235sysvmsg_modload(struct module *module, int cmd, void *arg)
236{
237	int error = 0;
238
239	switch (cmd) {
240	case MOD_LOAD:
241		msginit();
242		break;
243	case MOD_UNLOAD:
244		error = msgunload();
245		break;
246	case MOD_SHUTDOWN:
247		break;
248	default:
249		error = EINVAL;
250		break;
251	}
252	return (error);
253}
254
255static moduledata_t sysvmsg_mod = {
256	"sysvmsg",
257	&sysvmsg_modload,
258	NULL
259};
260
261SYSCALL_MODULE_HELPER(msgsys);
262SYSCALL_MODULE_HELPER(msgctl);
263SYSCALL_MODULE_HELPER(msgget);
264SYSCALL_MODULE_HELPER(msgsnd);
265SYSCALL_MODULE_HELPER(msgrcv);
266
267DECLARE_MODULE(sysvmsg, sysvmsg_mod,
268	SI_SUB_SYSV_MSG, SI_ORDER_FIRST);
269MODULE_VERSION(sysvmsg, 1);
270
271/*
272 * Entry point for all MSG calls
273 *
274 * MPSAFE
275 */
276int
277msgsys(td, uap)
278	struct thread *td;
279	/* XXX actually varargs. */
280	struct msgsys_args /* {
281		u_int	which;
282		int	a2;
283		int	a3;
284		int	a4;
285		int	a5;
286		int	a6;
287	} */ *uap;
288{
289	int error;
290
291	mtx_lock(&Giant);
292	if (!jail_sysvipc_allowed && jailed(td->td_proc->p_ucred)) {
293		error = ENOSYS;
294		goto done2;
295	}
296	if (uap->which >= sizeof(msgcalls)/sizeof(msgcalls[0])) {
297		error = EINVAL;
298		goto done2;
299	}
300	error = (*msgcalls[uap->which])(td, &uap->a2);
301done2:
302	mtx_unlock(&Giant);
303	return (error);
304}
305
306static void
307msg_freehdr(msghdr)
308	struct msg *msghdr;
309{
310	while (msghdr->msg_ts > 0) {
311		short next;
312		if (msghdr->msg_spot < 0 || msghdr->msg_spot >= msginfo.msgseg)
313			panic("msghdr->msg_spot out of range");
314		next = msgmaps[msghdr->msg_spot].next;
315		msgmaps[msghdr->msg_spot].next = free_msgmaps;
316		free_msgmaps = msghdr->msg_spot;
317		nfree_msgmaps++;
318		msghdr->msg_spot = next;
319		if (msghdr->msg_ts >= msginfo.msgssz)
320			msghdr->msg_ts -= msginfo.msgssz;
321		else
322			msghdr->msg_ts = 0;
323	}
324	if (msghdr->msg_spot != -1)
325		panic("msghdr->msg_spot != -1");
326	msghdr->msg_next = free_msghdrs;
327	free_msghdrs = msghdr;
328}
329
330#ifndef _SYS_SYSPROTO_H_
331struct msgctl_args {
332	int	msqid;
333	int	cmd;
334	struct	msqid_ds *buf;
335};
336#endif
337
338/*
339 * MPSAFE
340 */
341int
342msgctl(td, uap)
343	struct thread *td;
344	register struct msgctl_args *uap;
345{
346	int msqid = uap->msqid;
347	int cmd = uap->cmd;
348	struct msqid_ds *user_msqptr = uap->buf;
349	int rval, error;
350	struct msqid_ds msqbuf;
351	register struct msqid_ds *msqptr;
352
353#ifdef MSG_DEBUG_OK
354	printf("call to msgctl(%d, %d, 0x%x)\n", msqid, cmd, user_msqptr);
355#endif
356	mtx_lock(&Giant);
357	if (!jail_sysvipc_allowed && jailed(td->td_proc->p_ucred)) {
358		error = ENOSYS;
359		goto done2;
360	}
361
362	msqid = IPCID_TO_IX(msqid);
363
364	if (msqid < 0 || msqid >= msginfo.msgmni) {
365#ifdef MSG_DEBUG_OK
366		printf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
367		    msginfo.msgmni);
368#endif
369		error = EINVAL;
370		goto done2;
371	}
372
373	msqptr = &msqids[msqid];
374
375	if (msqptr->msg_qbytes == 0) {
376#ifdef MSG_DEBUG_OK
377		printf("no such msqid\n");
378#endif
379		error = EINVAL;
380		goto done2;
381	}
382	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
383#ifdef MSG_DEBUG_OK
384		printf("wrong sequence number\n");
385#endif
