xref: /netbsd-src/sys/kern/sys_select.c (revision 404ee5b9334f618040b6cdef96a0ff35a6fc4636)
1 /*	$NetBSD: sys_select.c,v 1.48 2019/09/20 15:00:47 kamil 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 1.48 2019/09/20 15:00:47 kamil 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 #include <sys/sysctl.h>
107 
108 /* Flags for lwp::l_selflag. */
109 #define	SEL_RESET	0	/* awoken, interrupted, or not yet polling */
110 #define	SEL_SCANNING	1	/* polling descriptors */
111 #define	SEL_BLOCKING	2	/* blocking and waiting for event */
112 #define	SEL_EVENT	3	/* interrupted, events set directly */
113 
114 /* Operations: either select() or poll(). */
115 #define	SELOP_SELECT	1
116 #define	SELOP_POLL	2
117 
118 /*
119  * Per-cluster state for select()/poll().  For a system with fewer
120  * than 32 CPUs, this gives us per-CPU clusters.
121  */
122 #define	SELCLUSTERS	32
123 #define	SELCLUSTERMASK	(SELCLUSTERS - 1)
124 
125 typedef struct selcluster {
126 	kmutex_t	*sc_lock;
127 	sleepq_t	sc_sleepq;
128 	int		sc_ncoll;
129 	uint32_t	sc_mask;
130 } selcluster_t;
131 
132 static inline int	selscan(char *, const int, const size_t, register_t *);
133 static inline int	pollscan(struct pollfd *, const int, register_t *);
134 static void		selclear(void);
135 
136 static const int sel_flag[] = {
137 	POLLRDNORM | POLLHUP | POLLERR,
138 	POLLWRNORM | POLLHUP | POLLERR,
139 	POLLRDBAND
140 };
141 
142 static syncobj_t select_sobj = {
143 	.sobj_flag	= SOBJ_SLEEPQ_FIFO,
144 	.sobj_unsleep	= sleepq_unsleep,
145 	.sobj_changepri	= sleepq_changepri,
146 	.sobj_lendpri	= sleepq_lendpri,
147 	.sobj_owner	= syncobj_noowner,
148 };
149 
150 static selcluster_t	*selcluster[SELCLUSTERS] __read_mostly;
151 static int		direct_select __read_mostly = 0;
152 
153 /*
154  * Select system call.
155  */
156 int
157 sys___pselect50(struct lwp *l, const struct sys___pselect50_args *uap,
158     register_t *retval)
159 {
160 	/* {
161 		syscallarg(int)				nd;
162 		syscallarg(fd_set *)			in;
163 		syscallarg(fd_set *)			ou;
164 		syscallarg(fd_set *)			ex;
165 		syscallarg(const struct timespec *)	ts;
166 		syscallarg(sigset_t *)			mask;
167 	} */
168 	struct timespec	ats, *ts = NULL;
169 	sigset_t	amask, *mask = NULL;
170 	int		error;
171 
172 	if (SCARG(uap, ts)) {
173 		error = copyin(SCARG(uap, ts), &ats, sizeof(ats));
174 		if (error)
175 			return error;
176 		ts = &ats;
177 	}
178 	if (SCARG(uap, mask) != NULL) {
179 		error = copyin(SCARG(uap, mask), &amask, sizeof(amask));
180 		if (error)
181 			return error;
182 		mask = &amask;
183 	}
184 
185 	return selcommon(retval, SCARG(uap, nd), SCARG(uap, in),
186 	    SCARG(uap, ou), SCARG(uap, ex), ts, mask);
187 }
188 
189 int
190 sys___select50(struct lwp *l, const struct sys___select50_args *uap,
191     register_t *retval)
192 {
193 	/* {
194 		syscallarg(int)			nd;
195 		syscallarg(fd_set *)		in;
196 		syscallarg(fd_set *)		ou;
197 		syscallarg(fd_set *)		ex;
198 		syscallarg(struct timeval *)	tv;
199 	} */
200 	struct timeval atv;
201 	struct timespec ats, *ts = NULL;
202 	int error;
203 
204 	if (SCARG(uap, tv)) {
205 		error = copyin(SCARG(uap, tv), (void *)&atv, sizeof(atv));
206 		if (error)
207 			return error;
208 
209 		if (atv.tv_usec < 0 || atv.tv_usec >= 1000000)
210 			return EINVAL;
211 
212 		TIMEVAL_TO_TIMESPEC(&atv, &ats);
213 		ts = &ats;
214 	}
215 
216 	return selcommon(retval, SCARG(uap, nd), SCARG(uap, in),
217 	    SCARG(uap, ou), SCARG(uap, ex), ts, NULL);
218 }
219 
220 /*
221  * sel_do_scan: common code to perform the scan on descriptors.
