xref: /illumos-gate/usr/src/lib/libc/port/gen/select.c (revision e5803b76927480e8f9b67b22201c484ccf4c2bcf)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*	Copyright (c) 1988 AT&T	*/
28 /*	  All Rights Reserved  	*/
29 
30 #pragma ident	"%Z%%M%	%I%	%E% SMI"
31 
32 /*
33  * Emulation of select() system call using poll() system call.
34  *
35  * Assumptions:
36  *	polling for input only is most common.
37  *	polling for exceptional conditions is very rare.
38  *
39  * Note that is it not feasible to emulate all error conditions,
40  * in particular conditions that would return EFAULT are far too
41  * difficult to check for in a library routine.
42  */
43 
44 #pragma weak _select = select
45 
46 #include "lint.h"
47 #include <values.h>
48 #include <pthread.h>
49 #include <errno.h>
50 #include <sys/time.h>
51 #include <sys/types.h>
52 #include <sys/select.h>
53 #include <sys/poll.h>
54 #include <alloca.h>
55 #include "libc.h"
56 
57 int
58 pselect(int nfds, fd_set *in0, fd_set *out0, fd_set *ex0,
59 	const timespec_t *tsp, const sigset_t *sigmask)
60 {
61 	long *in, *out, *ex;
62 	ulong_t m;	/* bit mask */
63 	int j;		/* loop counter */
64 	ulong_t b;	/* bits to test */
65 	int n, rv;
66 	struct pollfd *pfd;
67 	struct pollfd *p;
68 	int lastj = -1;
69 
70 	/* "zero" is read-only, it could go in the text segment */
71 	static fd_set zero = { 0 };
72 
73 	/*
74 	 * Check for invalid conditions at outset.
75 	 * Required for spec1170.
76 	 * SUSV3: We must behave as a cancellation point even if we fail early.
77 	 */
78 	if (nfds < 0 || nfds > FD_SETSIZE) {
79 		pthread_testcancel();
80 		errno = EINVAL;
81 		return (-1);
82 	}
83 	p = pfd = (struct pollfd *)alloca(nfds * sizeof (struct pollfd));
84 
85 	if (tsp != NULL) {
86 		/* check timespec validity */
87 		if (tsp->tv_nsec < 0 || tsp->tv_nsec >= NANOSEC ||
88 		    tsp->tv_sec < 0) {
89 			pthread_testcancel();
90 			errno = EINVAL;
91 			return (-1);
92 		}
93 	}
94 
95 	/*
96 	 * If any input args are null, point them at the null array.
97 	 */
98 	if (in0 == NULL)
99 		in0 = &zero;
100 	if (out0 == NULL)
101 		out0 = &zero;
102 	if (ex0 == NULL)
103 		ex0 = &zero;
104 
105 	/*
106 	 * For each fd, if any bits are set convert them into
107 	 * the appropriate pollfd struct.
108 	 */
109 	in = (long *)in0->fds_bits;
110 	out = (long *)out0->fds_bits;
111 	ex = (long *)ex0->fds_bits;
112 	for (n = 0; n < nfds; n += NFDBITS) {
113 		b = (ulong_t)(*in | *out | *ex);
114 		for (j = 0, m = 1; b != 0; j++, b >>= 1, m <<= 1) {
115 			if (b & 1) {
116 				p->fd = n + j;
117 				if (p->fd >= nfds)
118 					goto done;
119 				p->events = 0;
120 				if (*in & m)
121 					p->events |= POLLRDNORM;
122 				if (*out & m)
123 					p->events |= POLLWRNORM;
124 				if (*ex & m)
125 					p->events |= POLLRDBAND;
126 				p++;
127 			}
128 		}
129 		in++;
130 		out++;
131 		ex++;
132 	}
133 done:
134 	/*
135 	 * Now do the poll.
136 	 */
137 	n = (int)(p - pfd);		/* number of pollfd's */
138 	do {
139 		rv = _pollsys(pfd, (nfds_t)n, tsp, sigmask);
140 	} while (rv < 0 && errno == EAGAIN);
141 
142 	if (rv < 0)		/* no need to set bit masks */
143 		return (rv);
144 
145 	if (rv == 0) {
146 		/*
147 		 * Clear out bit masks, just in case.
148 		 * On the assumption that usually only
149 		 * one bit mask is set, use three loops.
150 		 */
151 		if (in0 != &zero) {
152 			in = (long *)in0->fds_bits;
153 			for (n = 0; n < nfds; n += NFDBITS)
154 				*in++ = 0;
155 		}
156 		if (out0 != &zero) {
157 			out = (long *)out0->fds_bits;
158 			for (n = 0; n < nfds; n += NFDBITS)
159 				*out++ = 0;
160 		}
161 		if (ex0 != &zero) {
162 			ex = (long *)ex0->fds_bits;
163 			for (n = 0; n < nfds; n += NFDBITS)
164 				*ex++ = 0;
165 		}
166 		return (0);
167 	}
168 
169 	/*
170 	 * Check for EINVAL error case first to avoid changing any bits
171 	 * if we're going to return an error.
