xref: /netbsd-src/sys/kern/sys_select.c (revision 0647473106440a964f74e4da0d9101cd009f0b0d)
1 /*	$NetBSD: sys_select.c,v 1.32 2011/05/18 14:48:04 christos Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007, 2008, 2009, 2010 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 and Mindaugas Rasiukevicius.
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 of synchronous I/O multiplexing subsystem.
70  *
71  * Locking
72  *
73  * Two locks are used: <object-lock> and selcluster_t::sc_lock.
74  *
75  * The <object-lock> might be a device driver or another subsystem, e.g.
76  * socket or pipe.  This lock is not exported, and thus invisible to this
77  * subsystem.  Mainly, synchronisation between selrecord() and selnotify()
78  * routines depends on this lock, as it will be described in the comments.
79  *
80  * Lock order
81  *
82  *	<object-lock> ->
83  *		selcluster_t::sc_lock
84  */
85 
86 #include <sys/cdefs.h>
87 __KERNEL_RCSID(0, "$NetBSD: sys_select.c,v 1.32 2011/05/18 14:48:04 christos Exp $");
88 
89 #include <sys/param.h>
90 #include <sys/systm.h>
91 #include <sys/filedesc.h>
92 #include <sys/file.h>
93 #include <sys/proc.h>
94 #include <sys/socketvar.h>
95 #include <sys/signalvar.h>
96 #include <sys/uio.h>
97 #include <sys/kernel.h>
98 #include <sys/lwp.h>
99 #include <sys/poll.h>
100 #include <sys/mount.h>
101 #include <sys/syscallargs.h>
102 #include <sys/cpu.h>
103 #include <sys/atomic.h>
104 #include <sys/socketvar.h>
105 #include <sys/sleepq.h>
106 
107 /* Flags for lwp::l_selflag. */
108 #define	SEL_RESET	0	/* awoken, interrupted, or not yet polling */
109 #define	SEL_SCANNING	1	/* polling descriptors */
110 #define	SEL_BLOCKING	2	/* blocking and waiting for event */
111 #define	SEL_EVENT	3	/* interrupted, events set directly */
112 
113 /* Operations: either select() or poll(). */
114 #define	SELOP_SELECT	1
115 #define	SELOP_POLL	2
116 
117 /*
118  * Per-cluster state for select()/poll().  For a system with fewer
119  * than 32 CPUs, this gives us per-CPU clusters.
120  */
121 #define	SELCLUSTERS	32
122 #define	SELCLUSTERMASK	(SELCLUSTERS - 1)
123 
124 typedef struct selcluster {
125 	kmutex_t	*sc_lock;
126 	sleepq_t	sc_sleepq;
127 	int		sc_ncoll;
128 	uint32_t	sc_mask;
129 } selcluster_t;
130 
131 static inline int	selscan(char *, const int, const size_t, register_t *);
132 static inline int	pollscan(struct pollfd *, const int, register_t *);
133 static void		selclear(void);
134 
135 static const int sel_flag[] = {
136 	POLLRDNORM | POLLHUP | POLLERR,
137 	POLLWRNORM | POLLHUP | POLLERR,
138 	POLLRDBAND
139 };
140 
141 static syncobj_t select_sobj = {
142 	SOBJ_SLEEPQ_FIFO,
143 	sleepq_unsleep,
144 	sleepq_changepri,
145 	sleepq_lendpri,
146 	syncobj_noowner,
147 };
148 
149 static selcluster_t	*selcluster[SELCLUSTERS] __read_mostly;
150 
151 /*
152  * Select system call.