386		error = EINVAL;
387		goto done2;
388	}
389
390	error = 0;
391	rval = 0;
392
393	switch (cmd) {
394
395	case IPC_RMID:
396	{
397		struct msg *msghdr;
398		if ((error = ipcperm(td, &msqptr->msg_perm, IPC_M)))
399			goto done2;
400		/* Free the message headers */
401		msghdr = msqptr->msg_first;
402		while (msghdr != NULL) {
403			struct msg *msghdr_tmp;
404
405			/* Free the segments of each message */
406			msqptr->msg_cbytes -= msghdr->msg_ts;
407			msqptr->msg_qnum--;
408			msghdr_tmp = msghdr;
409			msghdr = msghdr->msg_next;
410			msg_freehdr(msghdr_tmp);
411		}
412
413		if (msqptr->msg_cbytes != 0)
414			panic("msg_cbytes is screwed up");
415		if (msqptr->msg_qnum != 0)
416			panic("msg_qnum is screwed up");
417
418		msqptr->msg_qbytes = 0;	/* Mark it as free */
419
420		wakeup((caddr_t)msqptr);
421	}
422
423		break;
424
425	case IPC_SET:
426		if ((error = ipcperm(td, &msqptr->msg_perm, IPC_M)))
427			goto done2;
428		if ((error = copyin(user_msqptr, &msqbuf, sizeof(msqbuf))) != 0)
429			goto done2;
430		if (msqbuf.msg_qbytes > msqptr->msg_qbytes) {
431			error = suser_td(td);
432			if (error)
433				goto done2;
434		}
435		if (msqbuf.msg_qbytes > msginfo.msgmnb) {
436#ifdef MSG_DEBUG_OK
437			printf("can't increase msg_qbytes beyond %d (truncating)\n",
438			    msginfo.msgmnb);
439#endif
440			msqbuf.msg_qbytes = msginfo.msgmnb;	/* silently restrict qbytes to system limit */
441		}
442		if (msqbuf.msg_qbytes == 0) {
443#ifdef MSG_DEBUG_OK
444			printf("can't reduce msg_qbytes to 0\n");
445#endif
446			error = EINVAL;		/* non-standard errno! */
447			goto done2;
448		}
449		msqptr->msg_perm.uid = msqbuf.msg_perm.uid;	/* change the owner */
450		msqptr->msg_perm.gid = msqbuf.msg_perm.gid;	/* change the owner */
451		msqptr->msg_perm.mode = (msqptr->msg_perm.mode & ~0777) |
452		    (msqbuf.msg_perm.mode & 0777);
453		msqptr->msg_qbytes = msqbuf.msg_qbytes;
454		msqptr->msg_ctime = time_second;
455		break;
456
457	case IPC_STAT:
458		if ((error = ipcperm(td, &msqptr->msg_perm, IPC_R))) {
459#ifdef MSG_DEBUG_OK
460			printf("requester doesn't have read access\n");
461#endif
462			goto done2;
463		}
464		error = copyout((caddr_t)msqptr, user_msqptr,
465		    sizeof(struct msqid_ds));
466		break;
467
468	default:
469#ifdef MSG_DEBUG_OK
470		printf("invalid command %d\n", cmd);
471#endif
472		error = EINVAL;
473		goto done2;
474	}
475
476	if (error == 0)
477		td->td_retval[0] = rval;
478done2:
479	mtx_unlock(&Giant);
480	return(error);
481}
482
483#ifndef _SYS_SYSPROTO_H_
484struct msgget_args {
485	key_t	key;
486	int	msgflg;
487};
488#endif
489
490/*
491 * MPSAFE
492 */
493int
494msgget(td, uap)
495	struct thread *td;
496	register struct msgget_args *uap;
497{
498	int msqid, error = 0;
499	int key = uap->key;
500	int msgflg = uap->msgflg;
501	struct ucred *cred = td->td_proc->p_ucred;
502	register struct msqid_ds *msqptr = NULL;
503
504#ifdef MSG_DEBUG_OK
505	printf("msgget(0x%x, 0%o)\n", key, msgflg);
506#endif
507
508	mtx_lock(&Giant);
509	if (!jail_sysvipc_allowed && jailed(td->td_proc->p_ucred)) {
510		error = ENOSYS;
511		goto done2;
512	}
513
514	if (key != IPC_PRIVATE) {
515		for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
516			msqptr = &msqids[msqid];
517			if (msqptr->msg_qbytes != 0 &&
518			    msqptr->msg_perm.key == key)