222  */
223 static int
224 sel_do_scan(const int op, void *fds, const int nf, const size_t ni,
225     struct timespec *ts, sigset_t *mask, register_t *retval)
226 {
227 	lwp_t		* const l = curlwp;
228 	selcluster_t	*sc;
229 	kmutex_t	*lock;
230 	struct timespec	sleepts;
231 	int		error, timo;
232 
233 	timo = 0;
234 	if (ts && inittimeleft(ts, &sleepts) == -1) {
235 		return EINVAL;
236 	}
237 
238 	if (__predict_false(mask))
239 		sigsuspendsetup(l, mask);
240 
241 	sc = curcpu()->ci_data.cpu_selcluster;
242 	lock = sc->sc_lock;
243 	l->l_selcluster = sc;
244 	if (op == SELOP_SELECT) {
245 		l->l_selbits = fds;
246 		l->l_selni = ni;
247 	} else {
248 		l->l_selbits = NULL;
249 	}
250 
251 	for (;;) {
252 		int ncoll;
253 
254 		SLIST_INIT(&l->l_selwait);
255 		l->l_selret = 0;
256 
257 		/*
258 		 * No need to lock.  If this is overwritten by another value
259 		 * while scanning, we will retry below.  We only need to see
260 		 * exact state from the descriptors that we are about to poll,
261 		 * and lock activity resulting from fo_poll is enough to
262 		 * provide an up to date value for new polling activity.
263 		 */
264 		l->l_selflag = SEL_SCANNING;
265 		ncoll = sc->sc_ncoll;
266 
267 		if (op == SELOP_SELECT) {
268 			error = selscan((char *)fds, nf, ni, retval);
269 		} else {
270 			error = pollscan((struct pollfd *)fds, nf, retval);
271 		}
272 		if (error || *retval)
273 			break;
274 		if (ts && (timo = gettimeleft(ts, &sleepts)) <= 0)
275 			break;
276 		/*
277 		 * Acquire the lock and perform the (re)checks.  Note, if
278 		 * collision has occured, then our state does not matter,
279 		 * as we must perform re-scan.  Therefore, check it first.
280 		 */
281 state_check:
282 		mutex_spin_enter(lock);
283 		if (__predict_false(sc->sc_ncoll != ncoll)) {
284 			/* Collision: perform re-scan. */
285 			mutex_spin_exit(lock);
286 			selclear();
287 			continue;
288 		}
289 		if (__predict_true(l->l_selflag == SEL_EVENT)) {
290 			/* Events occured, they are set directly. */
291 			mutex_spin_exit(lock);
292 			break;
293 		}
294 		if (__predict_true(l->l_selflag == SEL_RESET)) {
295 			/* Events occured, but re-scan is requested. */
296 			mutex_spin_exit(lock);
297 			selclear();
298 			continue;
299 		}
300 		/* Nothing happen, therefore - sleep. */
301 		l->l_selflag = SEL_BLOCKING;
302 		l->l_kpriority = true;
303 		sleepq_enter(&sc->sc_sleepq, l, lock);
304 		sleepq_enqueue(&sc->sc_sleepq, sc, "select", &select_sobj);
305 		error = sleepq_block(timo, true);
306 		if (error != 0) {
307 			break;
308 		}
309 		/* Awoken: need to check the state. */
310 		goto state_check;
311 	}
312 	selclear();
313 
314 	/* Add direct events if any. */
315 	if (l->l_selflag == SEL_EVENT) {
316 		KASSERT(l->l_selret != 0);
317 		*retval += l->l_selret;
318 	}
319 
320 	if (__predict_false(mask))
321 		sigsuspendteardown(l);