153  */
154 int
155 sys___pselect50(struct lwp *l, const struct sys___pselect50_args *uap,
156     register_t *retval)
157 {
158 	/* {
159 		syscallarg(int)				nd;
160 		syscallarg(fd_set *)			in;
161 		syscallarg(fd_set *)			ou;
162 		syscallarg(fd_set *)			ex;
163 		syscallarg(const struct timespec *)	ts;
164 		syscallarg(sigset_t *)			mask;
165 	} */
166 	struct timespec	ats, *ts = NULL;
167 	sigset_t	amask, *mask = NULL;
168 	int		error;
169 
170 	if (SCARG(uap, ts)) {
171 		error = copyin(SCARG(uap, ts), &ats, sizeof(ats));
172 		if (error)
173 			return error;
174 		ts = &ats;
175 	}
176 	if (SCARG(uap, mask) != NULL) {
177 		error = copyin(SCARG(uap, mask), &amask, sizeof(amask));
178 		if (error)
179 			return error;
180 		mask = &amask;
181 	}
182 
183 	return selcommon(retval, SCARG(uap, nd), SCARG(uap, in),
184 	    SCARG(uap, ou), SCARG(uap, ex), ts, mask);
185 }
186 
187 int
188 sys___select50(struct lwp *l, const struct sys___select50_args *uap,
189     register_t *retval)
190 {
191 	/* {
192 		syscallarg(int)			nd;
193 		syscallarg(fd_set *)		in;
194 		syscallarg(fd_set *)		ou;
195 		syscallarg(fd_set *)		ex;
196 		syscallarg(struct timeval *)	tv;
197 	} */
198 	struct timeval atv;
199 	struct timespec ats, *ts = NULL;
200 	int error;
201 
202 	if (SCARG(uap, tv)) {
203 		error = copyin(SCARG(uap, tv), (void *)&atv, sizeof(atv));
204 		if (error)
205 			return error;
206 		TIMEVAL_TO_TIMESPEC(&atv, &ats);
207 		ts = &ats;
208 	}
209 
210 	return selcommon(retval, SCARG(uap, nd), SCARG(uap, in),
211 	    SCARG(uap, ou), SCARG(uap, ex), ts, NULL);
212 }
213 
214 /*
215  * sel_do_scan: common code to perform the scan on descriptors.
216  */
217 static int
218 sel_do_scan(const int op, void *fds, const int nf, const size_t ni,
219     struct timespec *ts, sigset_t *mask, register_t *retval)
220 {
221 	lwp_t		* const l = curlwp;
222 	selcluster_t	*sc;
223 	kmutex_t	*lock;
224 	struct timespec	sleepts;
225 	int		error, timo;
226 
227 	timo = 0;
228 	if (ts && inittimeleft(ts, &sleepts) == -1) {
229 		return EINVAL;
230 	}
231 
232 	if (__predict_false(mask))
233 		sigsuspendsetup(l, mask);
234 
235 	sc = curcpu()->ci_data.cpu_selcluster;
236 	lock = sc->sc_lock;
237 	l->l_selcluster = sc;
238 	SLIST_INIT(&l->l_selwait);
239 
240 	l->l_selret = 0;
241 	if (op == SELOP_SELECT) {
242 		l->l_selbits = fds;
243 		l->l_selni = ni;
244 	} else {
245 		l->l_selbits = NULL;
246 	}
247 	for (;;) {
248 		int ncoll;
249 
250 		/*
251 		 * No need to lock.  If this is overwritten by another value
252 		 * while scanning, we will retry below.  We only need to see
253 		 * exact state from the descriptors that we are about to poll,
254 		 * and lock activity resulting from fo_poll is enough to
255 		 * provide an up to date value for new polling activity.
256 		 */
257 		l->l_selflag = SEL_SCANNING;
258 		ncoll = sc->sc_ncoll;
259 
260 		if (op == SELOP_SELECT) {
261 			error = selscan((char *)fds, nf, ni, retval);
262 		} else {
263 			error = pollscan((struct pollfd *)fds, nf, retval);
264 		}
265 		if (error || *retval)
266 			break;
267 		if (ts && (timo = gettimeleft(ts, &sleepts)) <= 0)
268 			break;
269 		/*
270 		 * Acquire the lock and perform the (re)checks.  Note, if
271 		 * collision has occured, then our state does not matter,
272 		 * as we must perform re-scan.  Therefore, check it first.
