Home | History | Annotate | Line # | Download | only in kern
sys_generic.c revision 1.31
      1 /*	$NetBSD: sys_generic.c,v 1.31 1996/12/22 10:21:12 cgd Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 1982, 1986, 1989, 1993
      5  *	The Regents of the University of California.  All rights reserved.
      6  * (c) UNIX System Laboratories, Inc.
      7  * All or some portions of this file are derived from material licensed
      8  * to the University of California by American Telephone and Telegraph
      9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
     10  * the permission of UNIX System Laboratories, Inc.
     11  *
     12  * Redistribution and use in source and binary forms, with or without
     13  * modification, are permitted provided that the following conditions
     14  * are met:
     15  * 1. Redistributions of source code must retain the above copyright
     16  *    notice, this list of conditions and the following disclaimer.
     17  * 2. Redistributions in binary form must reproduce the above copyright
     18  *    notice, this list of conditions and the following disclaimer in the
     19  *    documentation and/or other materials provided with the distribution.
     20  * 3. All advertising materials mentioning features or use of this software
     21  *    must display the following acknowledgement:
     22  *	This product includes software developed by the University of
     23  *	California, Berkeley and its contributors.
     24  * 4. Neither the name of the University nor the names of its contributors
     25  *    may be used to endorse or promote products derived from this software
     26  *    without specific prior written permission.
     27  *
     28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     38  * SUCH DAMAGE.
     39  *
     40  *	@(#)sys_generic.c	8.5 (Berkeley) 1/21/94
     41  */
     42 
     43 #include <sys/param.h>
     44 #include <sys/systm.h>
     45 #include <sys/filedesc.h>
     46 #include <sys/ioctl.h>
     47 #include <sys/file.h>
     48 #include <sys/proc.h>
     49 #include <sys/socketvar.h>
     50 #include <sys/signalvar.h>
     51 #include <sys/uio.h>
     52 #include <sys/kernel.h>
     53 #include <sys/stat.h>
     54 #include <sys/malloc.h>
     55 #include <sys/poll.h>
     56 #ifdef KTRACE
     57 #include <sys/ktrace.h>
     58 #endif
     59 
     60 #include <sys/mount.h>
     61 #include <sys/syscallargs.h>
     62 
     63 int selscan __P((struct proc *, fd_mask *, fd_mask *, int, register_t *));
     64 int pollscan __P((struct proc *, struct pollfd *, int, register_t *));
     65 
     66 /*
     67  * Read system call.
     68  */
     69 /* ARGSUSED */
     70 int
     71 sys_read(p, v, retval)
     72 	struct proc *p;
     73 	void *v;
     74 	register_t *retval;
     75 {
     76 	register struct sys_read_args /* {
     77 		syscallarg(int) fd;
     78 		syscallarg(void *) buf;
     79 		syscallarg(size_t) nbyte;
     80 	} */ *uap = v;
     81 	register struct file *fp;
     82 	register struct filedesc *fdp = p->p_fd;
     83 	struct uio auio;
     84 	struct iovec aiov;
     85 	long cnt, error = 0;
     86 #ifdef KTRACE
     87 	struct iovec ktriov;
     88 #endif
     89 
     90 	if (((u_int)SCARG(uap, fd)) >= fdp->fd_nfiles ||
     91 	    (fp = fdp->fd_ofiles[SCARG(uap, fd)]) == NULL ||
     92 	    (fp->f_flag & FREAD) == 0)
     93 		return (EBADF);
     94 	aiov.iov_base = (caddr_t)SCARG(uap, buf);
     95 	aiov.iov_len = SCARG(uap, nbyte);
     96 	auio.uio_iov = &aiov;
     97 	auio.uio_iovcnt = 1;
     98 	auio.uio_resid = SCARG(uap, nbyte);
     99 	auio.uio_rw = UIO_READ;
    100 	auio.uio_segflg = UIO_USERSPACE;
    101 	auio.uio_procp = p;
    102 	if (auio.uio_resid < 0)
    103 		return EINVAL;
    104 #ifdef KTRACE
    105 	/*
    106 	 * if tracing, save a copy of iovec
    107 	 */
    108 	if (KTRPOINT(p, KTR_GENIO))
    109 		ktriov = aiov;
    110 #endif
    111 	cnt = SCARG(uap, nbyte);
    112 	error = (*fp->f_ops->fo_read)(fp, &auio, fp->f_cred);
    113 	if (error)
    114 		if (auio.uio_resid != cnt && (error == ERESTART ||
    115 		    error == EINTR || error == EWOULDBLOCK))
    116 			error = 0;
    117 	cnt -= auio.uio_resid;
    118 #ifdef KTRACE
    119 	if (KTRPOINT(p, KTR_GENIO) && error == 0)
    120 		ktrgenio(p->p_tracep, SCARG(uap, fd), UIO_READ, &ktriov,
    121 		    cnt, error);
    122 #endif
    123 	*retval = cnt;
    124 	return (error);
    125 }
    126 
    127 /*
    128  * Scatter read system call.