519				break;
520		}
521		if (msqid < msginfo.msgmni) {
522#ifdef MSG_DEBUG_OK
523			printf("found public key\n");
524#endif
525			if ((msgflg & IPC_CREAT) && (msgflg & IPC_EXCL)) {
526#ifdef MSG_DEBUG_OK
527				printf("not exclusive\n");
528#endif
529				error = EEXIST;
530				goto done2;
531			}
532			if ((error = ipcperm(td, &msqptr->msg_perm, msgflg & 0700 ))) {
533#ifdef MSG_DEBUG_OK
534				printf("requester doesn't have 0%o access\n",
535				    msgflg & 0700);
536#endif
537				goto done2;
538			}
539			goto found;
540		}
541	}
542
543#ifdef MSG_DEBUG_OK
544	printf("need to allocate the msqid_ds\n");
545#endif
546	if (key == IPC_PRIVATE || (msgflg & IPC_CREAT)) {
547		for (msqid = 0; msqid < msginfo.msgmni; msqid++) {
548			/*
549			 * Look for an unallocated and unlocked msqid_ds.
550			 * msqid_ds's can be locked by msgsnd or msgrcv while
551			 * they are copying the message in/out.  We can't
552			 * re-use the entry until they release it.
553			 */
554			msqptr = &msqids[msqid];
555			if (msqptr->msg_qbytes == 0 &&
556			    (msqptr->msg_perm.mode & MSG_LOCKED) == 0)
557				break;
558		}
559		if (msqid == msginfo.msgmni) {
560#ifdef MSG_DEBUG_OK
561			printf("no more msqid_ds's available\n");
562#endif
563			error = ENOSPC;
564			goto done2;
565		}
566#ifdef MSG_DEBUG_OK
567		printf("msqid %d is available\n", msqid);
568#endif
569		msqptr->msg_perm.key = key;
570		msqptr->msg_perm.cuid = cred->cr_uid;
571		msqptr->msg_perm.uid = cred->cr_uid;
572		msqptr->msg_perm.cgid = cred->cr_gid;
573		msqptr->msg_perm.gid = cred->cr_gid;
574		msqptr->msg_perm.mode = (msgflg & 0777);
575		/* Make sure that the returned msqid is unique */
576		msqptr->msg_perm.seq++;
577		msqptr->msg_first = NULL;
578		msqptr->msg_last = NULL;
579		msqptr->msg_cbytes = 0;
580		msqptr->msg_qnum = 0;
581		msqptr->msg_qbytes = msginfo.msgmnb;
582		msqptr->msg_lspid = 0;
583		msqptr->msg_lrpid = 0;
584		msqptr->msg_stime = 0;
585		msqptr->msg_rtime = 0;
586		msqptr->msg_ctime = time_second;
587	} else {
588#ifdef MSG_DEBUG_OK
589		printf("didn't find it and wasn't asked to create it\n");
590#endif
591		error = ENOENT;
592		goto done2;
593	}
594
595found:
596	/* Construct the unique msqid */
597	td->td_retval[0] = IXSEQ_TO_IPCID(msqid, msqptr->msg_perm);
598done2:
599	mtx_unlock(&Giant);
600	return (error);
601}
602
603#ifndef _SYS_SYSPROTO_H_
604struct msgsnd_args {
605	int	msqid;
606	void	*msgp;
607	size_t	msgsz;
608	int	msgflg;
609};
610#endif
611
612/*
613 * MPSAFE
614 */
615int
616msgsnd(td, uap)
617	struct thread *td;
618	register struct msgsnd_args *uap;
619{
620	int msqid = uap->msqid;
621	void *user_msgp = uap->msgp;
622	size_t msgsz = uap->msgsz;
623	int msgflg = uap->msgflg;
624	int segs_needed, error = 0;
625	register struct msqid_ds *msqptr;
626	register struct msg *msghdr;
627	short next;
628
629#ifdef MSG_DEBUG_OK
630	printf("call to msgsnd(%d, 0x%x, %d, %d)\n", msqid, user_msgp, msgsz,
631	    msgflg);
632#endif
633	mtx_lock(&Giant);
634	if (!jail_sysvipc_allowed && jailed(td->td_proc->p_ucred)) {
635		error = ENOSYS;
636		goto done2;
637	}
638
639	msqid = IPCID_TO_IX(msqid);
640
641	if (msqid < 0 || msqid >= msginfo.msgmni) {
642#ifdef MSG_DEBUG_OK
643		printf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
644		    msginfo.msgmni);