273 		 */
274 state_check:
275 		mutex_spin_enter(lock);
276 		if (__predict_false(sc->sc_ncoll != ncoll)) {
277 			/* Collision: perform re-scan. */
278 			mutex_spin_exit(lock);
279 			continue;
280 		}
281 		if (__predict_true(l->l_selflag == SEL_EVENT)) {
282 			/* Events occured, they are set directly. */
283 			mutex_spin_exit(lock);
284 			KASSERT(l->l_selret != 0);
285 			*retval = l->l_selret;
286 			break;
287 		}
288 		if (__predict_true(l->l_selflag == SEL_RESET)) {
289 			/* Events occured, but re-scan is requested. */
290 			mutex_spin_exit(lock);
291 			continue;
292 		}
293 		/* Nothing happen, therefore - sleep. */
294 		l->l_selflag = SEL_BLOCKING;
295 		l->l_kpriority = true;
296 		sleepq_enter(&sc->sc_sleepq, l, lock);
297 		sleepq_enqueue(&sc->sc_sleepq, sc, "select", &select_sobj);
298 		error = sleepq_block(timo, true);
299 		if (error != 0) {
300 			break;
301 		}
302 		/* Awoken: need to check the state. */
303 		goto state_check;
304 	}
305 	selclear();
306 
307 	/* select and poll are not restarted after signals... */
308 	if (error == ERESTART)
309 		return EINTR;
310 	if (error == EWOULDBLOCK)
311 		return 0;
312 	return error;
313 }
314 
315 int
316 selcommon(register_t *retval, int nd, fd_set *u_in, fd_set *u_ou,
317     fd_set *u_ex, struct timespec *ts, sigset_t *mask)
318 {
319 	char		smallbits[howmany(FD_SETSIZE, NFDBITS) *
320 			    sizeof(fd_mask) * 6];
321 	char 		*bits;
322 	int		error, nf;
323 	size_t		ni;
324 
325 	if (nd < 0)
326 		return (EINVAL);
327 	nf = curlwp->l_fd->fd_dt->dt_nfiles;
328 	if (nd > nf) {
329 		/* forgiving; slightly wrong */
330 		nd = nf;
331 	}
332 	ni = howmany(nd, NFDBITS) * sizeof(fd_mask);
333 	if (ni * 6 > sizeof(smallbits)) {
334 		bits = kmem_alloc(ni * 6, KM_SLEEP);
335 		if (bits == NULL)
336 			return ENOMEM;
337 	} else
338 		bits = smallbits;
339 
340 #define	getbits(name, x)						\
341 	if (u_ ## name) {						\
342 		error = copyin(u_ ## name, bits + ni * x, ni);		\
343 		if (error)						\
344 			goto fail;					\
345 	} else								\
346 		memset(bits + ni * x, 0, ni);
347 	getbits(in, 0);
348 	getbits(ou, 1);
349 	getbits(ex, 2);
350 #undef	getbits
351 
352 	error = sel_do_scan(SELOP_SELECT, bits, nd, ni, ts, mask, retval);
353 	if (error == 0 && u_in != NULL)
354 		error = copyout(bits + ni * 3, u_in, ni);
355 	if (error == 0 && u_ou != NULL)
356 		error = copyout(bits + ni * 4, u_ou, ni);
357 	if (error == 0 && u_ex != NULL)
358 		error = copyout(bits + ni * 5, u_ex, ni);
359  fail:
360 	if (bits != smallbits)
361 		kmem_free(bits, ni * 6);
362 	return (error);
363 }
364 
365 static inline int
366 selscan(char *bits, const int nfd, const size_t ni, register_t *retval)
367 {
368 	fd_mask *ibitp, *obitp;
369 	int msk, i, j, fd, n;
370 	file_t *fp;
371 
372 	ibitp = (fd_mask *)(bits + ni * 0);
373 	obitp = (fd_mask *)(bits + ni * 3);
374 	n = 0;
375 
376 	for (msk = 0; msk < 3; msk++) {
377 		for (i = 0; i < nfd; i += NFDBITS) {
378 			fd_mask ibits, obits;
379 
380 			ibits = *ibitp++;
381 			obits = 0;
382 			while ((j = ffs(ibits)) && (fd = i + --j) < nfd) {
383 				ibits &= ~(1 << j);
384 				if ((fp = fd_getfile(fd)) == NULL)
385 					return (EBADF);
386 				/*
387 				 * Setup an argument to selrecord(), which is
388 				 * a file descriptor number.
