xref: /netbsd-src/sys/kern/sys_select.c (revision 8e33eff89e26cf71871ead62f0d5063e1313c33a)
1 /*	$NetBSD: sys_select.c,v 1.66 2023/10/15 10:29:34 riastradh Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007, 2008, 2009, 2010, 2019, 2020, 2023
5  *     The NetBSD Foundation, Inc.
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by Andrew Doran and Mindaugas Rasiukevicius.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30  * POSSIBILITY OF SUCH DAMAGE.
31  */
32 
33 /*
34  * Copyright (c) 1982, 1986, 1989, 1993
35  *	The Regents of the University of California.  All rights reserved.
36  * (c) UNIX System Laboratories, Inc.
37  * All or some portions of this file are derived from material licensed
38  * to the University of California by American Telephone and Telegraph
39  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
40  * the permission of UNIX System Laboratories, Inc.
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *	@(#)sys_generic.c	8.9 (Berkeley) 2/14/95
67  */
68 
69 /*
70  * System calls of synchronous I/O multiplexing subsystem.
71  *
72  * Locking
73  *
74  * Two locks are used: <object-lock> and selcluster_t::sc_lock.
75  *
76  * The <object-lock> might be a device driver or another subsystem, e.g.
77  * socket or pipe.  This lock is not exported, and thus invisible to this
78  * subsystem.  Mainly, synchronisation between selrecord() and selnotify()
79  * routines depends on this lock, as it will be described in the comments.
80  *
81  * Lock order
82  *
83  *	<object-lock> ->
84  *		selcluster_t::sc_lock
85  */
86 
87 #include <sys/cdefs.h>
88 __KERNEL_RCSID(0, "$NetBSD: sys_select.c,v 1.66 2023/10/15 10:29:34 riastradh Exp $");
89 
90 #include <sys/param.h>
91 
92 #include <sys/atomic.h>
93 #include <sys/bitops.h>
94 #include <sys/cpu.h>
95 #include <sys/file.h>
96 #include <sys/filedesc.h>
97 #include <sys/kernel.h>
98 #include <sys/lwp.h>
99 #include <sys/mount.h>
100 #include <sys/poll.h>
101 #include <sys/proc.h>
102 #include <sys/signalvar.h>
103 #include <sys/sleepq.h>
104 #include <sys/socketvar.h>
105 #include <sys/socketvar.h>
106 #include <sys/syncobj.h>
107 #include <sys/syscallargs.h>
108 #include <sys/sysctl.h>
109 #include <sys/systm.h>
110 #include <sys/uio.h>
111 
112 /* Flags for lwp::l_selflag. */
113 #define	SEL_RESET	0	/* awoken, interrupted, or not yet polling */
114 #define	SEL_SCANNING	1	/* polling descriptors */
115 #define	SEL_BLOCKING	2	/* blocking and waiting for event */
116 #define	SEL_EVENT	3	/* interrupted, events set directly */
117 
118 /*
119  * Per-cluster state for select()/poll().  For a system with fewer
120  * than 64 CPUs, this gives us per-CPU clusters.
121  */
122 #define	SELCLUSTERS	64
123 #define	SELCLUSTERMASK	(SELCLUSTERS - 1)
124 
125 typedef struct selcluster {
126 	kmutex_t	*sc_lock;
127 	sleepq_t	sc_sleepq;
128 	uint64_t	sc_mask;
129 	int		sc_ncoll;
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 /*
143  * LWPs are woken using the sleep queue only due to a collision, the case
144  * with the maximum Suck Factor.  Save the cost of sorting for named waiters
145  * by inserting in LIFO order.  In the future it would be preferable to not
146  * enqueue LWPs at all, unless subject to a collision.
147  */
148 syncobj_t select_sobj = {
149 	.sobj_name	= "select",
150 	.sobj_flag	= SOBJ_SLEEPQ_LIFO,
151 	.sobj_boostpri  = PRI_KERNEL,
152 	.sobj_unsleep	= sleepq_unsleep,
153 	.sobj_changepri	= sleepq_changepri,
154 	.sobj_lendpri	= sleepq_lendpri,
155 	.sobj_owner	= syncobj_noowner,
156 };
157 
158 static selcluster_t	*selcluster[SELCLUSTERS] __read_mostly;
159 static int		direct_select __read_mostly = 0;
160 
161 /* Operations: either select() or poll(). */
162 const char		selop_select[] = "select";
163 const char		selop_poll[] = "poll";
164 
165 /*
166  * Select system call.
