xref: /netbsd-src/sys/kern/sys_select.c (revision 09afef20633f5fe63d92dfe43ee3a9380dc06883)
1 /*	$NetBSD: sys_select.c,v 1.19 2009/11/11 09:48:51 rmind Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007, 2008, 2009 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Copyright (c) 1982, 1986, 1989, 1993
34  *	The Regents of the University of California.  All rights reserved.
35  * (c) UNIX System Laboratories, Inc.
36  * All or some portions of this file are derived from material licensed
37  * to the University of California by American Telephone and Telegraph
38  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
39  * the permission of UNIX System Laboratories, Inc.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. Neither the name of the University nor the names of its contributors
50  *    may be used to endorse or promote products derived from this software
51  *    without specific prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
54  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
55  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
56  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
57  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
58  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
59  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  *
65  *	@(#)sys_generic.c	8.9 (Berkeley) 2/14/95
66  */
67 
68 /*
69  * System calls relating to files.
70  */
71 
72 #include <sys/cdefs.h>
73 __KERNEL_RCSID(0, "$NetBSD: sys_select.c,v 1.19 2009/11/11 09:48:51 rmind Exp $");
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/filedesc.h>
78 #include <sys/ioctl.h>
79 #include <sys/file.h>
80 #include <sys/proc.h>
81 #include <sys/socketvar.h>
82 #include <sys/signalvar.h>
83 #include <sys/uio.h>
84 #include <sys/kernel.h>
85 #include <sys/stat.h>
86 #include <sys/poll.h>
87 #include <sys/vnode.h>
88 #include <sys/mount.h>
89 #include <sys/syscallargs.h>
90 #include <sys/cpu.h>
91 #include <sys/atomic.h>
92 #include <sys/socketvar.h>
93 #include <sys/sleepq.h>
94 
95 /* Flags for lwp::l_selflag. */
96 #define	SEL_RESET	0	/* awoken, interrupted, or not yet polling */
97 #define	SEL_SCANNING	1	/* polling descriptors */
98 #define	SEL_BLOCKING	2	/* about to block on select_cv */
99 
100 /* Per-CPU state for select()/poll(). */
101 #if MAXCPUS > 32
102 #error adjust this code
103 #endif
104 typedef struct selcpu {
105 	kmutex_t	*sc_lock;
106 	sleepq_t	sc_sleepq;
107 	int		sc_ncoll;
108 	uint32_t	sc_mask;
109 } selcpu_t;
110 
111 static inline int	selscan(char *, u_int, register_t *);
112 static inline int	pollscan(struct pollfd *, u_int, register_t *);
113 static void		selclear(void);
114 
115 static syncobj_t select_sobj = {
116 	SOBJ_SLEEPQ_FIFO,
117 	sleepq_unsleep,
118 	sleepq_changepri,
119 	sleepq_lendpri,
120 	syncobj_noowner,
121 };
122 
123 /*
124  * Select system call.
125  */
126 int
127 sys___pselect50(struct lwp *l, const struct sys___pselect50_args *uap,
128     register_t *retval)
129 {
130 	/* {
131 		syscallarg(int)				nd;
132 		syscallarg(fd_set *)			in;
133 		syscallarg(fd_set *)			ou;
134 		syscallarg(fd_set *)			ex;
135 		syscallarg(const struct timespec *)	ts;
136 		syscallarg(sigset_t *)			mask;
137 	} */
138 	struct timespec	ats, *ts = NULL;
139 	sigset_t	amask, *mask = NULL;
140 	int		error;
141 
142 	if (SCARG(uap, ts)) {
143 		error = copyin(SCARG(uap, ts), &ats, sizeof(ats));
144 		if (error)
145 			return error;
146 		ts = &ats;
147 	}
148 	if (SCARG(uap, mask) != NULL) {
149 		error = copyin(SCARG(uap, mask), &amask, sizeof(amask));
150 		if (error)
151 			return error;
152 		mask = &amask;
153 	}
154 
155 	return selcommon(retval, SCARG(uap, nd), SCARG(uap, in),
156 	    SCARG(uap, ou), SCARG(uap, ex), ts, mask);
157 }
158 
159 int
160 sys___select50(struct lwp *l, const struct sys___select50_args *uap,
161     register_t *retval)
162 {
163 	/* {
164 		syscallarg(int)			nd;
165 		syscallarg(fd_set *)		in;
166 		syscallarg(fd_set *)		ou;
167 		syscallarg(fd_set *)		ex;
168 		syscallarg(struct timeval *)	tv;
169 	} */
170 	struct timeval atv;
171 	struct timespec ats, *ts = NULL;
172 	int error;
173 
174 	if (SCARG(uap, tv)) {
175 		error = copyin(SCARG(uap, tv), (void *)&atv, sizeof(atv));
176 		if (error)
177 			return error;
178 		TIMEVAL_TO_TIMESPEC(&atv, &ats);
179 		ts = &ats;
180 	}
181 
182 	return selcommon(retval, SCARG(uap, nd), SCARG(uap, in),
183 	    SCARG(uap, ou), SCARG(uap, ex), ts, NULL);
184 }
185 
186 /*
187  * sel_do_scan: common code to perform the scan on descriptors.