389 				 */
390 				curlwp->l_selrec = fd;
391 				if ((*fp->f_ops->fo_poll)(fp, sel_flag[msk])) {
392 					obits |= (1 << j);
393 					n++;
394 				}
395 				fd_putfile(fd);
396 			}
397 			*obitp++ = obits;
398 		}
399 	}
400 	*retval = n;
401 	return (0);
402 }
403 
404 /*
405  * Poll system call.
406  */
407 int
408 sys_poll(struct lwp *l, const struct sys_poll_args *uap, register_t *retval)
409 {
410 	/* {
411 		syscallarg(struct pollfd *)	fds;
412 		syscallarg(u_int)		nfds;
413 		syscallarg(int)			timeout;
414 	} */
415 	struct timespec	ats, *ts = NULL;
416 
417 	if (SCARG(uap, timeout) != INFTIM) {
418 		ats.tv_sec = SCARG(uap, timeout) / 1000;
419 		ats.tv_nsec = (SCARG(uap, timeout) % 1000) * 1000000;
420 		ts = &ats;
421 	}
422 
423 	return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, NULL);
424 }
425 
426 /*
427  * Poll system call.
428  */
429 int
430 sys___pollts50(struct lwp *l, const struct sys___pollts50_args *uap,
431     register_t *retval)
432 {
433 	/* {
434 		syscallarg(struct pollfd *)		fds;
435 		syscallarg(u_int)			nfds;
436 		syscallarg(const struct timespec *)	ts;
437 		syscallarg(const sigset_t *)		mask;
438 	} */
439 	struct timespec	ats, *ts = NULL;
440 	sigset_t	amask, *mask = NULL;
441 	int		error;
442 
443 	if (SCARG(uap, ts)) {
444 		error = copyin(SCARG(uap, ts), &ats, sizeof(ats));
445 		if (error)
446 			return error;
447 		ts = &ats;
448 	}
449 	if (SCARG(uap, mask)) {
450 		error = copyin(SCARG(uap, mask), &amask, sizeof(amask));
451 		if (error)
452 			return error;
453 		mask = &amask;
454 	}
455 
456 	return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, mask);
457 }
458 
459 int
460 pollcommon(register_t *retval, struct pollfd *u_fds, u_int nfds,
461     struct timespec *ts, sigset_t *mask)
462 {
463 	struct pollfd	smallfds[32];
464 	struct pollfd	*fds;
465 	int		error;
466 	size_t		ni;
467 
468 	if (nfds > 1000 + curlwp->l_fd->fd_dt->dt_nfiles) {
469 		/*
470 		 * Either the user passed in a very sparse 'fds' or junk!
471 		 * The kmem_alloc() call below would be bad news.
472 		 * We could process the 'fds' array in chunks, but that
473 		 * is a lot of code that isn't normally useful.
474 		 * (Or just move the copyin/out into pollscan().)
475 		 * Historically the code silently truncated 'fds' to
476 		 * dt_nfiles entries - but that does cause issues.