167  */
168 int
169 sys___pselect50(struct lwp *l, const struct sys___pselect50_args *uap,
170     register_t *retval)
171 {
172 	/* {
173 		syscallarg(int)				nd;
174 		syscallarg(fd_set *)			in;
175 		syscallarg(fd_set *)			ou;
176 		syscallarg(fd_set *)			ex;
177 		syscallarg(const struct timespec *)	ts;
178 		syscallarg(sigset_t *)			mask;
179 	} */
180 	struct timespec	ats, *ts = NULL;
181 	sigset_t	amask, *mask = NULL;
182 	int		error;
183 
184 	if (SCARG(uap, ts)) {
185 		error = copyin(SCARG(uap, ts), &ats, sizeof(ats));
186 		if (error)
187 			return error;
188 		ts = &ats;
189 	}
190 	if (SCARG(uap, mask) != NULL) {
191 		error = copyin(SCARG(uap, mask), &amask, sizeof(amask));
192 		if (error)
193 			return error;
194 		mask = &amask;
195 	}
196 
197 	return selcommon(retval, SCARG(uap, nd), SCARG(uap, in),
198 	    SCARG(uap, ou), SCARG(uap, ex), ts, mask);
199 }
200 
201 int
202 sys___select50(struct lwp *l, const struct sys___select50_args *uap,
203     register_t *retval)
204 {
205 	/* {
206 		syscallarg(int)			nd;
207 		syscallarg(fd_set *)		in;
208 		syscallarg(fd_set *)		ou;
209 		syscallarg(fd_set *)		ex;
210 		syscallarg(struct timeval *)	tv;
211 	} */
212 	struct timeval atv;
213 	struct timespec ats, *ts = NULL;
214 	int error;
215 
216 	if (SCARG(uap, tv)) {
217 		error = copyin(SCARG(uap, tv), (void *)&atv, sizeof(atv));
218 		if (error)
219 			return error;
220 
221 		if (atv.tv_usec < 0 || atv.tv_usec >= 1000000)
222 			return EINVAL;
223 
224 		TIMEVAL_TO_TIMESPEC(&atv, &ats);
225 		ts = &ats;
226 	}
227 
228 	return selcommon(retval, SCARG(uap, nd), SCARG(uap, in),
229 	    SCARG(uap, ou), SCARG(uap, ex), ts, NULL);
230 }
231 
232 /*
233  * sel_do_scan: common code to perform the scan on descriptors.
234  */
235 static int
236 sel_do_scan(const char *opname, void *fds, const int nf, const size_t ni,
237     struct timespec *ts, sigset_t *mask, register_t *retval)
238 {
239 	lwp_t		* const l = curlwp;
240 	selcluster_t	*sc;
241 	kmutex_t	*lock;
242 	struct timespec	sleepts;
243 	int		error, timo;
244 
245 	timo = 0;
246 	if (ts && inittimeleft(ts, &sleepts) == -1) {
247 		return EINVAL;
248 	}
249 
250 	if (__predict_false(mask))
251 		sigsuspendsetup(l, mask);
252 
253 	/*
254 	 * We may context switch during or at any time after picking a CPU
255 	 * and cluster to associate with, but it doesn't matter.  In the
256 	 * unlikely event we migrate elsewhere all we risk is a little lock
257 	 * contention; correctness is not sacrificed.
258 	 */
259 	sc = curcpu()->ci_data.cpu_selcluster;
260 	lock = sc->sc_lock;
261 	l->l_selcluster = sc;
262 
263 	if (opname == selop_select) {
264 		l->l_selbits = fds;
265 		l->l_selni = ni;
266 	} else {
267 		l->l_selbits = NULL;
268 	}
269 
270 	for (;;) {
271 		int ncoll;
272 
273 		SLIST_INIT(&l->l_selwait);
274 		l->l_selret = 0;
275 
276 		/*
277 		 * No need to lock.  If this is overwritten by another value
278 		 * while scanning, we will retry below.  We only need to see
279 		 * exact state from the descriptors that we are about to poll,
280 		 * and lock activity resulting from fo_poll is enough to
281 		 * provide an up to date value for new polling activity.