645#endif
646		error = EINVAL;
647		goto done2;
648	}
649
650	msqptr = &msqids[msqid];
651	if (msqptr->msg_qbytes == 0) {
652#ifdef MSG_DEBUG_OK
653		printf("no such message queue id\n");
654#endif
655		error = EINVAL;
656		goto done2;
657	}
658	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
659#ifdef MSG_DEBUG_OK
660		printf("wrong sequence number\n");
661#endif
662		error = EINVAL;
663		goto done2;
664	}
665
666	if ((error = ipcperm(td, &msqptr->msg_perm, IPC_W))) {
667#ifdef MSG_DEBUG_OK
668		printf("requester doesn't have write access\n");
669#endif
670		goto done2;
671	}
672
673	segs_needed = (msgsz + msginfo.msgssz - 1) / msginfo.msgssz;
674#ifdef MSG_DEBUG_OK
675	printf("msgsz=%d, msgssz=%d, segs_needed=%d\n", msgsz, msginfo.msgssz,
676	    segs_needed);
677#endif
678	for (;;) {
679		int need_more_resources = 0;
680
681		/*
682		 * check msgsz
683		 * (inside this loop in case msg_qbytes changes while we sleep)
684		 */
685
686		if (msgsz > msqptr->msg_qbytes) {
687#ifdef MSG_DEBUG_OK
688			printf("msgsz > msqptr->msg_qbytes\n");
689#endif
690			error = EINVAL;
691			goto done2;
692		}
693
694		if (msqptr->msg_perm.mode & MSG_LOCKED) {
695#ifdef MSG_DEBUG_OK
696			printf("msqid is locked\n");
697#endif
698			need_more_resources = 1;
699		}
700		if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes) {
701#ifdef MSG_DEBUG_OK
702			printf("msgsz + msg_cbytes > msg_qbytes\n");
703#endif
704			need_more_resources = 1;
705		}
706		if (segs_needed > nfree_msgmaps) {
707#ifdef MSG_DEBUG_OK
708			printf("segs_needed > nfree_msgmaps\n");
709#endif
710			need_more_resources = 1;
711		}
712		if (free_msghdrs == NULL) {
713#ifdef MSG_DEBUG_OK
714			printf("no more msghdrs\n");
715#endif
716			need_more_resources = 1;
717		}
718
719		if (need_more_resources) {
720			int we_own_it;
721
722			if ((msgflg & IPC_NOWAIT) != 0) {
723#ifdef MSG_DEBUG_OK
724				printf("need more resources but caller doesn't want to wait\n");
725#endif
726				error = EAGAIN;
727				goto done2;
728			}
729
730			if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0) {
731#ifdef MSG_DEBUG_OK
732				printf("we don't own the msqid_ds\n");
733#endif
734				we_own_it = 0;
735			} else {
736				/* Force later arrivals to wait for our
737				   request */
738#ifdef MSG_DEBUG_OK
739				printf("we own the msqid_ds\n");
740#endif
741				msqptr->msg_perm.mode |= MSG_LOCKED;
742				we_own_it = 1;
743			}
744#ifdef MSG_DEBUG_OK
745			printf("goodnight\n");
746#endif
747			error = tsleep((caddr_t)msqptr, (PZERO - 4) | PCATCH,
748			    "msgwait", 0);
749#ifdef MSG_DEBUG_OK
750			printf("good morning, error=%d\n", error);
751#endif
752			if (we_own_it)
753				msqptr->msg_perm.mode &= ~MSG_LOCKED;
754			if (error != 0) {
755#ifdef MSG_DEBUG_OK
756				printf("msgsnd:  interrupted system call\n");
757#endif
758				error = EINTR;
759				goto done2;
760			}
761
762			/*
763			 * Make sure that the msq queue still exists
764			 */
765
766			if (msqptr->msg_qbytes == 0) {
767#ifdef MSG_DEBUG_OK
768				printf("msqid deleted\n");
769#endif
770				error = EIDRM;
771				goto done2;
772			}
773
774		} else {
775#ifdef MSG_DEBUG_OK
776			printf("got all the resources that we need\n");
777#endif
778			break;
779		}
780	}
781
782	/*
783	 * We have the resources that we need.