477 		 */
478 		return EINVAL;
479 	}
480 	ni = nfds * sizeof(struct pollfd);
481 	if (ni > sizeof(smallfds)) {
482 		fds = kmem_alloc(ni, KM_SLEEP);
483 		if (fds == NULL)
484 			return ENOMEM;
485 	} else
486 		fds = smallfds;
487 
488 	error = copyin(u_fds, fds, ni);
489 	if (error)
490 		goto fail;
491 
492 	error = sel_do_scan(SELOP_POLL, fds, nfds, ni, ts, mask, retval);
493 	if (error == 0)
494 		error = copyout(fds, u_fds, ni);
495  fail:
496 	if (fds != smallfds)
497 		kmem_free(fds, ni);
498 	return (error);
499 }
500 
501 static inline int
502 pollscan(struct pollfd *fds, const int nfd, register_t *retval)
503 {
504 	file_t *fp;
505 	int i, n = 0;
506 
507 	for (i = 0; i < nfd; i++, fds++) {
508 		if (fds->fd < 0) {
509 			fds->revents = 0;
510 		} else if ((fp = fd_getfile(fds->fd)) == NULL) {
511 			fds->revents = POLLNVAL;
512 			n++;
513 		} else {
514 			/*
515 			 * Perform poll: registers select request or returns
516 			 * the events which are set.  Setup an argument for
517 			 * selrecord(), which is a pointer to struct pollfd.
518 			 */
519 			curlwp->l_selrec = (uintptr_t)fds;
520 			fds->revents = (*fp->f_ops->fo_poll)(fp,
521 			    fds->events | POLLERR | POLLHUP);
522 			if (fds->revents != 0)
523 				n++;
524 			fd_putfile(fds->fd);
525 		}
526 	}
527 	*retval = n;
528 	return (0);
529 }
530 
531 int
532 seltrue(dev_t dev, int events, lwp_t *l)
533 {
534 
535 	return (events & (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM));
536 }
537 
538 /*
539  * Record a select request.  Concurrency issues:
540  *
541  * The caller holds the same lock across calls to selrecord() and
542  * selnotify(), so we don't need to consider a concurrent wakeup
543  * while in this routine.
544  *
545  * The only activity we need to guard against is selclear(), called by
546  * another thread that is exiting sel_do_scan().
547  * `sel_lwp' can only become non-NULL while the caller's lock is held,
548  * so it cannot become non-NULL due to a change made by another thread
549  * while we are in this routine.  It can only become _NULL_ due to a
550  * call to selclear().
551  *
552  * If it is non-NULL and != selector there is the potential for
553  * selclear() to be called by another thread.  If either of those
554  * conditions are true, we're not interested in touching the `named
555  * waiter' part of the selinfo record because we need to record a
556  * collision.  Hence there is no need for additional locking in this
557  * routine.
558  */
559 void
560 selrecord(lwp_t *selector, struct selinfo *sip)
561 {
562 	selcluster_t *sc;
563 	lwp_t *other;
564 
565 	KASSERT(selector == curlwp);
566 
567 	sc = selector->l_selcluster;
568 	other = sip->sel_lwp;
569 
570 	if (other == selector) {
571 		/* 1. We (selector) already claimed to be the first LWP. */
572 		KASSERT(sip->sel_cluster = sc);
573 	} else if (other == NULL) {
574 		/*
575 		 * 2. No first LWP, therefore we (selector) are the first.
576 		 *
577 		 * There may be unnamed waiters (collisions).  Issue a memory
578 		 * barrier to ensure that we access sel_lwp (above) before
579 		 * other fields - this guards against a call to selclear().
580 		 */
581 		membar_enter();
582 		sip->sel_lwp = selector;
583 		SLIST_INSERT_HEAD(&selector->l_selwait, sip, sel_chain);
584 		/* Copy the argument, which is for selnotify(). */
585 		sip->sel_fdinfo = selector->l_selrec;
586 		/* Replace selinfo's lock with the chosen cluster's lock. */
587 		sip->sel_cluster = sc;
588 	} else {
589 		/* 3. Multiple waiters: record a collision. */
590 		sip->sel_collision |= sc->sc_mask;
591 		KASSERT(sip->sel_cluster != NULL);
592 	}
593 }
594 
595 /*
596  * sel_setevents: a helper function for selnotify(), to set the events
597  * for LWP sleeping in selcommon() or pollcommon().