188  */
189 static int
190 sel_do_scan(void *fds, u_int nfds, struct timespec *ts, sigset_t *mask,
191     register_t *retval, int selpoll)
192 {
193 	lwp_t		* const l = curlwp;
194 	proc_t		* const p = l->l_proc;
195 	selcpu_t	*sc;
196 	kmutex_t	*lock;
197 	sigset_t	oldmask;
198 	struct timespec	sleepts;
199 	int		error, timo;
200 
201 	timo = 0;
202 	if (ts && inittimeleft(ts, &sleepts) == -1) {
203 		return EINVAL;
204 	}
205 
206 	if (__predict_false(mask)) {
207 		sigminusset(&sigcantmask, mask);
208 		mutex_enter(p->p_lock);
209 		oldmask = l->l_sigmask;
210 		l->l_sigmask = *mask;
211 		mutex_exit(p->p_lock);
212 	} else {
213 		/* XXXgcc */
214 		oldmask = l->l_sigmask;
215 	}
216 
217 	sc = curcpu()->ci_data.cpu_selcpu;
218 	lock = sc->sc_lock;
219 	l->l_selcpu = sc;
220 	SLIST_INIT(&l->l_selwait);
221 	for (;;) {
222 		int ncoll;
223 
224 		/*
225 		 * No need to lock.  If this is overwritten by another value
226 		 * while scanning, we will retry below.  We only need to see
227 		 * exact state from the descriptors that we are about to poll,
228 		 * and lock activity resulting from fo_poll is enough to
229 		 * provide an up to date value for new polling activity.
230 		 */
231 		l->l_selflag = SEL_SCANNING;
232 		ncoll = sc->sc_ncoll;
233 
234 		if (selpoll) {
235 			error = selscan((char *)fds, nfds, retval);
236 		} else {
237 			error = pollscan((struct pollfd *)fds, nfds, retval);
238 		}
239 
240 		if (error || *retval)
241 			break;
242 		if (ts && (timo = gettimeleft(ts, &sleepts)) <= 0)
243 			break;
244 		mutex_spin_enter(lock);
245 		if (l->l_selflag != SEL_SCANNING || sc->sc_ncoll != ncoll) {
246 			mutex_spin_exit(lock);
247 			continue;
248 		}
249 		l->l_selflag = SEL_BLOCKING;
250 		l->l_kpriority = true;
251 		sleepq_enter(&sc->sc_sleepq, l, lock);
252 		sleepq_enqueue(&sc->sc_sleepq, sc, "select", &select_sobj);
253 		error = sleepq_block(timo, true);
254 		if (error != 0)
255 			break;
256 	}
257 	selclear();
258 
259 	if (__predict_false(mask)) {
260 		mutex_enter(p->p_lock);
261 		l->l_sigmask = oldmask;
262 		mutex_exit(p->p_lock);
263 	}
264 	return error;
265 }
266 
267 int
268 selcommon(register_t *retval, int nd, fd_set *u_in, fd_set *u_ou,
269     fd_set *u_ex, struct timespec *ts, sigset_t *mask)
270 {
271 	char		smallbits[howmany(FD_SETSIZE, NFDBITS) *
272 			    sizeof(fd_mask) * 6];
273 	char 		*bits;
274 	int		error, nf;
275 	size_t		ni;
276 
277 	if (nd < 0)
278 		return (EINVAL);
279 	nf = curlwp->l_fd->fd_dt->dt_nfiles;
280 	if (nd > nf) {
281 		/* forgiving; slightly wrong */
282 		nd = nf;
283 	}
284 	ni = howmany(nd, NFDBITS) * sizeof(fd_mask);