    129  */
    130 int
    131 sys_readv(p, v, retval)
    132 	struct proc *p;
    133 	void *v;
    134 	register_t *retval;
    135 {
    136 	register struct sys_readv_args /* {
    137 		syscallarg(int) fd;
    138 		syscallarg(const struct iovec *) iovp;
    139 		syscallarg(u_int) iovcnt;
    140 	} */ *uap = v;
    141 	register struct file *fp;
    142 	register struct filedesc *fdp = p->p_fd;
    143 	struct uio auio;
    144 	register struct iovec *iov;
    145 	struct iovec *needfree;
    146 	struct iovec aiov[UIO_SMALLIOV];
    147 	long i, cnt, error = 0;
    148 	u_int iovlen;
    149 #ifdef KTRACE
    150 	struct iovec *ktriov = NULL;
    151 #endif
    152 
    153 	if (((u_int)SCARG(uap, fd)) >= fdp->fd_nfiles ||
    154 	    (fp = fdp->fd_ofiles[SCARG(uap, fd)]) == NULL ||
    155 	    (fp->f_flag & FREAD) == 0)
    156 		return (EBADF);
    157 	/* note: can't use iovlen until iovcnt is validated */
    158 	iovlen = SCARG(uap, iovcnt) * sizeof (struct iovec);
    159 	if (SCARG(uap, iovcnt) > UIO_SMALLIOV) {
    160 		if (SCARG(uap, iovcnt) > UIO_MAXIOV)
    161 			return (EINVAL);
    162 		MALLOC(iov, struct iovec *, iovlen, M_IOV, M_WAITOK);
    163 		needfree = iov;
    164 	} else {
    165 		iov = aiov;
    166 		needfree = NULL;
    167 	}
    168 	auio.uio_iov = iov;
    169 	auio.uio_iovcnt = SCARG(uap, iovcnt);
    170 	auio.uio_rw = UIO_READ;
    171 	auio.uio_segflg = UIO_USERSPACE;
    172 	auio.uio_procp = p;
    173 	error = copyin(SCARG(uap, iovp), iov, iovlen);
    174 	if (error)
    175 		goto done;
    176 	auio.uio_resid = 0;
    177 	for (i = 0; i < SCARG(uap, iovcnt); i++) {
    178 #if 0
    179 		/* Cannot happen iov_len is unsigned */
    180 		if (iov->iov_len < 0) {
    181 			error = EINVAL;
    182 			goto done;
    183 		}
    184 #endif
    185 		auio.uio_resid += iov->iov_len;
    186 		if (auio.uio_resid < 0) {
    187 			error = EINVAL;
    188 			goto done;
    189 		}
    190 		iov++;
    191 	}
    192 #ifdef KTRACE
    193 	/*
    194 	 * if tracing, save a copy of iovec
    195 	 */
    196 	if (KTRPOINT(p, KTR_GENIO))  {
    197 		MALLOC(ktriov, struct iovec *, iovlen, M_TEMP, M_WAITOK);
    198 		bcopy((caddr_t)auio.uio_iov, (caddr_t)ktriov, iovlen);
    199 	}
    200 #endif
    201 	cnt = auio.uio_resid;
    202 	error = (*fp->f_ops->fo_read)(fp, &auio, fp->f_cred);
    203 	if (error)
    204 		if (auio.uio_resid != cnt && (error == ERESTART ||
    205 		    error == EINTR || error == EWOULDBLOCK))
    206 			error = 0;
    207 	cnt -= auio.uio_resid;
    208 #ifdef KTRACE
    209 	if (ktriov != NULL) {
    210 		if (error == 0)
    211 			ktrgenio(p->p_tracep, SCARG(uap, fd), UIO_READ, ktriov,
    212 			    cnt, error);
    213 		FREE(ktriov, M_TEMP);
    214 	}
    215 #endif
    216 	*retval = cnt;
    217 done:
    218 	if (needfree)
    219 		FREE(needfree, M_IOV);
    220 	return (error);
    221 }
    222 
    223 /*
    224  * Write system call
    225  */
    226 int
    227 sys_write(p, v, retval)
    228 	struct proc *p;
    229 	void *v;
    230 	register_t *retval;