598  */
599 static inline bool
600 sel_setevents(lwp_t *l, struct selinfo *sip, const int events)
601 {
602 	const int oflag = l->l_selflag;
603 	int ret = 0;
604 
605 	/*
606 	 * If we require re-scan or it was required by somebody else,
607 	 * then just (re)set SEL_RESET and return.
608 	 */
609 	if (__predict_false(events == 0 || oflag == SEL_RESET)) {
610 		l->l_selflag = SEL_RESET;
611 		return true;
612 	}
613 	/*
614 	 * Direct set.  Note: select state of LWP is locked.  First,
615 	 * determine whether it is selcommon() or pollcommon().
616 	 */
617 	if (l->l_selbits != NULL) {
618 		const size_t ni = l->l_selni;
619 		fd_mask *fds = (fd_mask *)l->l_selbits;
620 		fd_mask *ofds = (fd_mask *)((char *)fds + ni * 3);
621 		const int fd = sip->sel_fdinfo, fbit = 1 << (fd & __NFDMASK);
622 		const int idx = fd >> __NFDSHIFT;
623 		int n;
624 
625 		for (n = 0; n < 3; n++) {
626 			if ((fds[idx] & fbit) != 0 && (sel_flag[n] & events)) {
627 				ofds[idx] |= fbit;
628 				ret++;
629 			}
630 			fds = (fd_mask *)((char *)fds + ni);
631 			ofds = (fd_mask *)((char *)ofds + ni);
632 		}
633 	} else {
634 		struct pollfd *pfd = (void *)sip->sel_fdinfo;
635 		int revents = events & (pfd->events | POLLERR | POLLHUP);
636 
637 		if (revents) {
638 			pfd->revents |= revents;
639 			ret = 1;
640 		}
641 	}
642 	/* Check whether there are any events to return. */
643 	if (!ret) {
644 		return false;
645 	}
646 	/* Indicate direct set and note the event (cluster lock is held). */
647 	l->l_selflag = SEL_EVENT;
648 	l->l_selret += ret;
649 	return true;
650 }
651 
652 /*
653  * Do a wakeup when a selectable event occurs.  Concurrency issues:
654  *
655  * As per selrecord(), the caller's object lock is held.  If there
656  * is a named waiter, we must acquire the associated selcluster's lock
657  * in order to synchronize with selclear() and pollers going to sleep
658  * in sel_do_scan().
659  *
660  * sip->sel_cluser cannot change at this point, as it is only changed
661  * in selrecord(), and concurrent calls to selrecord() are locked
662  * out by the caller.
663  */
664 void
665 selnotify(struct selinfo *sip, int events, long knhint)
666 {
667 	selcluster_t *sc;
668 	uint32_t mask;
669 	int index, oflag;
670 	lwp_t *l;
671 	kmutex_t *lock;
672 
673 	KNOTE(&sip->sel_klist, knhint);
674 
675 	if (sip->sel_lwp != NULL) {
676 		/* One named LWP is waiting. */
677 		sc = sip->sel_cluster;
678 		lock = sc->sc_lock;
679 		mutex_spin_enter(lock);
680 		/* Still there? */
681 		if (sip->sel_lwp != NULL) {
682 			/*
683 			 * Set the events for our LWP and indicate that.
684 			 * Otherwise, request for a full re-scan.