285 	if (ni * 6 > sizeof(smallbits)) {
286 		bits = kmem_alloc(ni * 6, KM_SLEEP);
287 		if (bits == NULL)
288 			return ENOMEM;
289 	} else
290 		bits = smallbits;
291 
292 #define	getbits(name, x)						\
293 	if (u_ ## name) {						\
294 		error = copyin(u_ ## name, bits + ni * x, ni);		\
295 		if (error)						\
296 			goto done;					\
297 	} else								\
298 		memset(bits + ni * x, 0, ni);
299 	getbits(in, 0);
300 	getbits(ou, 1);
301 	getbits(ex, 2);
302 #undef	getbits
303 
304 	error = sel_do_scan(bits, nd, ts, mask, retval, 1);
305  done:
306 	/* select is not restarted after signals... */
307 	if (error == ERESTART)
308 		error = EINTR;
309 	if (error == EWOULDBLOCK)
310 		error = 0;
311 	if (error == 0 && u_in != NULL)
312 		error = copyout(bits + ni * 3, u_in, ni);
313 	if (error == 0 && u_ou != NULL)
314 		error = copyout(bits + ni * 4, u_ou, ni);
315 	if (error == 0 && u_ex != NULL)
316 		error = copyout(bits + ni * 5, u_ex, ni);
317 	if (bits != smallbits)
318 		kmem_free(bits, ni * 6);
319 	return (error);
320 }
321 
322 static inline int
323 selscan(char *bits, u_int nfd, register_t *retval)
324 {
325 	static const int flag[3] = { POLLRDNORM | POLLHUP | POLLERR,
326 			       POLLWRNORM | POLLHUP | POLLERR,
327 			       POLLRDBAND };
328 	fd_mask *ibitp, *obitp;
329 	int msk, i, j, fd, ni, n;
330 	fd_mask ibits, obits;
331 	file_t *fp;
332 
333 	ni = howmany(nfd, NFDBITS) * sizeof(fd_mask);
334 	ibitp = (fd_mask *)(bits + ni * 0);
335 	obitp = (fd_mask *)(bits + ni * 3);
336 	n = 0;
337 
338 	for (msk = 0; msk < 3; msk++) {
339 		for (i = 0; i < nfd; i += NFDBITS) {
340 			ibits = *ibitp++;
341 			obits = 0;
342 			while ((j = ffs(ibits)) && (fd = i + --j) < nfd) {
343 				ibits &= ~(1 << j);
344 				if ((fp = fd_getfile(fd)) == NULL)
345 					return (EBADF);
346 				if ((*fp->f_ops->fo_poll)(fp, flag[msk])) {
347 					obits |= (1 << j);
348 					n++;
349 				}
350 				fd_putfile(fd);
351 			}
352 			*obitp++ = obits;
353 		}
354 	}
355 	*retval = n;
356 	return (0);
357 }
358 
359 /*
360  * Poll system call.
361  */
362 int
363 sys_poll(struct lwp *l, const struct sys_poll_args *uap, register_t *retval)
364 {
365 	/* {
366 		syscallarg(struct pollfd *)	fds;
367 		syscallarg(u_int)		nfds;
368 		syscallarg(int)			timeout;
369 	} */
370 	struct timespec	ats, *ts = NULL;
371 
372 	if (SCARG(uap, timeout) != INFTIM) {
373 		ats.tv_sec = SCARG(uap, timeout) / 1000;
374 		ats.tv_nsec = (SCARG(uap, timeout) % 1000) * 1000000;
375 		ts = &ats;
376 	}
377 
378 	return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, NULL);
379 }
380 
381 /*
382  * Poll system call.