    231 {
    232 	register struct sys_write_args /* {
    233 		syscallarg(int) fd;
    234 		syscallarg(const void *) buf;
    235 		syscallarg(size_t) nbyte;
    236 	} */ *uap = v;
    237 	register struct file *fp;
    238 	register struct filedesc *fdp = p->p_fd;
    239 	struct uio auio;
    240 	struct iovec aiov;
    241 	long cnt, error = 0;
    242 #ifdef KTRACE
    243 	struct iovec ktriov;
    244 #endif
    245 
    246 	if (((u_int)SCARG(uap, fd)) >= fdp->fd_nfiles ||
    247 	    (fp = fdp->fd_ofiles[SCARG(uap, fd)]) == NULL ||
    248 	    (fp->f_flag & FWRITE) == 0)
    249 		return (EBADF);
    250 	aiov.iov_base = (char *)SCARG(uap, buf);	/* XXX kills const */
    251 	aiov.iov_len = SCARG(uap, nbyte);
    252 	auio.uio_iov = &aiov;
    253 	auio.uio_iovcnt = 1;
    254 	auio.uio_resid = SCARG(uap, nbyte);
    255 	auio.uio_rw = UIO_WRITE;
    256 	auio.uio_segflg = UIO_USERSPACE;
    257 	auio.uio_procp = p;
    258 	if (auio.uio_resid < 0)
    259 		return EINVAL;
    260 #ifdef KTRACE
    261 	/*
    262 	 * if tracing, save a copy of iovec
    263 	 */
    264 	if (KTRPOINT(p, KTR_GENIO))
    265 		ktriov = aiov;
    266 #endif
    267 	cnt = SCARG(uap, nbyte);
    268 	error = (*fp->f_ops->fo_write)(fp, &auio, fp->f_cred);
    269 	if (error) {
    270 		if (auio.uio_resid != cnt && (error == ERESTART ||
    271 		    error == EINTR || error == EWOULDBLOCK))
    272 			error = 0;
    273 		if (error == EPIPE)
    274 			psignal(p, SIGPIPE);
    275 	}
    276 	cnt -= auio.uio_resid;
    277 #ifdef KTRACE
    278 	if (KTRPOINT(p, KTR_GENIO) && error == 0)
    279 		ktrgenio(p->p_tracep, SCARG(uap, fd), UIO_WRITE,
    280 		    &ktriov, cnt, error);
    281 #endif
    282 	*retval = cnt;
    283 	return (error);
    284 }
    285 
    286 /*
    287  * Gather write system call
    288  */
    289 int
    290 sys_writev(p, v, retval)
    291 	struct proc *p;
    292 	void *v;
    293 	register_t *retval;
    294 {
    295 	register struct sys_writev_args /* {
    296 		syscallarg(int) fd;
    297 		syscallarg(const struct iovec *) iovp;
    298 		syscallarg(u_int) iovcnt;
    299 	} */ *uap = v;
    300 	register struct file *fp;
    301 	register struct filedesc *fdp = p->p_fd;
    302 	struct uio auio;
    303 	register struct iovec *iov;
    304 	struct iovec *needfree;
    305 	struct iovec aiov[UIO_SMALLIOV];
    306 	long i, cnt, error = 0;
    307 	u_int iovlen;
    308 #ifdef KTRACE
    309 	struct iovec *ktriov = NULL;
    310 #endif
    311 
    312 	if (((u_int)SCARG(uap, fd)) >= fdp->fd_nfiles ||
    313 	    (fp = fdp->fd_ofiles[SCARG(uap, fd)]) == NULL ||
    314 	    (fp->f_flag & FWRITE) == 0)
    315 		return (EBADF);
    316 	/* note: can't use iovlen until iovcnt is validated */
    317 	iovlen = SCARG(uap, iovcnt) * sizeof (struct iovec);
    318 	if (SCARG(uap, iovcnt) > UIO_SMALLIOV) {
    319 		if (SCARG(uap, iovcnt) > UIO_MAXIOV)
    320 			return (EINVAL);
    321 		MALLOC(iov, struct iovec *, iovlen, M_IOV, M_WAITOK);
    322 		needfree = iov;
    323 	} else {
    324 		iov = aiov;