172 	 */
173 	for (p = pfd, j = n; j-- > 0; p++) {
174 		/*
175 		 * select will return EBADF immediately if any fd's
176 		 * are bad.  poll will complete the poll on the
177 		 * rest of the fd's and include the error indication
178 		 * in the returned bits.  This is a rare case so we
179 		 * accept this difference and return the error after
180 		 * doing more work than select would've done.
181 		 */
182 		if (p->revents & POLLNVAL) {
183 			errno = EBADF;
184 			return (-1);
185 		}
186 		/*
187 		 * We would like to make POLLHUP available to select,
188 		 * checking to see if we have pending data to be read.
189 		 * BUT until we figure out how not to break Xsun's
190 		 * dependencies on select's existing features...
191 		 * This is what we _thought_ would work ... sigh!
192 		 */
193 		/*
194 		 * if ((p->revents & POLLHUP) &&
195 		 *	!(p->revents & (POLLRDNORM|POLLRDBAND))) {
196 		 *	errno = EINTR;
197 		 *	return (-1);
198 		 * }
199 		 */
200 	}
201 
202 	/*
203 	 * Convert results of poll back into bits
204 	 * in the argument arrays.
205 	 *
206 	 * We assume POLLRDNORM, POLLWRNORM, and POLLRDBAND will only be set
207 	 * on return from poll if they were set on input, thus we don't
208 	 * worry about accidentally setting the corresponding bits in the
209 	 * zero array if the input bit masks were null.
210 	 *
211 	 * Must return number of bits set, not number of ready descriptors
212 	 * (as the man page says, and as poll() does).
213 	 */
214 	rv = 0;
215 	for (p = pfd; n-- > 0; p++) {
216 		j = (int)(p->fd / NFDBITS);
217 		/* have we moved into another word of the bit mask yet? */
218 		if (j != lastj) {
219 			/* clear all output bits to start with */
220 			in = (long *)&in0->fds_bits[j];
221 			out = (long *)&out0->fds_bits[j];
222 			ex = (long *)&ex0->fds_bits[j];
223 			/*
224 			 * In case we made "zero" read-only (e.g., with
225 			 * cc -R), avoid actually storing into it.
226 			 */
227 			if (in0 != &zero)
228 				*in = 0;
229 			if (out0 != &zero)
230 				*out = 0;
231 			if (ex0 != &zero)
232 				*ex = 0;
233 			lastj = j;
234 		}
235 		if (p->revents) {
236 			m = 1L << (p->fd % NFDBITS);
237 			if (p->revents & POLLRDNORM) {
238 				*in |= m;
239 				rv++;
240 			}
241 			if (p->revents & POLLWRNORM) {
242 				*out |= m;
243 				rv++;
244 			}
245 			if (p->revents & POLLRDBAND) {
246 				*ex |= m;
247 				rv++;
248 			}
249 			/*
250 			 * Only set this bit on return if we asked about
251 			 * input conditions.
252 			 */
253 			if ((p->revents & (POLLHUP|POLLERR)) &&
254 			    (p->events & POLLRDNORM)) {
255 				if ((*in & m) == 0)
256 					rv++;	/* wasn't already set */
257 				*in |= m;
258 			}
259 			/*
260 			 * Only set this bit on return if we asked about
261 			 * output conditions.
262 			 */
263 			if ((p->revents & (POLLHUP|POLLERR)) &&
264 			    (p->events & POLLWRNORM)) {
265 				if ((*out & m) == 0)
266 					rv++;	/* wasn't already set */
267 				*out |= m;
268 			}
269 			/*
270 			 * Only set this bit on return if we asked about
271 			 * output conditions.
272 			 */
273 			if ((p->revents & (POLLHUP|POLLERR)) &&
274 			    (p->events & POLLRDBAND)) {
275 				if ((*ex & m) == 0)
276 					rv++;	/* wasn't already set */
277 				*ex |= m;
278 			}
279 		}
280 	}
281 	return (rv);
282 }
283 
284 int
285 select(int nfds, fd_set *in0, fd_set *out0, fd_set *ex0, struct timeval *tv)
286 {
287 	timespec_t ts;
288 	timespec_t *tsp;
289 
290 	if (tv == NULL)
291 		tsp = NULL;
292 	else {
293 		/* check timeval validity */
294 		if (tv->tv_usec < 0 || tv->tv_usec >= MICROSEC) {
295 			errno = EINVAL;
296 			return (-1);
297 		}
298 		/*
299 		 * Convert timeval to timespec.
300 		 * To preserve compatibility with past behavior,
301 		 * when select was built upon poll(2), which has a
302 		 * minimum non-zero timeout of 1 millisecond, force
303 		 * a minimum non-zero timeout of 500 microseconds.
304 		 */
305 		ts.tv_sec = tv->tv_sec;
306 		ts.tv_nsec = tv->tv_usec * 1000;
307 		if (ts.tv_nsec != 0 && ts.tv_nsec < 500000)
308 			ts.tv_nsec = 500000;
309 		tsp = &ts;
310 	}
311 
312 	return (pselect(nfds, in0, out0, ex0, tsp, NULL));
313 }
314