383  */
384 int
385 sys___pollts50(struct lwp *l, const struct sys___pollts50_args *uap,
386     register_t *retval)
387 {
388 	/* {
389 		syscallarg(struct pollfd *)		fds;
390 		syscallarg(u_int)			nfds;
391 		syscallarg(const struct timespec *)	ts;
392 		syscallarg(const sigset_t *)		mask;
393 	} */
394 	struct timespec	ats, *ts = NULL;
395 	sigset_t	amask, *mask = NULL;
396 	int		error;
397 
398 	if (SCARG(uap, ts)) {
399 		error = copyin(SCARG(uap, ts), &ats, sizeof(ats));
400 		if (error)
401 			return error;
402 		ts = &ats;
403 	}
404 	if (SCARG(uap, mask)) {
405 		error = copyin(SCARG(uap, mask), &amask, sizeof(amask));
406 		if (error)
407 			return error;
408 		mask = &amask;
409 	}
410 
411 	return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, mask);
412 }
413 
414 int
415 pollcommon(register_t *retval, struct pollfd *u_fds, u_int nfds,
416     struct timespec *ts, sigset_t *mask)
417 {
418 	struct pollfd	smallfds[32];
419 	struct pollfd	*fds;
420 	int		error;
421 	size_t		ni, nf;
422 
423 	nf = curlwp->l_fd->fd_dt->dt_nfiles;
424 	if (nfds > nf) {
425 		/* forgiving; slightly wrong */
426 		nfds = nf;
427 	}
428 	ni = nfds * sizeof(struct pollfd);
429 	if (ni > sizeof(smallfds)) {
430 		fds = kmem_alloc(ni, KM_SLEEP);
431 		if (fds == NULL)
432 			return ENOMEM;
433 	} else
434 		fds = smallfds;
435 
436 	error = copyin(u_fds, fds, ni);
437 	if (error)
438 		goto done;
439 
440 	error = sel_do_scan(fds, nfds, ts, mask, retval, 0);
441  done:
442 	/* poll is not restarted after signals... */
443 	if (error == ERESTART)
444 		error = EINTR;
445 	if (error == EWOULDBLOCK)
446 		error = 0;
447 	if (error == 0)
448 		error = copyout(fds, u_fds, ni);
449 	if (fds != smallfds)
450 		kmem_free(fds, ni);
451 	return (error);
452 }
453 
454 static inline int
455 pollscan(struct pollfd *fds, u_int nfd, register_t *retval)
456 {
457 	int i, n;
458 	file_t *fp;
459 
460 	n = 0;
461 	for (i = 0; i < nfd; i++, fds++) {
462 		if (fds->fd < 0) {
463 			fds->revents = 0;
464 		} else if ((fp = fd_getfile(fds->fd)) == NULL) {
465 			fds->revents = POLLNVAL;
466 			n++;
467 		} else {
468 			fds->revents = (*fp->f_ops->fo_poll)(fp,
469 			    fds->events | POLLERR | POLLHUP);
470 			if (fds->revents != 0)
471 				n++;
472 			fd_putfile(fds->fd);
473 		}
474 	}
475 	*retval = n;
476 	return (0);
477 }
478 
479 /*ARGSUSED*/
480 int
481 seltrue(dev_t dev, int events, lwp_t *l)
482 {
483 
484 	return (events & (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM));
485 }
486 
487 /*
488  * Record a select request.  Concurrency issues:
489  *
490  * The caller holds the same lock across calls to selrecord() and
491  * selnotify(), so we don't need to consider a concurrent wakeup
492  * while in this routine.
493  *
494  * The only activity we need to guard against is selclear(), called by
495  * another thread that is exiting sel_do_scan().
496  * `sel_lwp' can only become non-NULL while the caller's lock is held,
497  * so it cannot become non-NULL due to a change made by another thread
498  * while we are in this routine.  It can only become _NULL_ due to a
499  * call to selclear().
500  *
501  * If it is non-NULL and != selector there is the potential for
502  * selclear() to be called by another thread.  If either of those
503  * conditions are true, we're not interested in touching the `named
504  * waiter' part of the selinfo record because we need to record a
505  * collision.  Hence there is no need for additional locking in this
506  * routine.
507  */
508 void
509 selrecord(lwp_t *selector, struct selinfo *sip)
510 {
511 	selcpu_t *sc;
512 	lwp_t *other;
513 
514 	KASSERT(selector == curlwp);
515 
516 	sc = selector->l_selcpu;
517 	other = sip->sel_lwp;
518 
519 	if (other == selector) {
520 		/* `selector' has already claimed it. */
521 		KASSERT(sip->sel_cpu = sc);
522 	} else if (other == NULL) {
523 		/*
524 		 * First named waiter, although there may be unnamed
525 		 * waiters (collisions).  Issue a memory barrier to
526 		 * ensure that we access sel_lwp (above) before other
527 		 * fields - this guards against a call to selclear().