    325 		needfree = NULL;
    326 	}
    327 	auio.uio_iov = iov;
    328 	auio.uio_iovcnt = SCARG(uap, iovcnt);
    329 	auio.uio_rw = UIO_WRITE;
    330 	auio.uio_segflg = UIO_USERSPACE;
    331 	auio.uio_procp = p;
    332 	error = copyin(SCARG(uap, iovp), iov, iovlen);
    333 	if (error)
    334 		goto done;
    335 	auio.uio_resid = 0;
    336 	for (i = 0; i < SCARG(uap, iovcnt); i++) {
    337 #if 0
    338 		/* Cannot happen iov_len is unsigned */
    339 		if (iov->iov_len < 0) {
    340 			error = EINVAL;
    341 			goto done;
    342 		}
    343 #endif
    344 		auio.uio_resid += iov->iov_len;
    345 		if (auio.uio_resid < 0) {
    346 			error = EINVAL;
    347 			goto done;
    348 		}
    349 		iov++;
    350 	}
    351 #ifdef KTRACE
    352 	/*
    353 	 * if tracing, save a copy of iovec
    354 	 */
    355 	if (KTRPOINT(p, KTR_GENIO))  {
    356 		MALLOC(ktriov, struct iovec *, iovlen, M_TEMP, M_WAITOK);
    357 		bcopy((caddr_t)auio.uio_iov, (caddr_t)ktriov, iovlen);
    358 	}
    359 #endif
    360 	cnt = auio.uio_resid;
    361 	error = (*fp->f_ops->fo_write)(fp, &auio, fp->f_cred);
    362 	if (error) {
    363 		if (auio.uio_resid != cnt && (error == ERESTART ||
    364 		    error == EINTR || error == EWOULDBLOCK))
    365 			error = 0;
    366 		if (error == EPIPE)
    367 			psignal(p, SIGPIPE);
    368 	}
    369 	cnt -= auio.uio_resid;
    370 #ifdef KTRACE
    371 	if (ktriov != NULL) {
    372 		if (error == 0)
    373 			ktrgenio(p->p_tracep, SCARG(uap, fd), UIO_WRITE,
    374 				ktriov, cnt, error);
    375 		FREE(ktriov, M_TEMP);
    376 	}
    377 #endif
    378 	*retval = cnt;
    379 done:
    380 	if (needfree)
    381 		FREE(needfree, M_IOV);
    382 	return (error);
    383 }
    384 
    385 /*
    386  * Ioctl system call
    387  */
    388 /* ARGSUSED */
    389 int
    390 sys_ioctl(p, v, retval)
    391 	struct proc *p;
    392 	void *v;
    393 	register_t *retval;
    394 {
    395 	register struct sys_ioctl_args /* {
    396 		syscallarg(int) fd;
    397 		syscallarg(u_long) com;
    398 		syscallarg(caddr_t) data;
    399 	} */ *uap = v;
    400 	register struct file *fp;
    401 	register struct filedesc *fdp;
    402 	register u_long com;
    403 	register int error;
    404 	register u_int size;
    405 	caddr_t data, memp;
    406 	int tmp;
    407 #define STK_PARAMS	128
    408 	char stkbuf[STK_PARAMS];
    409 
    410 	fdp = p->p_fd;
    411 	if ((u_int)SCARG(uap, fd) >= fdp->fd_nfiles ||
    412 	    (fp = fdp->fd_ofiles[SCARG(uap, fd)]) == NULL)
    413 		return (EBADF);
    414 
    415 	if ((fp->f_flag & (FREAD | FWRITE)) == 0)
    416 		return (EBADF);
    417 
    418 	switch (com = SCARG(uap, com)) {
    419 	case FIONCLEX:
    420 		fdp->fd_ofileflags[SCARG(uap, fd)] &= ~UF_EXCLOSE;
    421 		return (0);
    422 	case FIOCLEX:
    423 		fdp->fd_ofileflags[SCARG(uap, fd)] |= UF_EXCLOSE;
    424 		return (0);
    425 	}
    426 
    427 	/*
    428 	 * Interpret high order word to find amount of data to be
    429 	 * copied to/from the user's address space.