784	 * Make sure!
785	 */
786
787	if (msqptr->msg_perm.mode & MSG_LOCKED)
788		panic("msg_perm.mode & MSG_LOCKED");
789	if (segs_needed > nfree_msgmaps)
790		panic("segs_needed > nfree_msgmaps");
791	if (msgsz + msqptr->msg_cbytes > msqptr->msg_qbytes)
792		panic("msgsz + msg_cbytes > msg_qbytes");
793	if (free_msghdrs == NULL)
794		panic("no more msghdrs");
795
796	/*
797	 * Re-lock the msqid_ds in case we page-fault when copying in the
798	 * message
799	 */
800
801	if ((msqptr->msg_perm.mode & MSG_LOCKED) != 0)
802		panic("msqid_ds is already locked");
803	msqptr->msg_perm.mode |= MSG_LOCKED;
804
805	/*
806	 * Allocate a message header
807	 */
808
809	msghdr = free_msghdrs;
810	free_msghdrs = msghdr->msg_next;
811	msghdr->msg_spot = -1;
812	msghdr->msg_ts = msgsz;
813
814	/*
815	 * Allocate space for the message
816	 */
817
818	while (segs_needed > 0) {
819		if (nfree_msgmaps <= 0)
820			panic("not enough msgmaps");
821		if (free_msgmaps == -1)
822			panic("nil free_msgmaps");
823		next = free_msgmaps;
824		if (next <= -1)
825			panic("next too low #1");
826		if (next >= msginfo.msgseg)
827			panic("next out of range #1");
828#ifdef MSG_DEBUG_OK
829		printf("allocating segment %d to message\n", next);
830#endif
831		free_msgmaps = msgmaps[next].next;
832		nfree_msgmaps--;
833		msgmaps[next].next = msghdr->msg_spot;
834		msghdr->msg_spot = next;
835		segs_needed--;
836	}
837
838	/*
839	 * Copy in the message type
840	 */
841
842	if ((error = copyin(user_msgp, &msghdr->msg_type,
843	    sizeof(msghdr->msg_type))) != 0) {
844#ifdef MSG_DEBUG_OK
845		printf("error %d copying the message type\n", error);
846#endif
847		msg_freehdr(msghdr);
848		msqptr->msg_perm.mode &= ~MSG_LOCKED;
849		wakeup((caddr_t)msqptr);
850		goto done2;
851	}
852	user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
853
854	/*
855	 * Validate the message type
856	 */
857
858	if (msghdr->msg_type < 1) {
859		msg_freehdr(msghdr);
860		msqptr->msg_perm.mode &= ~MSG_LOCKED;
861		wakeup((caddr_t)msqptr);
862#ifdef MSG_DEBUG_OK
863		printf("mtype (%d) < 1\n", msghdr->msg_type);
864#endif
865		error = EINVAL;
866		goto done2;
867	}
868
869	/*
870	 * Copy in the message body
871	 */
872
873	next = msghdr->msg_spot;
874	while (msgsz > 0) {
875		size_t tlen;
876		if (msgsz > msginfo.msgssz)
877			tlen = msginfo.msgssz;
878		else
879			tlen = msgsz;
880		if (next <= -1)
881			panic("next too low #2");
882		if (next >= msginfo.msgseg)
883			panic("next out of range #2");
884		if ((error = copyin(user_msgp, &msgpool[next * msginfo.msgssz],
885		    tlen)) != 0) {
886#ifdef MSG_DEBUG_OK
887			printf("error %d copying in message segment\n", error);
888#endif
889			msg_freehdr(msghdr);
890			msqptr->msg_perm.mode &= ~MSG_LOCKED;
891			wakeup((caddr_t)msqptr);
892			goto done2;
893		}
894		msgsz -= tlen;
895		user_msgp = (char *)user_msgp + tlen;
896		next = msgmaps[next].next;
897	}
898	if (next != -1)
899		panic("didn't use all the msg segments");
900
901	/*
902	 * We've got the message.  Unlock the msqid_ds.