322 
323 	/* select and poll are not restarted after signals... */
324 	if (error == ERESTART)
325 		return EINTR;
326 	if (error == EWOULDBLOCK)
327 		return 0;
328 	return error;
329 }
330 
331 int
332 selcommon(register_t *retval, int nd, fd_set *u_in, fd_set *u_ou,
333     fd_set *u_ex, struct timespec *ts, sigset_t *mask)
334 {
335 	char		smallbits[howmany(FD_SETSIZE, NFDBITS) *
336 			    sizeof(fd_mask) * 6];
337 	char 		*bits;
338 	int		error, nf;
339 	size_t		ni;
340 
341 	if (nd < 0)
342 		return (EINVAL);
343 	nf = curlwp->l_fd->fd_dt->dt_nfiles;
344 	if (nd > nf) {
345 		/* forgiving; slightly wrong */
346 		nd = nf;
347 	}
348 	ni = howmany(nd, NFDBITS) * sizeof(fd_mask);
349 	if (ni * 6 > sizeof(smallbits))
350 		bits = kmem_alloc(ni * 6, KM_SLEEP);
351 	else
352 		bits = smallbits;
353 
354 #define	getbits(name, x)						\
355 	if (u_ ## name) {						\
356 		error = copyin(u_ ## name, bits + ni * x, ni);		\
357 		if (error)						\
358 			goto fail;					\
359 	} else								\
360 		memset(bits + ni * x, 0, ni);
361 	getbits(in, 0);
362 	getbits(ou, 1);
363 	getbits(ex, 2);
364 #undef	getbits
365 
366 	error = sel_do_scan(SELOP_SELECT, bits, nd, ni, ts, mask, retval);
367 	if (error == 0 && u_in != NULL)
368 		error = copyout(bits + ni * 3, u_in, ni);
369 	if (error == 0 && u_ou != NULL)
370 		error = copyout(bits + ni * 4, u_ou, ni);
371 	if (error == 0 && u_ex != NULL)
372 		error = copyout(bits + ni * 5, u_ex, ni);
373  fail:
374 	if (bits != smallbits)
375 		kmem_free(bits, ni * 6);
376 	return (error);
377 }
378 
379 static inline int
380 selscan(char *bits, const int nfd, const size_t ni, register_t *retval)
381 {
382 	fd_mask *ibitp, *obitp;
383 	int msk, i, j, fd, n;
384 	file_t *fp;
385 
386 	ibitp = (fd_mask *)(bits + ni * 0);
387 	obitp = (fd_mask *)(bits + ni * 3);
388 	n = 0;
389 
390 	memset(obitp, 0, ni * 3);
391 	for (msk = 0; msk < 3; msk++) {
392 		for (i = 0; i < nfd; i += NFDBITS) {
393 			fd_mask ibits, obits;
394 
395 			ibits = *ibitp;
396 			obits = 0;
397 			while ((j = ffs(ibits)) && (fd = i + --j) < nfd) {
398 				ibits &= ~(1U << j);
399 				if ((fp = fd_getfile(fd)) == NULL)
400 					return (EBADF);
401 				/*
402 				 * Setup an argument to selrecord(), which is
403 				 * a file descriptor number.
404 				 */
405 				curlwp->l_selrec = fd;
406 				if ((*fp->f_ops->fo_poll)(fp, sel_flag[msk])) {
407 					obits |= (1U << j);
408 					n++;
409 				}
410 				fd_putfile(fd);
411 			}
412 			if (obits != 0) {
413 				if (direct_select) {
414 					kmutex_t *lock;
415 					lock = curlwp->l_selcluster->sc_lock;
416 					mutex_spin_enter(lock);
417 					*obitp |= obits;
418 					mutex_spin_exit(lock);
419 				} else {
420 					*obitp |= obits;
421 				}
422 			}
423 			ibitp++;
424 			obitp++;
425 		}
426 	}
427 	*retval = n;
428 	return (0);
429 }
430 
431 /*
432  * Poll system call.