528 		 */
529 		membar_enter();
530 		sip->sel_lwp = selector;
531 		SLIST_INSERT_HEAD(&selector->l_selwait, sip, sel_chain);
532 		/* Replace selinfo's lock with our chosen CPU's lock. */
533 		sip->sel_cpu = sc;
534 	} else {
535 		/* Multiple waiters: record a collision. */
536 		sip->sel_collision |= sc->sc_mask;
537 		KASSERT(sip->sel_cpu != NULL);
538 	}
539 }
540 
541 /*
542  * Do a wakeup when a selectable event occurs.  Concurrency issues:
543  *
544  * As per selrecord(), the caller's object lock is held.  If there
545  * is a named waiter, we must acquire the associated selcpu's lock
546  * in order to synchronize with selclear() and pollers going to sleep
547  * in sel_do_scan().
548  *
549  * sip->sel_cpu cannot change at this point, as it is only changed
550  * in selrecord(), and concurrent calls to selrecord() are locked
551  * out by the caller.
552  */
553 void
554 selnotify(struct selinfo *sip, int events, long knhint)
555 {
556 	selcpu_t *sc;
557 	uint32_t mask;
558 	int index, oflag;
559 	lwp_t *l;
560 	kmutex_t *lock;
561 
562 	KNOTE(&sip->sel_klist, knhint);
563 
564 	if (sip->sel_lwp != NULL) {
565 		/* One named LWP is waiting. */
566 		sc = sip->sel_cpu;
567 		lock = sc->sc_lock;
568 		mutex_spin_enter(lock);
569 		/* Still there? */
570 		if (sip->sel_lwp != NULL) {
571 			l = sip->sel_lwp;
572 			/*
573 			 * If thread is sleeping, wake it up.  If it's not
574 			 * yet asleep, it will notice the change in state
575 			 * and will re-poll the descriptors.
576 			 */
577 			oflag = l->l_selflag;
578 			l->l_selflag = SEL_RESET;
579 			if (oflag == SEL_BLOCKING && l->l_mutex == lock) {
580 				KASSERT(l->l_wchan == sc);
581 				sleepq_unsleep(l, false);
582 			}
583 		}
584 		mutex_spin_exit(lock);
585 	}
586 
587 	if ((mask = sip->sel_collision) != 0) {
588 		/*
589 		 * There was a collision (multiple waiters): we must
590 		 * inform all potentially interested waiters.
591 		 */
592 		sip->sel_collision = 0;
593 		do {
594 			index = ffs(mask) - 1;
595 			mask &= ~(1 << index);
596 			sc = cpu_lookup(index)->ci_data.cpu_selcpu;
597 			lock = sc->sc_lock;
598 			mutex_spin_enter(lock);
599 			sc->sc_ncoll++;
600 			sleepq_wake(&sc->sc_sleepq, sc, (u_int)-1, lock);
601 		} while (__predict_false(mask != 0));
602 	}
603 }
604 
605 /*
606  * Remove an LWP from all objects that it is waiting for.  Concurrency
607  * issues:
608  *
609  * The object owner's (e.g. device driver) lock is not held here.  Calls
610  * can be made to selrecord() and we do not synchronize against those
611  * directly using locks.  However, we use `sel_lwp' to lock out changes.
612  * Before clearing it we must use memory barriers to ensure that we can
613  * safely traverse the list of selinfo records.
614  */
615 static void
616 selclear(void)
617 {
618 	struct selinfo *sip, *next;
619 	selcpu_t *sc;
620 	lwp_t *l;
621 	kmutex_t *lock;
622 
623 	l = curlwp;
624 	sc = l->l_selcpu;
625 	lock = sc->sc_lock;
626 
627 	mutex_spin_enter(lock);
628 	for (sip = SLIST_FIRST(&l->l_selwait); sip != NULL; sip = next) {
629 		KASSERT(sip->sel_lwp == l);
630 		KASSERT(sip->sel_cpu == l->l_selcpu);
631 		/*
632 		 * Read link to next selinfo record, if any.
633 		 * It's no longer safe to touch `sip' after clearing
634 		 * `sel_lwp', so ensure that the read of `sel_chain'
635 		 * completes before the clearing of sel_lwp becomes
636 		 * globally visible.
637 		 */
638 		next = SLIST_NEXT(sip, sel_chain);
639 		membar_exit();
640 		/* Release the record for another named waiter to use. */
641 		sip->sel_lwp = NULL;
642 	}
643 	mutex_spin_exit(lock);
644 }
645 
646 /*
647  * Initialize the select/poll system calls.  Called once for each
648  * CPU in the system, as they are attached.