282 		 */
283 		if (ts && (ts->tv_sec | ts->tv_nsec | direct_select) == 0) {
284 			/* Non-blocking: no need for selrecord()/selclear() */
285 			l->l_selflag = SEL_RESET;
286 		} else {
287 			l->l_selflag = SEL_SCANNING;
288 		}
289 		ncoll = sc->sc_ncoll;
290 		membar_release();
291 
292 		if (opname == selop_select) {
293 			error = selscan((char *)fds, nf, ni, retval);
294 		} else {
295 			error = pollscan((struct pollfd *)fds, nf, retval);
296 		}
297 		if (error || *retval)
298 			break;
299 		if (ts && (timo = gettimeleft(ts, &sleepts)) <= 0)
300 			break;
301 		/*
302 		 * Acquire the lock and perform the (re)checks.  Note, if
303 		 * collision has occurred, then our state does not matter,
304 		 * as we must perform re-scan.  Therefore, check it first.
305 		 */
306 state_check:
307 		mutex_spin_enter(lock);
308 		if (__predict_false(sc->sc_ncoll != ncoll)) {
309 			/* Collision: perform re-scan. */
310 			mutex_spin_exit(lock);
311 			selclear();
312 			continue;
313 		}
314 		if (__predict_true(l->l_selflag == SEL_EVENT)) {
315 			/* Events occurred, they are set directly. */
316 			mutex_spin_exit(lock);
317 			break;
318 		}
319 		if (__predict_true(l->l_selflag == SEL_RESET)) {
320 			/* Events occurred, but re-scan is requested. */
321 			mutex_spin_exit(lock);
322 			selclear();
323 			continue;
324 		}
325 		/* Nothing happen, therefore - sleep. */
326 		l->l_selflag = SEL_BLOCKING;
327 		KASSERT(l->l_blcnt == 0);
328 		(void)sleepq_enter(&sc->sc_sleepq, l, lock);
329 		sleepq_enqueue(&sc->sc_sleepq, sc, opname, &select_sobj, true);
330 		error = sleepq_block(timo, true, &select_sobj, 0);
331 		if (error != 0) {
332 			break;
333 		}
334 		/* Awoken: need to check the state. */
335 		goto state_check;
336 	}
337 	selclear();
338 
339 	/* Add direct events if any. */
340 	if (l->l_selflag == SEL_EVENT) {
341 		KASSERT(l->l_selret != 0);
342 		*retval += l->l_selret;
343 	}
344 
345 	if (__predict_false(mask))
346 		sigsuspendteardown(l);
347 
348 	/* select and poll are not restarted after signals... */
349 	if (error == ERESTART)
350 		return EINTR;
351 	if (error == EWOULDBLOCK)
352 		return 0;
353 	return error;
354 }
355 
356 int
357 selcommon(register_t *retval, int nd, fd_set *u_in, fd_set *u_ou,
358     fd_set *u_ex, struct timespec *ts, sigset_t *mask)
359 {
360 	char		smallbits[howmany(FD_SETSIZE, NFDBITS) *
361 			    sizeof(fd_mask) * 6];
362 	char 		*bits;
363 	int		error, nf;
364 	size_t		ni;
365 
366 	if (nd < 0)
367 		return (EINVAL);
368 	nf = atomic_load_consume(&curlwp->l_fd->fd_dt)->dt_nfiles;
369 	if (nd > nf) {
370 		/* forgiving; slightly wrong */
371 		nd = nf;
372 	}
373 	ni = howmany(nd, NFDBITS) * sizeof(fd_mask);
374 	if (ni * 6 > sizeof(smallbits))
375 		bits = kmem_alloc(ni * 6, KM_SLEEP);
376 	else
377 		bits = smallbits;
378 
379 #define	getbits(name, x)						\
380 	if (u_ ## name) {						\
381 		error = copyin(u_ ## name, bits + ni * x, ni);		\
382 		if (error)						\
383 			goto fail;					\
384 	} else								\
385 		memset(bits + ni * x, 0, ni);
386 	getbits(in, 0);
387 	getbits(ou, 1);
388 	getbits(ex, 2);
389 #undef	getbits
390 
391 	error = sel_do_scan(selop_select, bits, nd, ni, ts, mask, retval);
392 	if (error == 0 && u_in != NULL)
393 		error = copyout(bits + ni * 3, u_in, ni);
394 	if (error == 0 && u_ou != NULL)
395 		error = copyout(bits + ni * 4, u_ou, ni);
396 	if (error == 0 && u_ex != NULL)
397 		error = copyout(bits + ni * 5, u_ex, ni);
398  fail:
399 	if (bits != smallbits)
400 		kmem_free(bits, ni * 6);
401 	return (error);
402 }
403 
404 static inline int
405 selscan(char *bits, const int nfd, const size_t ni, register_t *retval)
406 {
407 	fd_mask *ibitp, *obitp;
408 	int msk, i, j, fd, n;
409 	file_t *fp;
410 	lwp_t *l;
411 
412 	ibitp = (fd_mask *)(bits + ni * 0);
413 	obitp = (fd_mask *)(bits + ni * 3);
414 	n = 0;
415 	l = curlwp;
416 
417 	memset(obitp, 0, ni * 3);
418 	for (msk = 0; msk < 3; msk++) {
419 		for (i = 0; i < nfd; i += NFDBITS) {
420 			fd_mask ibits, obits;
421 
422 			ibits = *ibitp;
423 			obits = 0;
424 			while ((j = ffs(ibits)) && (fd = i + --j) < nfd) {
425 				ibits &= ~(1U << j);
426 				if ((fp = fd_getfile(fd)) == NULL)
427 					return (EBADF);
428 				/*
429 				 * Setup an argument to selrecord(), which is
430 				 * a file descriptor number.