433  */
434 int
435 sys_poll(struct lwp *l, const struct sys_poll_args *uap, register_t *retval)
436 {
437 	/* {
438 		syscallarg(struct pollfd *)	fds;
439 		syscallarg(u_int)		nfds;
440 		syscallarg(int)			timeout;
441 	} */
442 	struct timespec	ats, *ts = NULL;
443 
444 	if (SCARG(uap, timeout) != INFTIM) {
445 		ats.tv_sec = SCARG(uap, timeout) / 1000;
446 		ats.tv_nsec = (SCARG(uap, timeout) % 1000) * 1000000;
447 		ts = &ats;
448 	}
449 
450 	return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, NULL);
451 }
452 
453 /*
454  * Poll system call.
455  */
456 int
457 sys___pollts50(struct lwp *l, const struct sys___pollts50_args *uap,
458     register_t *retval)
459 {
460 	/* {
461 		syscallarg(struct pollfd *)		fds;
462 		syscallarg(u_int)			nfds;
463 		syscallarg(const struct timespec *)	ts;
464 		syscallarg(const sigset_t *)		mask;
465 	} */
466 	struct timespec	ats, *ts = NULL;
467 	sigset_t	amask, *mask = NULL;
468 	int		error;
469 
470 	if (SCARG(uap, ts)) {
471 		error = copyin(SCARG(uap, ts), &ats, sizeof(ats));
472 		if (error)
473 			return error;
474 		ts = &ats;
475 	}
476 	if (SCARG(uap, mask)) {
477 		error = copyin(SCARG(uap, mask), &amask, sizeof(amask));
478 		if (error)
479 			return error;
480 		mask = &amask;
481 	}
482 
483 	return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, mask);
484 }
485 
486 int
487 pollcommon(register_t *retval, struct pollfd *u_fds, u_int nfds,
488     struct timespec *ts, sigset_t *mask)
489 {
490 	struct pollfd	smallfds[32];
491 	struct pollfd	*fds;
492 	int		error;
493 	size_t		ni;
494 
495 	if (nfds > curlwp->l_proc->p_rlimit[RLIMIT_NOFILE].rlim_max + 1000) {
496 		/*
497 		 * Prevent userland from causing over-allocation.
498 		 * Raising the default limit too high can still cause
499 		 * a lot of memory to be allocated, but this also means
500 		 * that the file descriptor array will also be large.
501 		 *
502 		 * To reduce the memory requirements here, we could
503 		 * process the 'fds' array in chunks, but that
504 		 * is a lot of code that isn't normally useful.
505 		 * (Or just move the copyin/out into pollscan().)
506 		 *
507 		 * Historically the code silently truncated 'fds' to
508 		 * dt_nfiles entries - but that does cause issues.
509 		 *
510 		 * Using the max limit equivalent to sysctl
511 		 * kern.maxfiles is the moral equivalent of OPEN_MAX
512 		 * as specified by POSIX.
513 		 *
514 		 * We add a slop of 1000 in case the resource limit was
515 		 * changed after opening descriptors or the same descriptor
516 		 * was specified more than once.
517 		 */
518 		return EINVAL;
519 	}
520 	ni = nfds * sizeof(struct pollfd);
521 	if (ni > sizeof(smallfds))
522 		fds = kmem_alloc(ni, KM_SLEEP);
523 	else
524 		fds = smallfds;
525 
526 	error = copyin(u_fds, fds, ni);
527 	if (error)
528 		goto fail;
529 
530 	error = sel_do_scan(SELOP_POLL, fds, nfds, ni, ts, mask, retval);
531 	if (error == 0)
532 		error = copyout(fds, u_fds, ni);
533  fail:
534 	if (fds != smallfds)
535 		kmem_free(fds, ni);
536 	return (error);
537 }
538 
539 static inline int
540 pollscan(struct pollfd *fds, const int nfd, register_t *retval)
541 {
542 	file_t *fp;
543 	int i, n = 0, revents;
544 
545 	for (i = 0; i < nfd; i++, fds++) {
546 		fds->revents = 0;
547 		if (fds->fd < 0) {
548 			revents = 0;
549 		} else if ((fp = fd_getfile(fds->fd)) == NULL) {
550 			revents = POLLNVAL;
551 		} else {
552 			/*
553 			 * Perform poll: registers select request or returns
554 			 * the events which are set.  Setup an argument for
555 			 * selrecord(), which is a pointer to struct pollfd.