903	 */
904
905	msqptr->msg_perm.mode &= ~MSG_LOCKED;
906
907	/*
908	 * Make sure that the msqid_ds is still allocated.
909	 */
910
911	if (msqptr->msg_qbytes == 0) {
912		msg_freehdr(msghdr);
913		wakeup((caddr_t)msqptr);
914		error = EIDRM;
915		goto done2;
916	}
917
918	/*
919	 * Put the message into the queue
920	 */
921
922	if (msqptr->msg_first == NULL) {
923		msqptr->msg_first = msghdr;
924		msqptr->msg_last = msghdr;
925	} else {
926		msqptr->msg_last->msg_next = msghdr;
927		msqptr->msg_last = msghdr;
928	}
929	msqptr->msg_last->msg_next = NULL;
930
931	msqptr->msg_cbytes += msghdr->msg_ts;
932	msqptr->msg_qnum++;
933	msqptr->msg_lspid = td->td_proc->p_pid;
934	msqptr->msg_stime = time_second;
935
936	wakeup((caddr_t)msqptr);
937	td->td_retval[0] = 0;
938done2:
939	mtx_unlock(&Giant);
940	return (error);
941}
942
943#ifndef _SYS_SYSPROTO_H_
944struct msgrcv_args {
945	int	msqid;
946	void	*msgp;
947	size_t	msgsz;
948	long	msgtyp;
949	int	msgflg;
950};
951#endif
952
953/*
954 * MPSAFE
955 */
956int
957msgrcv(td, uap)
958	struct thread *td;
959	register struct msgrcv_args *uap;
960{
961	int msqid = uap->msqid;
962	void *user_msgp = uap->msgp;
963	size_t msgsz = uap->msgsz;
964	long msgtyp = uap->msgtyp;
965	int msgflg = uap->msgflg;
966	size_t len;
967	register struct msqid_ds *msqptr;
968	register struct msg *msghdr;
969	int error = 0;
970	short next;
971
972#ifdef MSG_DEBUG_OK
973	printf("call to msgrcv(%d, 0x%x, %d, %ld, %d)\n", msqid, user_msgp,
974	    msgsz, msgtyp, msgflg);
975#endif
976
977	mtx_lock(&Giant);
978	if (!jail_sysvipc_allowed && jailed(td->td_proc->p_ucred)) {
979		error = ENOSYS;
980		goto done2;
981	}
982
983	msqid = IPCID_TO_IX(msqid);
984
985	if (msqid < 0 || msqid >= msginfo.msgmni) {
986#ifdef MSG_DEBUG_OK
987		printf("msqid (%d) out of range (0<=msqid<%d)\n", msqid,
988		    msginfo.msgmni);
989#endif
990		error = EINVAL;
991		goto done2;
992	}
993
994	msqptr = &msqids[msqid];
995	if (msqptr->msg_qbytes == 0) {
996#ifdef MSG_DEBUG_OK
997		printf("no such message queue id\n");
998#endif
999		error = EINVAL;
1000		goto done2;
1001	}
1002	if (msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
1003#ifdef MSG_DEBUG_OK
1004		printf("wrong sequence number\n");
1005#endif
1006		error = EINVAL;
1007		goto done2;
1008	}
1009
1010	if ((error = ipcperm(td, &msqptr->msg_perm, IPC_R))) {
1011#ifdef MSG_DEBUG_OK
1012		printf("requester doesn't have read access\n");
1013#endif
1014		goto done2;
1015	}
1016
1017	msghdr = NULL;
1018	while (msghdr == NULL) {
1019		if (msgtyp == 0) {
1020			msghdr = msqptr->msg_first;
1021			if (msghdr != NULL) {
1022				if (msgsz < msghdr->msg_ts &&
1023				    (msgflg & MSG_NOERROR) == 0) {
1024#ifdef MSG_DEBUG_OK
1025					printf("first message on the queue is too big (want %d, got %d)\n",
1026					    msgsz, msghdr->msg_ts);
1027#endif
1028					error = E2BIG;
1029					goto done2;
1030				}
1031				if (msqptr->msg_first == msqptr->msg_last) {
1032					msqptr->msg_first = NULL;
1033					msqptr->msg_last = NULL;
1034				} else {
1035					msqptr->msg_first = msghdr->msg_next;
1036					if (msqptr->msg_first == NULL)
1037						panic("msg_first/last screwed up #1");
1038				}
1039			}
1040		} else {
1041			struct msg *previous;
1042			struct msg **prev;
1043
1044			previous = NULL;
1045			prev = &(msqptr->msg_first);
1046			while ((msghdr = *prev) != NULL) {
1047				/*
1048				 * Is this message's type an exact match or is
1049				 * this message's type less than or equal to
1050				 * the absolute value of a negative msgtyp?