    430 	 */
    431 	size = IOCPARM_LEN(com);
    432 	if (size > IOCPARM_MAX)
    433 		return (ENOTTY);
    434 	memp = NULL;
    435 	if (size > sizeof (stkbuf)) {
    436 		memp = (caddr_t)malloc((u_long)size, M_IOCTLOPS, M_WAITOK);
    437 		data = memp;
    438 	} else
    439 		data = stkbuf;
    440 	if (com&IOC_IN) {
    441 		if (size) {
    442 			error = copyin(SCARG(uap, data), data, size);
    443 			if (error) {
    444 				if (memp)
    445 					free(memp, M_IOCTLOPS);
    446 				return (error);
    447 			}
    448 		} else
    449 			*(caddr_t *)data = SCARG(uap, data);
    450 	} else if ((com&IOC_OUT) && size)
    451 		/*
    452 		 * Zero the buffer so the user always
    453 		 * gets back something deterministic.
    454 		 */
    455 		bzero(data, size);
    456 	else if (com&IOC_VOID)
    457 		*(caddr_t *)data = SCARG(uap, data);
    458 
    459 	switch (com) {
    460 
    461 	case FIONBIO:
    462 		if ((tmp = *(int *)data) != 0)
    463 			fp->f_flag |= FNONBLOCK;
    464 		else
    465 			fp->f_flag &= ~FNONBLOCK;
    466 		error = (*fp->f_ops->fo_ioctl)(fp, FIONBIO, (caddr_t)&tmp, p);
    467 		break;
    468 
    469 	case FIOASYNC:
    470 		if ((tmp = *(int *)data) != 0)
    471 			fp->f_flag |= FASYNC;
    472 		else
    473 			fp->f_flag &= ~FASYNC;
    474 		error = (*fp->f_ops->fo_ioctl)(fp, FIOASYNC, (caddr_t)&tmp, p);
    475 		break;
    476 
    477 	case FIOSETOWN:
    478 		tmp = *(int *)data;
    479 		if (fp->f_type == DTYPE_SOCKET) {
    480 			((struct socket *)fp->f_data)->so_pgid = tmp;
    481 			error = 0;
    482 			break;
    483 		}
    484 		if (tmp <= 0) {
    485 			tmp = -tmp;
    486 		} else {
    487 			struct proc *p1 = pfind(tmp);
    488 			if (p1 == 0) {
    489 				error = ESRCH;
    490 				break;
    491 			}
    492 			tmp = p1->p_pgrp->pg_id;
    493 		}
    494 		error = (*fp->f_ops->fo_ioctl)
    495 			(fp, TIOCSPGRP, (caddr_t)&tmp, p);
    496 		break;
    497 
    498 	case FIOGETOWN:
    499 		if (fp->f_type == DTYPE_SOCKET) {
    500 			error = 0;
    501 			*(int *)data = ((struct socket *)fp->f_data)->so_pgid;
    502 			break;
    503 		}
    504 		error = (*fp->f_ops->fo_ioctl)(fp, TIOCGPGRP, data, p);
    505 		*(int *)data = -*(int *)data;
    506 		break;
    507 
    508 	default:
    509 		error = (*fp->f_ops->fo_ioctl)(fp, com, data, p);
    510 		/*
    511 		 * Copy any data to user, size was
    512 		 * already set and checked above.