556 			 */
557 			curlwp->l_selrec = (uintptr_t)fds;
558 			revents = (*fp->f_ops->fo_poll)(fp,
559 			    fds->events | POLLERR | POLLHUP);
560 			fd_putfile(fds->fd);
561 		}
562 		if (revents) {
563 			fds->revents = revents;
564 			n++;
565 		}
566 	}
567 	*retval = n;
568 	return (0);
569 }
570 
571 int
572 seltrue(dev_t dev, int events, lwp_t *l)
573 {
574 
575 	return (events & (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM));
576 }
577 
578 /*
579  * Record a select request.  Concurrency issues:
580  *
581  * The caller holds the same lock across calls to selrecord() and
582  * selnotify(), so we don't need to consider a concurrent wakeup
583  * while in this routine.
584  *
585  * The only activity we need to guard against is selclear(), called by
586  * another thread that is exiting sel_do_scan().
587  * `sel_lwp' can only become non-NULL while the caller's lock is held,
588  * so it cannot become non-NULL due to a change made by another thread
589  * while we are in this routine.  It can only become _NULL_ due to a
590  * call to selclear().
591  *
592  * If it is non-NULL and != selector there is the potential for
593  * selclear() to be called by another thread.  If either of those
594  * conditions are true, we're not interested in touching the `named
595  * waiter' part of the selinfo record because we need to record a
596  * collision.  Hence there is no need for additional locking in this
597  * routine.
598  */
599 void
600 selrecord(lwp_t *selector, struct selinfo *sip)
601 {
602 	selcluster_t *sc;
603 	lwp_t *other;
604 
605 	KASSERT(selector == curlwp);
606 
607 	sc = selector->l_selcluster;
608 	other = sip->sel_lwp;
609 
610 	if (other == selector) {
611 		/* 1. We (selector) already claimed to be the first LWP. */
612 		KASSERT(sip->sel_cluster == sc);
613 	} else if (other == NULL) {
614 		/*
615 		 * 2. No first LWP, therefore we (selector) are the first.
616 		 *
617 		 * There may be unnamed waiters (collisions).  Issue a memory
618 		 * barrier to ensure that we access sel_lwp (above) before
619 		 * other fields - this guards against a call to selclear().
620 		 */
621 		membar_enter();
622 		sip->sel_lwp = selector;
623 		SLIST_INSERT_HEAD(&selector->l_selwait, sip, sel_chain);
624 		/* Copy the argument, which is for selnotify(). */
625 		sip->sel_fdinfo = selector->l_selrec;
626 		/* Replace selinfo's lock with the chosen cluster's lock. */
627 		sip->sel_cluster = sc;
628 	} else {
629 		/* 3. Multiple waiters: record a collision. */
630 		sip->sel_collision |= sc->sc_mask;
631 		KASSERT(sip->sel_cluster != NULL);
632 	}
633 }
634 
635 /*
636  * sel_setevents: a helper function for selnotify(), to set the events
637  * for LWP sleeping in selcommon() or pollcommon().
638  */
639 static inline bool
640 sel_setevents(lwp_t *l, struct selinfo *sip, const int events)
641 {
642 	const int oflag = l->l_selflag;
643 	int ret = 0;
644 
645 	/*
646 	 * If we require re-scan or it was required by somebody else,
647 	 * then just (re)set SEL_RESET and return.
648 	 */
649 	if (__predict_false(events == 0 || oflag == SEL_RESET)) {
650 		l->l_selflag = SEL_RESET;
651 		return true;
652 	}
653 	/*
654 	 * Direct set.  Note: select state of LWP is locked.  First,
655 	 * determine whether it is selcommon() or pollcommon().