649  */
650 void
651 selsysinit(struct cpu_info *ci)
652 {
653 	selcpu_t *sc;
654 
655 	sc = kmem_alloc(roundup2(sizeof(selcpu_t), coherency_unit) +
656 	    coherency_unit, KM_SLEEP);
657 	sc = (void *)roundup2((uintptr_t)sc, coherency_unit);
658 	sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SCHED);
659 	sleepq_init(&sc->sc_sleepq);
660 	sc->sc_ncoll = 0;
661 	sc->sc_mask = (1 << cpu_index(ci));
662 	ci->ci_data.cpu_selcpu = sc;
663 }
664 
665 /*
666  * Initialize a selinfo record.
667  */
668 void
669 selinit(struct selinfo *sip)
670 {
671 
672 	memset(sip, 0, sizeof(*sip));
673 }
674 
675 /*
676  * Destroy a selinfo record.  The owning object must not gain new
677  * references while this is in progress: all activity on the record
678  * must be stopped.
679  *
680  * Concurrency issues: we only need guard against a call to selclear()
681  * by a thread exiting sel_do_scan().  The caller has prevented further
682  * references being made to the selinfo record via selrecord(), and it
683  * won't call selwakeup() again.
684  */
685 void
686 seldestroy(struct selinfo *sip)
687 {
688 	selcpu_t *sc;
689 	kmutex_t *lock;
690 	lwp_t *l;
691 
692 	if (sip->sel_lwp == NULL)
693 		return;
694 
695 	/*
696 	 * Lock out selclear().  The selcpu pointer can't change while
697 	 * we are here since it is only ever changed in selrecord(),
698 	 * and that will not be entered again for this record because
699 	 * it is dying.
700 	 */
701 	KASSERT(sip->sel_cpu != NULL);
702 	sc = sip->sel_cpu;
703 	lock = sc->sc_lock;
704 	mutex_spin_enter(lock);
705 	if ((l = sip->sel_lwp) != NULL) {
706 		/*
707 		 * This should rarely happen, so although SLIST_REMOVE()
708 		 * is slow, using it here is not a problem.
709 		 */
710 		KASSERT(l->l_selcpu == sc);
711 		SLIST_REMOVE(&l->l_selwait, sip, selinfo, sel_chain);
712 		sip->sel_lwp = NULL;
713 	}
714 	mutex_spin_exit(lock);
715 }
716 
717 int
718 pollsock(struct socket *so, const struct timespec *tsp, int events)
719 {
720 	int		ncoll, error, timo;
721 	struct timespec	sleepts, ts;
722 	selcpu_t	*sc;
723 	lwp_t		*l;
724 	kmutex_t	*lock;
725 
726 	timo = 0;
727 	if (tsp != NULL) {
728 		ts = *tsp;
729 		if (inittimeleft(&ts, &sleepts) == -1)
730 			return EINVAL;
731 	}
732 
733 	l = curlwp;
734 	sc = l->l_cpu->ci_data.cpu_selcpu;
735 	lock = sc->sc_lock;
736 	l->l_selcpu = sc;
737 	SLIST_INIT(&l->l_selwait);
738 	error = 0;
739 	for (;;) {
740 		/*
741 		 * No need to lock.  If this is overwritten by another
742 		 * value while scanning, we will retry below.  We only
743 		 * need to see exact state from the descriptors that
744 		 * we are about to poll, and lock activity resulting
745 		 * from fo_poll is enough to provide an up to date value
746 		 * for new polling activity.
747 		 */
748 		ncoll = sc->sc_ncoll;
749 		l->l_selflag = SEL_SCANNING;
750 		if (sopoll(so, events) != 0)
751 			break;
752 		if (tsp && (timo = gettimeleft(&ts, &sleepts)) <= 0)
753 			break;
754 		mutex_spin_enter(lock);
755 		if (l->l_selflag != SEL_SCANNING || sc->sc_ncoll != ncoll) {
756 			mutex_spin_exit(lock);
757 			continue;
758 		}
759 		l->l_selflag = SEL_BLOCKING;
760 		sleepq_enter(&sc->sc_sleepq, l, lock);
761 		sleepq_enqueue(&sc->sc_sleepq, sc, "pollsock", &select_sobj);
762 		error = sleepq_block(timo, true);
763 		if (error != 0)
764 			break;
765 	}
766 	selclear();
767 	/* poll is not restarted after signals... */
768 	if (error == ERESTART)
769 		error = EINTR;
770 	if (error == EWOULDBLOCK)
771 		error = 0;
772 	return (error);
773 }
774