1051				 * Note that the second half of this test can
1052				 * NEVER be true if msgtyp is positive since
1053				 * msg_type is always positive!
1054				 */
1055
1056				if (msgtyp == msghdr->msg_type ||
1057				    msghdr->msg_type <= -msgtyp) {
1058#ifdef MSG_DEBUG_OK
1059					printf("found message type %d, requested %d\n",
1060					    msghdr->msg_type, msgtyp);
1061#endif
1062					if (msgsz < msghdr->msg_ts &&
1063					    (msgflg & MSG_NOERROR) == 0) {
1064#ifdef MSG_DEBUG_OK
1065						printf("requested message on the queue is too big (want %d, got %d)\n",
1066						    msgsz, msghdr->msg_ts);
1067#endif
1068						error = E2BIG;
1069						goto done2;
1070					}
1071					*prev = msghdr->msg_next;
1072					if (msghdr == msqptr->msg_last) {
1073						if (previous == NULL) {
1074							if (prev !=
1075							    &msqptr->msg_first)
1076								panic("msg_first/last screwed up #2");
1077							msqptr->msg_first =
1078							    NULL;
1079							msqptr->msg_last =
1080							    NULL;
1081						} else {
1082							if (prev ==
1083							    &msqptr->msg_first)
1084								panic("msg_first/last screwed up #3");
1085							msqptr->msg_last =
1086							    previous;
1087						}
1088					}
1089					break;
1090				}
1091				previous = msghdr;
1092				prev = &(msghdr->msg_next);
1093			}
1094		}
1095
1096		/*
1097		 * We've either extracted the msghdr for the appropriate
1098		 * message or there isn't one.
1099		 * If there is one then bail out of this loop.
1100		 */
1101
1102		if (msghdr != NULL)
1103			break;
1104
1105		/*
1106		 * Hmph!  No message found.  Does the user want to wait?
1107		 */
1108
1109		if ((msgflg & IPC_NOWAIT) != 0) {
1110#ifdef MSG_DEBUG_OK
1111			printf("no appropriate message found (msgtyp=%d)\n",
1112			    msgtyp);
1113#endif
1114			/* The SVID says to return ENOMSG. */
1115			error = ENOMSG;
1116			goto done2;
1117		}
1118
1119		/*
1120		 * Wait for something to happen
1121		 */
1122
1123#ifdef MSG_DEBUG_OK
1124		printf("msgrcv:  goodnight\n");
1125#endif
1126		error = tsleep((caddr_t)msqptr, (PZERO - 4) | PCATCH, "msgwait",
1127		    0);
1128#ifdef MSG_DEBUG_OK
1129		printf("msgrcv:  good morning (error=%d)\n", error);
1130#endif
1131
1132		if (error != 0) {
1133#ifdef MSG_DEBUG_OK
1134			printf("msgsnd:  interrupted system call\n");
1135#endif
1136			error = EINTR;
1137			goto done2;
1138		}
1139
1140		/*
1141		 * Make sure that the msq queue still exists
1142		 */
1143
1144		if (msqptr->msg_qbytes == 0 ||
1145		    msqptr->msg_perm.seq != IPCID_TO_SEQ(uap->msqid)) {
1146#ifdef MSG_DEBUG_OK
1147			printf("msqid deleted\n");
1148#endif
1149			error = EIDRM;
1150			goto done2;
1151		}
1152	}
1153
1154	/*
1155	 * Return the message to the user.