656 	 */
657 	if (l->l_selbits != NULL) {
658 		const size_t ni = l->l_selni;
659 		fd_mask *fds = (fd_mask *)l->l_selbits;
660 		fd_mask *ofds = (fd_mask *)((char *)fds + ni * 3);
661 		const int fd = sip->sel_fdinfo, fbit = 1 << (fd & __NFDMASK);
662 		const int idx = fd >> __NFDSHIFT;
663 		int n;
664 
665 		for (n = 0; n < 3; n++) {
666 			if ((fds[idx] & fbit) != 0 &&
667 			    (ofds[idx] & fbit) == 0 &&
668 			    (sel_flag[n] & events)) {
669 				ofds[idx] |= fbit;
670 				ret++;
671 			}
672 			fds = (fd_mask *)((char *)fds + ni);
673 			ofds = (fd_mask *)((char *)ofds + ni);
674 		}
675 	} else {
676 		struct pollfd *pfd = (void *)sip->sel_fdinfo;
677 		int revents = events & (pfd->events | POLLERR | POLLHUP);
678 
679 		if (revents) {
680 			if (pfd->revents == 0)
681 				ret = 1;
682 			pfd->revents |= revents;
683 		}
684 	}
685 	/* Check whether there are any events to return. */
686 	if (!ret) {
687 		return false;
688 	}
689 	/* Indicate direct set and note the event (cluster lock is held). */
690 	l->l_selflag = SEL_EVENT;
691 	l->l_selret += ret;
692 	return true;
693 }
694 
695 /*
696  * Do a wakeup when a selectable event occurs.  Concurrency issues:
697  *
698  * As per selrecord(), the caller's object lock is held.  If there
699  * is a named waiter, we must acquire the associated selcluster's lock
700  * in order to synchronize with selclear() and pollers going to sleep
701  * in sel_do_scan().
702  *
703  * sip->sel_cluser cannot change at this point, as it is only changed
704  * in selrecord(), and concurrent calls to selrecord() are locked
705  * out by the caller.
706  */
707 void
708 selnotify(struct selinfo *sip, int events, long knhint)
709 {
710 	selcluster_t *sc;
711 	uint32_t mask;
712 	int index, oflag;
713 	lwp_t *l;
714 	kmutex_t *lock;
715 
716 	KNOTE(&sip->sel_klist, knhint);
717 
718 	if (sip->sel_lwp != NULL) {
719 		/* One named LWP is waiting. */
720 		sc = sip->sel_cluster;
721 		lock = sc->sc_lock;
722 		mutex_spin_enter(lock);
723 		/* Still there? */
724 		if (sip->sel_lwp != NULL) {
725 			/*
726 			 * Set the events for our LWP and indicate that.
727 			 * Otherwise, request for a full re-scan.
728 			 */
729 			l = sip->sel_lwp;
730 			oflag = l->l_selflag;
731 
732 			if (!direct_select) {
733 				l->l_selflag = SEL_RESET;
734 			} else if (!sel_setevents(l, sip, events)) {
735 				/* No events to return. */
736 				mutex_spin_exit(lock);
737 				return;
738 			}
739 
740 			/*
741 			 * If thread is sleeping, wake it up.  If it's not
742 			 * yet asleep, it will notice the change in state
743 			 * and will re-poll the descriptors.
744 			 */
745 			if (oflag == SEL_BLOCKING && l->l_mutex == lock) {
746 				KASSERT(l->l_wchan == sc);
747 				sleepq_unsleep(l, false);
748 			}
749 		}
750 		mutex_spin_exit(lock);
751 	}
752 
753 	if ((mask = sip->sel_collision) != 0) {
754 		/*
755 		 * There was a collision (multiple waiters): we must
756 		 * inform all potentially interested waiters.
757 		 */
758 		sip->sel_collision = 0;
759 		do {
760 			index = ffs(mask) - 1;
761 			mask &= ~(1U << index);
762 			sc = selcluster[index];
763 			lock = sc->sc_lock;
764 			mutex_spin_enter(lock);
765 			sc->sc_ncoll++;
766 			sleepq_wake(&sc->sc_sleepq, sc, (u_int)-1, lock);
767 		} while (__predict_false(mask != 0));
768 	}
769 }
770 
771 /*
772  * Remove an LWP from all objects that it is waiting for.  Concurrency
773  * issues:
774  *
775  * The object owner's (e.g. device driver) lock is not held here.  Calls
776  * can be made to selrecord() and we do not synchronize against those
777  * directly using locks.  However, we use `sel_lwp' to lock out changes.
778  * Before clearing it we must use memory barriers to ensure that we can
779  * safely traverse the list of selinfo records.