    513 		 */
    514 		if (error == 0 && (com&IOC_OUT) && size)
    515 			error = copyout(data, SCARG(uap, data), size);
    516 		break;
    517 	}
    518 	if (memp)
    519 		free(memp, M_IOCTLOPS);
    520 	return (error);
    521 }
    522 
    523 int	selwait, nselcoll;
    524 
    525 /*
    526  * Select system call.
    527  */
    528 int
    529 sys_select(p, v, retval)
    530 	register struct proc *p;
    531 	void *v;
    532 	register_t *retval;
    533 {
    534 	register struct sys_select_args /* {
    535 		syscallarg(u_int) nd;
    536 		syscallarg(fd_set *) in;
    537 		syscallarg(fd_set *) ou;
    538 		syscallarg(fd_set *) ex;
    539 		syscallarg(struct timeval *) tv;
    540 	} */ *uap = v;
    541 	caddr_t bits;
    542 	char smallbits[howmany(FD_SETSIZE, NFDBITS) * sizeof(fd_mask) * 6];
    543 	struct timeval atv;
    544 	int s, ncoll, error = 0, timo;
    545 	size_t ni;
    546 
    547 	if (SCARG(uap, nd) > p->p_fd->fd_nfiles) {
    548 		/* forgiving; slightly wrong */
    549 		SCARG(uap, nd) = p->p_fd->fd_nfiles;
    550 	}
    551 	ni = howmany(SCARG(uap, nd), NFDBITS) * sizeof(fd_mask);
    552 	if (ni * 6 > sizeof(smallbits))
    553 		bits = malloc(ni * 6, M_TEMP, M_WAITOK);
    554 	else
    555 		bits = smallbits;
    556 
    557 #define	getbits(name, x) \
    558 	if (SCARG(uap, name)) { \
    559 		error = copyin(SCARG(uap, name), bits + ni * x, ni); \
    560 		if (error) \
    561 			goto done; \
    562 	} else \
    563 		bzero(bits + ni * x, ni);
    564 	getbits(in, 0);
    565 	getbits(ou, 1);
    566 	getbits(ex, 2);
    567 #undef	getbits
    568 
    569 	if (SCARG(uap, tv)) {
    570 		error = copyin(SCARG(uap, tv), (caddr_t)&atv,
    571 			sizeof (atv));
    572 		if (error)
    573 			goto done;
    574 		if (itimerfix(&atv)) {
    575 			error = EINVAL;
    576 			goto done;
    577 		}
    578 		s = splclock();
    579 		timeradd(&atv, &time, &atv);
    580 		timo = hzto(&atv);
    581 		/*
    582 		 * Avoid inadvertently sleeping forever.
    583 		 */
    584 		if (timo == 0)
    585 			timo = 1;
    586 		splx(s);
    587 	} else
    588 		timo = 0;
    589 retry:
    590 	ncoll = nselcoll;
    591 	p->p_flag |= P_SELECT;
    592 	error = selscan(p, (fd_mask *)(bits + ni * 0),
    593 			   (fd_mask *)(bits + ni * 3), SCARG(uap, nd), retval);
    594 	if (error || *retval)
    595 		goto done;
    596 	s = splhigh();
    597 	if (timo && timercmp(&time, &atv, >=)) {
    598 		splx(s);
    599 		goto done;
    600 	}
    601 	if ((p->p_flag & P_SELECT) == 0 || nselcoll != ncoll) {
    602 		splx(s);
    603 		goto retry;
    604 	}
    605 	p->p_flag &= ~P_SELECT;
    606 	error = tsleep((caddr_t)&selwait, PSOCK | PCATCH, "select", timo);
    607 	splx(s);
    608 	if (error == 0)
    609 		goto retry;
    610 done:
    611 	p->p_flag &= ~P_SELECT;
    612 	/* select is not restarted after signals... */
    613 	if (error == ERESTART)
    614 		error = EINTR;