431 				 */
432 				l->l_selrec = fd;
433 				if ((*fp->f_ops->fo_poll)(fp, sel_flag[msk])) {
434 					if (!direct_select) {
435 						/*
436 						 * Have events: do nothing in
437 						 * selrecord().
438 						 */
439 						l->l_selflag = SEL_RESET;
440 					}
441 					obits |= (1U << j);
442 					n++;
443 				}
444 				fd_putfile(fd);
445 			}
446 			if (obits != 0) {
447 				if (direct_select) {
448 					kmutex_t *lock;
449 					lock = l->l_selcluster->sc_lock;
450 					mutex_spin_enter(lock);
451 					*obitp |= obits;
452 					mutex_spin_exit(lock);
453 				} else {
454 					*obitp |= obits;
455 				}
456 			}
457 			ibitp++;
458 			obitp++;
459 		}
460 	}
461 	*retval = n;
462 	return (0);
463 }
464 
465 /*
466  * Poll system call.
467  */
468 int
469 sys_poll(struct lwp *l, const struct sys_poll_args *uap, register_t *retval)
470 {
471 	/* {
472 		syscallarg(struct pollfd *)	fds;
473 		syscallarg(u_int)		nfds;
474 		syscallarg(int)			timeout;
475 	} */
476 	struct timespec	ats, *ts = NULL;
477 
478 	if (SCARG(uap, timeout) != INFTIM) {
479 		ats.tv_sec = SCARG(uap, timeout) / 1000;
480 		ats.tv_nsec = (SCARG(uap, timeout) % 1000) * 1000000;
481 		ts = &ats;
482 	}
483 
484 	return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, NULL);
485 }
486 
487 /*
488  * Poll system call.
489  */
490 int
491 sys___pollts50(struct lwp *l, const struct sys___pollts50_args *uap,
492     register_t *retval)
493 {
494 	/* {
495 		syscallarg(struct pollfd *)		fds;
496 		syscallarg(u_int)			nfds;
497 		syscallarg(const struct timespec *)	ts;
498 		syscallarg(const sigset_t *)		mask;
499 	} */
500 	struct timespec	ats, *ts = NULL;
501 	sigset_t	amask, *mask = NULL;
502 	int		error;
503 
504 	if (SCARG(uap, ts)) {
505 		error = copyin(SCARG(uap, ts), &ats, sizeof(ats));
506 		if (error)
507 			return error;
508 		ts = &ats;
509 	}
510 	if (SCARG(uap, mask)) {
511 		error = copyin(SCARG(uap, mask), &amask, sizeof(amask));
512 		if (error)
513 			return error;
514 		mask = &amask;
515 	}
516 
517 	return pollcommon(retval, SCARG(uap, fds), SCARG(uap, nfds), ts, mask);
518 }
519 
520 int
521 pollcommon(register_t *retval, struct pollfd *u_fds, u_int nfds,
522     struct timespec *ts, sigset_t *mask)
523 {
524 	struct pollfd	smallfds[32];
525 	struct pollfd	*fds;
526 	int		error;
527 	size_t		ni;
528 
529 	if (nfds > curlwp->l_proc->p_rlimit[RLIMIT_NOFILE].rlim_max + 1000) {
530 		/*
531 		 * Prevent userland from causing over-allocation.
532 		 * Raising the default limit too high can still cause
533 		 * a lot of memory to be allocated, but this also means
534 		 * that the file descriptor array will also be large.
535 		 *
536 		 * To reduce the memory requirements here, we could
537 		 * process the 'fds' array in chunks, but that
538 		 * is a lot of code that isn't normally useful.
539 		 * (Or just move the copyin/out into pollscan().)