780  */
781 static void
782 selclear(void)
783 {
784 	struct selinfo *sip, *next;
785 	selcluster_t *sc;
786 	lwp_t *l;
787 	kmutex_t *lock;
788 
789 	l = curlwp;
790 	sc = l->l_selcluster;
791 	lock = sc->sc_lock;
792 
793 	mutex_spin_enter(lock);
794 	for (sip = SLIST_FIRST(&l->l_selwait); sip != NULL; sip = next) {
795 		KASSERT(sip->sel_lwp == l);
796 		KASSERT(sip->sel_cluster == l->l_selcluster);
797 
798 		/*
799 		 * Read link to next selinfo record, if any.
800 		 * It's no longer safe to touch `sip' after clearing
801 		 * `sel_lwp', so ensure that the read of `sel_chain'
802 		 * completes before the clearing of sel_lwp becomes
803 		 * globally visible.
804 		 */
805 		next = SLIST_NEXT(sip, sel_chain);
806 		membar_exit();
807 		/* Release the record for another named waiter to use. */
808 		sip->sel_lwp = NULL;
809 	}
810 	mutex_spin_exit(lock);
811 }
812 
813 /*
814  * Initialize the select/poll system calls.  Called once for each
815  * CPU in the system, as they are attached.
816  */
817 void
818 selsysinit(struct cpu_info *ci)
819 {
820 	selcluster_t *sc;
821 	u_int index;
822 
823 	/* If already a cluster in place for this bit, re-use. */
824 	index = cpu_index(ci) & SELCLUSTERMASK;
825 	sc = selcluster[index];
826 	if (sc == NULL) {
827 		sc = kmem_alloc(roundup2(sizeof(selcluster_t),
828 		    coherency_unit) + coherency_unit, KM_SLEEP);
829 		sc = (void *)roundup2((uintptr_t)sc, coherency_unit);
830 		sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SCHED);
831 		sleepq_init(&sc->sc_sleepq);
832 		sc->sc_ncoll = 0;
833 		sc->sc_mask = __BIT(index);
834 		selcluster[index] = sc;
835 	}
836 	ci->ci_data.cpu_selcluster = sc;
837 }
838 
839 /*
840  * Initialize a selinfo record.
841  */
842 void
843 selinit(struct selinfo *sip)
844 {
845 
846 	memset(sip, 0, sizeof(*sip));
847 }
848 
849 /*
850  * Destroy a selinfo record.  The owning object must not gain new
851  * references while this is in progress: all activity on the record
852  * must be stopped.
853  *
854  * Concurrency issues: we only need guard against a call to selclear()
855  * by a thread exiting sel_do_scan().  The caller has prevented further
856  * references being made to the selinfo record via selrecord(), and it
857  * will not call selnotify() again.
858  */
859 void
860 seldestroy(struct selinfo *sip)
861 {
862 	selcluster_t *sc;
863 	kmutex_t *lock;
864 	lwp_t *l;
865 
866 	if (sip->sel_lwp == NULL)
867 		return;
868 
869 	/*
870 	 * Lock out selclear().  The selcluster pointer can't change while
871 	 * we are here since it is only ever changed in selrecord(),
872 	 * and that will not be entered again for this record because
873 	 * it is dying.
874 	 */
875 	KASSERT(sip->sel_cluster != NULL);
876 	sc = sip->sel_cluster;
877 	lock = sc->sc_lock;
878 	mutex_spin_enter(lock);
879 	if ((l = sip->sel_lwp) != NULL) {
880 		/*
881 		 * This should rarely happen, so although SLIST_REMOVE()
882 		 * is slow, using it here is not a problem.
883 		 */
884 		KASSERT(l->l_selcluster == sc);
885 		SLIST_REMOVE(&l->l_selwait, sip, selinfo, sel_chain);
886 		sip->sel_lwp = NULL;
887 	}
888 	mutex_spin_exit(lock);
889 }
890 
891 /*
892  * System control nodes.
893  */
894 SYSCTL_SETUP(sysctl_select_setup, "sysctl select setup")
895 {
896 
897 	sysctl_createv(clog, 0, NULL, NULL,
898 		CTLFLAG_PERMANENT | CTLFLAG_READWRITE,
899 		CTLTYPE_INT, "direct_select",
900 		SYSCTL_DESCR("Enable/disable direct select (for testing)"),
901 		NULL, 0, &direct_select, 0,
902 		CTL_KERN, CTL_CREATE, CTL_EOL);
903 }
904