    615 	if (error == EWOULDBLOCK)
    616 		error = 0;
    617 	if (error == 0) {
    618 #define	putbits(name, x) \
    619 		if (SCARG(uap, name)) { \
    620 			error = copyout(bits + ni * x, SCARG(uap, name), ni); \
    621 			if (error) \
    622 				goto out; \
    623 		}
    624 		putbits(in, 3);
    625 		putbits(ou, 4);
    626 		putbits(ex, 5);
    627 #undef putbits
    628 	}
    629 out:
    630 	if (ni * 6 > sizeof(smallbits))
    631 		free(bits, M_TEMP);
    632 	return (error);
    633 }
    634 
    635 int
    636 selscan(p, ibitp, obitp, nfd, retval)
    637 	struct proc *p;
    638 	fd_mask *ibitp, *obitp;
    639 	int nfd;
    640 	register_t *retval;
    641 {
    642 	register struct filedesc *fdp = p->p_fd;
    643 	register int msk, i, j, fd;
    644 	register fd_mask ibits, obits;
    645 	struct file *fp;
    646 	int n = 0;
    647 	static int flag[3] = { POLLRDNORM | POLLHUP | POLLERR,
    648 			       POLLWRNORM | POLLHUP | POLLERR,
    649 			       POLLRDBAND };
    650 
    651 	for (msk = 0; msk < 3; msk++) {
    652 		for (i = 0; i < nfd; i += NFDBITS) {
    653 			ibits = *ibitp++;
    654 			obits = 0;
    655 			while ((j = ffs(ibits)) && (fd = i + --j) < nfd) {
    656 				ibits &= ~(1 << j);
    657 				fp = fdp->fd_ofiles[fd];
    658 				if (fp == NULL)
    659 					return (EBADF);
    660 				if ((*fp->f_ops->fo_poll)(fp, flag[msk], p)) {
    661 					obits |= (1 << j);
    662 					n++;
    663 				}
    664 			}
    665 			*obitp++ = obits;
    666 		}
    667 	}
    668 	*retval = n;
    669 	return (0);
    670 }
    671 
    672 /*
    673  * Poll system call.
    674  */
    675 int
    676 sys_poll(p, v, retval)
    677 	register struct proc *p;
    678 	void *v;
    679 	register_t *retval;
    680 {
    681 	register struct sys_poll_args /* {
    682 		syscallarg(struct pollfd *) fds;
    683 		syscallarg(u_int) nfds;
    684 		syscallarg(int) timeout;
    685 	} */ *uap = v;
    686 	caddr_t bits;
    687 	char smallbits[32 * sizeof(struct pollfd)];
    688 	struct timeval atv;
    689 	int s, ncoll, error = 0, timo;
    690 	size_t ni;
    691 
    692 	if (SCARG(uap, nfds) > p->p_fd->fd_nfiles) {
    693 		/* forgiving; slightly wrong */
    694 		SCARG(uap, nfds) = p->p_fd->fd_nfiles;
    695 	}
    696 	ni = SCARG(uap, nfds) * sizeof(struct pollfd);
    697 	if (ni > sizeof(smallbits))
    698 		bits = malloc(ni, M_TEMP, M_WAITOK);
    699 	else
    700 		bits = smallbits;
    701 
    702 	error = copyin(SCARG(uap, fds), bits, ni);
    703 	if (error)
    704 		goto done;
    705 
    706 	if (SCARG(uap, timeout) != INFTIM) {
    707 		atv.tv_sec = SCARG(uap, timeout) / 1000;
    708 		atv.tv_usec = (SCARG(uap, timeout) % 1000) * 1000;
    709 		if (itimerfix(&atv)) {
    710 			error = EINVAL;
    711 			goto done;
    712 		}
    713 		s = splclock();
    714 		timeradd(&atv, &time, &atv);
    715 		timo = hzto(&atv);
    716 		/*
    717 		 * Avoid inadvertently sleeping forever.