685 			 */
686 			l = sip->sel_lwp;
687 			oflag = l->l_selflag;
688 #ifndef NO_DIRECT_SELECT
689 			if (!sel_setevents(l, sip, events)) {
690 				/* No events to return. */
691 				mutex_spin_exit(lock);
692 				return;
693 			}
694 #else
695 			l->l_selflag = SEL_RESET;
696 #endif
697 			/*
698 			 * If thread is sleeping, wake it up.  If it's not
699 			 * yet asleep, it will notice the change in state
700 			 * and will re-poll the descriptors.
701 			 */
702 			if (oflag == SEL_BLOCKING && l->l_mutex == lock) {
703 				KASSERT(l->l_wchan == sc);
704 				sleepq_unsleep(l, false);
705 			}
706 		}
707 		mutex_spin_exit(lock);
708 	}
709 
710 	if ((mask = sip->sel_collision) != 0) {
711 		/*
712 		 * There was a collision (multiple waiters): we must
713 		 * inform all potentially interested waiters.
714 		 */
715 		sip->sel_collision = 0;
716 		do {
717 			index = ffs(mask) - 1;
718 			mask &= ~(1 << index);
719 			sc = selcluster[index];
720 			lock = sc->sc_lock;
721 			mutex_spin_enter(lock);
722 			sc->sc_ncoll++;
723 			sleepq_wake(&sc->sc_sleepq, sc, (u_int)-1, lock);
724 		} while (__predict_false(mask != 0));
725 	}
726 }
727 
728 /*
729  * Remove an LWP from all objects that it is waiting for.  Concurrency
730  * issues:
731  *
732  * The object owner's (e.g. device driver) lock is not held here.  Calls
733  * can be made to selrecord() and we do not synchronize against those
734  * directly using locks.  However, we use `sel_lwp' to lock out changes.
735  * Before clearing it we must use memory barriers to ensure that we can
736  * safely traverse the list of selinfo records.
737  */
738 static void
739 selclear(void)
740 {
741 	struct selinfo *sip, *next;
742 	selcluster_t *sc;
743 	lwp_t *l;
744 	kmutex_t *lock;
745 
746 	l = curlwp;
747 	sc = l->l_selcluster;
748 	lock = sc->sc_lock;
749 
750 	mutex_spin_enter(lock);
751 	for (sip = SLIST_FIRST(&l->l_selwait); sip != NULL; sip = next) {
752 		KASSERT(sip->sel_lwp == l);
753 		KASSERT(sip->sel_cluster == l->l_selcluster);
754 
755 		/*
756 		 * Read link to next selinfo record, if any.
757 		 * It's no longer safe to touch `sip' after clearing
758 		 * `sel_lwp', so ensure that the read of `sel_chain'
759 		 * completes before the clearing of sel_lwp becomes
760 		 * globally visible.
761 		 */
762 		next = SLIST_NEXT(sip, sel_chain);
763 		membar_exit();
764 		/* Release the record for another named waiter to use. */
765 		sip->sel_lwp = NULL;
766 	}
767 	mutex_spin_exit(lock);
768 }
769 
770 /*
771  * Initialize the select/poll system calls.  Called once for each
772  * CPU in the system, as they are attached.
773  */
774 void
775 selsysinit(struct cpu_info *ci)
776 {
777 	selcluster_t *sc;
778 	u_int index;
779 
780 	/* If already a cluster in place for this bit, re-use. */
781 	index = cpu_index(ci) & SELCLUSTERMASK;
782 	sc = selcluster[index];
783 	if (sc == NULL) {
784 		sc = kmem_alloc(roundup2(sizeof(selcluster_t),
785 		    coherency_unit) + coherency_unit, KM_SLEEP);
786 		sc = (void *)roundup2((uintptr_t)sc, coherency_unit);
787 		sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SCHED);
788 		sleepq_init(&sc->sc_sleepq);
789 		sc->sc_ncoll = 0;
790 		sc->sc_mask = (1 << index);
791 		selcluster[index] = sc;
792 	}
793 	ci->ci_data.cpu_selcluster = sc;
794 }
795 
796 /*
797  * Initialize a selinfo record.