1156	 *
1157	 * First, do the bookkeeping (before we risk being interrupted).
1158	 */
1159
1160	msqptr->msg_cbytes -= msghdr->msg_ts;
1161	msqptr->msg_qnum--;
1162	msqptr->msg_lrpid = td->td_proc->p_pid;
1163	msqptr->msg_rtime = time_second;
1164
1165	/*
1166	 * Make msgsz the actual amount that we'll be returning.
1167	 * Note that this effectively truncates the message if it is too long
1168	 * (since msgsz is never increased).
1169	 */
1170
1171#ifdef MSG_DEBUG_OK
1172	printf("found a message, msgsz=%d, msg_ts=%d\n", msgsz,
1173	    msghdr->msg_ts);
1174#endif
1175	if (msgsz > msghdr->msg_ts)
1176		msgsz = msghdr->msg_ts;
1177
1178	/*
1179	 * Return the type to the user.
1180	 */
1181
1182	error = copyout((caddr_t)&(msghdr->msg_type), user_msgp,
1183	    sizeof(msghdr->msg_type));
1184	if (error != 0) {
1185#ifdef MSG_DEBUG_OK
1186		printf("error (%d) copying out message type\n", error);
1187#endif
1188		msg_freehdr(msghdr);
1189		wakeup((caddr_t)msqptr);
1190		goto done2;
1191	}
1192	user_msgp = (char *)user_msgp + sizeof(msghdr->msg_type);
1193
1194	/*
1195	 * Return the segments to the user
1196	 */
1197
1198	next = msghdr->msg_spot;
1199	for (len = 0; len < msgsz; len += msginfo.msgssz) {
1200		size_t tlen;
1201
1202		if (msgsz - len > msginfo.msgssz)
1203			tlen = msginfo.msgssz;
1204		else
1205			tlen = msgsz - len;
1206		if (next <= -1)
1207			panic("next too low #3");
1208		if (next >= msginfo.msgseg)
1209			panic("next out of range #3");
1210		error = copyout((caddr_t)&msgpool[next * msginfo.msgssz],
1211		    user_msgp, tlen);
1212		if (error != 0) {
1213#ifdef MSG_DEBUG_OK
1214			printf("error (%d) copying out message segment\n",
1215			    error);
1216#endif
1217			msg_freehdr(msghdr);
1218			wakeup((caddr_t)msqptr);
1219			goto done2;
1220		}
1221		user_msgp = (char *)user_msgp + tlen;
1222		next = msgmaps[next].next;
1223	}
1224
1225	/*
1226	 * Done, return the actual number of bytes copied out.
1227	 */
1228
1229	msg_freehdr(msghdr);
1230	wakeup((caddr_t)msqptr);
1231	td->td_retval[0] = msgsz;
1232done2:
1233	mtx_unlock(&Giant);
1234	return (error);
1235}
1236
1237static int
1238sysctl_msqids(SYSCTL_HANDLER_ARGS)
1239{
1240
1241	return (SYSCTL_OUT(req, msqids,
1242	    sizeof(struct msqid_ds) * msginfo.msgmni));
1243}
1244
1245SYSCTL_DECL(_kern_ipc);
1246SYSCTL_INT(_kern_ipc, OID_AUTO, msgmax, CTLFLAG_RD, &msginfo.msgmax, 0, "");
1247SYSCTL_INT(_kern_ipc, OID_AUTO, msgmni, CTLFLAG_RD, &msginfo.msgmni, 0, "");
1248SYSCTL_INT(_kern_ipc, OID_AUTO, msgmnb, CTLFLAG_RD, &msginfo.msgmnb, 0, "");
1249SYSCTL_INT(_kern_ipc, OID_AUTO, msgtql, CTLFLAG_RD, &msginfo.msgtql, 0, "");
1250SYSCTL_INT(_kern_ipc, OID_AUTO, msgssz, CTLFLAG_RD, &msginfo.msgssz, 0, "");
1251SYSCTL_INT(_kern_ipc, OID_AUTO, msgseg, CTLFLAG_RD, &msginfo.msgseg, 0, "")
1252SYSCTL_PROC(_kern_ipc, OID_AUTO, msqids, CTLFLAG_RD,
1253    NULL, 0, sysctl_msqids, "", "Message queue IDs");
1254