    718 		 */
    719 		if (timo == 0)
    720 			timo = 1;
    721 		splx(s);
    722 	} else
    723 		timo = 0;
    724 retry:
    725 	ncoll = nselcoll;
    726 	p->p_flag |= P_SELECT;
    727 	error = pollscan(p, (struct pollfd *)bits, SCARG(uap, nfds), retval);
    728 	if (error || *retval)
    729 		goto done;
    730 	s = splhigh();
    731 	if (timo && timercmp(&time, &atv, >=)) {
    732 		splx(s);
    733 		goto done;
    734 	}
    735 	if ((p->p_flag & P_SELECT) == 0 || nselcoll != ncoll) {
    736 		splx(s);
    737 		goto retry;
    738 	}
    739 	p->p_flag &= ~P_SELECT;
    740 	error = tsleep((caddr_t)&selwait, PSOCK | PCATCH, "select", timo);
    741 	splx(s);
    742 	if (error == 0)
    743 		goto retry;
    744 done:
    745 	p->p_flag &= ~P_SELECT;
    746 	/* poll is not restarted after signals... */
    747 	if (error == ERESTART)
    748 		error = EINTR;
    749 	if (error == EWOULDBLOCK)
    750 		error = 0;
    751 	if (error == 0) {
    752 		error = copyout(bits, SCARG(uap, fds), ni);
    753 		if (error)
    754 			goto out;
    755 	}
    756 out:
    757 	if (ni > sizeof(smallbits))
    758 		free(bits, M_TEMP);
    759 	return (error);
    760 }
    761 
    762 int
    763 pollscan(p, fds, nfd, retval)
    764 	struct proc *p;
    765 	struct pollfd *fds;
    766 	int nfd;
    767 	register_t *retval;
    768 {
    769 	register struct filedesc *fdp = p->p_fd;
    770 	int i;
    771 	struct file *fp;
    772 	int n = 0;
    773 
    774 	for (i = 0; i < nfd; i++, fds++) {
    775 		fp = fdp->fd_ofiles[fds->fd];
    776 		if (fp == 0) {
    777 			fds->revents = POLLNVAL;
    778 			n++;
    779 		} else {
    780 			fds->revents = (*fp->f_ops->fo_poll)(fp,
    781 			    fds->events | POLLERR | POLLHUP, p);
    782 			if (fds->revents != 0)
    783 				n++;
    784 		}
    785 	}
    786 	*retval = n;
    787 	return (0);
    788 }
    789 
    790 /*ARGSUSED*/
    791 int
    792 seltrue(dev, events, p)
    793 	dev_t dev;
    794 	int events;
    795 	struct proc *p;
    796 {
    797 
    798 	return (events & (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM));
    799 }
    800 
    801 /*
    802  * Record a select request.
    803  */
    804 void
    805 selrecord(selector, sip)
    806 	struct proc *selector;
    807 	struct selinfo *sip;
    808 {
    809 	struct proc *p;
    810 	pid_t mypid;
    811 
    812 	mypid = selector->p_pid;
    813 	if (sip->si_pid == mypid)
    814 		return;
    815 	if (sip->si_pid && (p = pfind(sip->si_pid)) &&
    816 	    p->p_wchan == (caddr_t)&selwait)
    817 		sip->si_flags |= SI_COLL;
    818 	else
    819 		sip->si_pid = mypid;
    820 }
    821 
    822 /*
    823  * Do a wakeup when a selectable event occurs.
    824  */
    825 void
    826 selwakeup(sip)
    827 	register struct selinfo *sip;
    828 {
    829 	register struct proc *p;
    830 	int s;
    831 
    832 	if (sip->si_pid == 0)
    833 		return;
    834 	if (sip->si_flags & SI_COLL) {
    835 		nselcoll++;
    836 		sip->si_flags &= ~SI_COLL;
    837 		wakeup((caddr_t)&selwait);
    838 	}
    839 	p = pfind(sip->si_pid);
    840 	sip->si_pid = 0;
    841 	if (p != NULL) {
    842 		s = splhigh();
    843 		if (p->p_wchan == (caddr_t)&selwait) {
    844 			if (p->p_stat == SSLEEP)
    845 				setrunnable(p);
    846 			else
    847 				unsleep(p);
    848 		} else if (p->p_flag & P_SELECT)
    849 			p->p_flag &= ~P_SELECT;
    850 		splx(s);
    851 	}
    852 }
    853