798  */
799 void
800 selinit(struct selinfo *sip)
801 {
802 
803 	memset(sip, 0, sizeof(*sip));
804 }
805 
806 /*
807  * Destroy a selinfo record.  The owning object must not gain new
808  * references while this is in progress: all activity on the record
809  * must be stopped.
810  *
811  * Concurrency issues: we only need guard against a call to selclear()
812  * by a thread exiting sel_do_scan().  The caller has prevented further
813  * references being made to the selinfo record via selrecord(), and it
814  * will not call selnotify() again.
815  */
816 void
817 seldestroy(struct selinfo *sip)
818 {
819 	selcluster_t *sc;
820 	kmutex_t *lock;
821 	lwp_t *l;
822 
823 	if (sip->sel_lwp == NULL)
824 		return;
825 
826 	/*
827 	 * Lock out selclear().  The selcluster pointer can't change while
828 	 * we are here since it is only ever changed in selrecord(),
829 	 * and that will not be entered again for this record because
830 	 * it is dying.
831 	 */
832 	KASSERT(sip->sel_cluster != NULL);
833 	sc = sip->sel_cluster;
834 	lock = sc->sc_lock;
835 	mutex_spin_enter(lock);
836 	if ((l = sip->sel_lwp) != NULL) {
837 		/*
838 		 * This should rarely happen, so although SLIST_REMOVE()
839 		 * is slow, using it here is not a problem.
840 		 */
841 		KASSERT(l->l_selcluster == sc);
842 		SLIST_REMOVE(&l->l_selwait, sip, selinfo, sel_chain);
843 		sip->sel_lwp = NULL;
844 	}
845 	mutex_spin_exit(lock);
846 }
847 
848 int
849 pollsock(struct socket *so, const struct timespec *tsp, int events)
850 {
851 	int		ncoll, error, timo;
852 	struct timespec	sleepts, ts;
853 	selcluster_t	*sc;
854 	lwp_t		*l;
855 	kmutex_t	*lock;
856 
857 	timo = 0;
858 	if (tsp != NULL) {
859 		ts = *tsp;
860 		if (inittimeleft(&ts, &sleepts) == -1)
861 			return EINVAL;
862 	}
863 
864 	l = curlwp;
865 	sc = curcpu()->ci_data.cpu_selcluster;
866 	lock = sc->sc_lock;
867 	l->l_selcluster = sc;
868 	SLIST_INIT(&l->l_selwait);
869 	error = 0;
870 	for (;;) {
871 		/*
872 		 * No need to lock.  If this is overwritten by another
873 		 * value while scanning, we will retry below.  We only
874 		 * need to see exact state from the descriptors that
875 		 * we are about to poll, and lock activity resulting
876 		 * from fo_poll is enough to provide an up to date value
877 		 * for new polling activity.
878 		 */
879 		ncoll = sc->sc_ncoll;
880 		l->l_selflag = SEL_SCANNING;
881 		if (sopoll(so, events) != 0)
882 			break;
883 		if (tsp && (timo = gettimeleft(&ts, &sleepts)) <= 0)
884 			break;
885 		mutex_spin_enter(lock);
886 		if (l->l_selflag != SEL_SCANNING || sc->sc_ncoll != ncoll) {
887 			mutex_spin_exit(lock);
888 			continue;
889 		}
890 		l->l_selflag = SEL_BLOCKING;
891 		sleepq_enter(&sc->sc_sleepq, l, lock);
892 		sleepq_enqueue(&sc->sc_sleepq, sc, "pollsock", &select_sobj);
893 		error = sleepq_block(timo, true);
894 		if (error != 0)
895 			break;
896 	}
897 	selclear();
898 	/* poll is not restarted after signals... */
899 	if (error == ERESTART)
900 		error = EINTR;
901 	if (error == EWOULDBLOCK)
902 		error = 0;
903 	return (error);
904 }
905