540 		 *
541 		 * Historically the code silently truncated 'fds' to
542 		 * dt_nfiles entries - but that does cause issues.
543 		 *
544 		 * Using the max limit equivalent to sysctl
545 		 * kern.maxfiles is the moral equivalent of OPEN_MAX
546 		 * as specified by POSIX.
547 		 *
548 		 * We add a slop of 1000 in case the resource limit was
549 		 * changed after opening descriptors or the same descriptor
550 		 * was specified more than once.
551 		 */
552 		return EINVAL;
553 	}
554 	ni = nfds * sizeof(struct pollfd);
555 	if (ni > sizeof(smallfds))
556 		fds = kmem_alloc(ni, KM_SLEEP);
557 	else
558 		fds = smallfds;
559 
560 	error = copyin(u_fds, fds, ni);
561 	if (error)
562 		goto fail;
563 
564 	error = sel_do_scan(selop_poll, fds, nfds, ni, ts, mask, retval);
565 	if (error == 0)
566 		error = copyout(fds, u_fds, ni);
567  fail:
568 	if (fds != smallfds)
569 		kmem_free(fds, ni);
570 	return (error);
571 }
572 
573 static inline int
574 pollscan(struct pollfd *fds, const int nfd, register_t *retval)
575 {
576 	file_t *fp;
577 	int i, n = 0, revents;
578 
579 	for (i = 0; i < nfd; i++, fds++) {
580 		fds->revents = 0;
581 		if (fds->fd < 0) {
582 			revents = 0;
583 		} else if ((fp = fd_getfile(fds->fd)) == NULL) {
584 			revents = POLLNVAL;
585 		} else {
586 			/*
587 			 * Perform poll: registers select request or returns
588 			 * the events which are set.  Setup an argument for
589 			 * selrecord(), which is a pointer to struct pollfd.
590 			 */
591 			curlwp->l_selrec = (uintptr_t)fds;
592 			revents = (*fp->f_ops->fo_poll)(fp,
593 			    fds->events | POLLERR | POLLHUP);
594 			fd_putfile(fds->fd);
595 		}
596 		if (revents) {
597 			if (!direct_select)  {
598 				/* Have events: do nothing in selrecord(). */
599 				curlwp->l_selflag = SEL_RESET;
600 			}
601 			fds->revents = revents;
602 			n++;
603 		}
604 	}
605 	*retval = n;
606 	return (0);
607 }
608 
609 int
610 seltrue(dev_t dev, int events, lwp_t *l)
611 {
612 
613 	return (events & (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM));
614 }
615 
616 /*
617  * Record a select request.  Concurrency issues:
618  *
619  * The caller holds the same lock across calls to selrecord() and
620  * selnotify(), so we don't need to consider a concurrent wakeup
621  * while in this routine.
622  *
623  * The only activity we need to guard against is selclear(), called by
624  * another thread that is exiting sel_do_scan().
625  * `sel_lwp' can only become non-NULL while the caller's lock is held,
626  * so it cannot become non-NULL due to a change made by another thread
627  * while we are in this routine.  It can only become _NULL_ due to a
628  * call to selclear().
629  *
630  * If it is non-NULL and != selector there is the potential for
631  * selclear() to be called by another thread.  If either of those
632  * conditions are true, we're not interested in touching the `named
633  * waiter' part of the selinfo record because we need to record a
634  * collision.  Hence there is no need for additional locking in this
635  * routine.
636  */
637 void
638 selrecord(lwp_t *selector, struct selinfo *sip)
639 {
640 	selcluster_t *sc;
641 	lwp_t *other;
642 
643 	KASSERT(selector == curlwp);
644 
645 	sc = selector->l_selcluster;
646 	other = sip->sel_lwp;
647 
648 	if (selector->l_selflag == SEL_RESET) {
649 		/* 0. We're not going to block - will poll again if needed. */
650 	} else if (other == selector) {
651 		/* 1. We (selector) already claimed to be the first LWP. */
652 		KASSERT(sip->sel_cluster == sc);
653 	} else if (other == NULL) {
654 		/*
655 		 * 2. No first LWP, therefore we (selector) are the first.
656 		 *
657 		 * There may be unnamed waiters (collisions).  Issue a memory
658 		 * barrier to ensure that we access sel_lwp (above) before
659 		 * other fields - this guards against a call to selclear().
660 		 */
661 		membar_acquire();
662 		sip->sel_lwp = selector;
663 		SLIST_INSERT_HEAD(&selector->l_selwait, sip, sel_chain);
664 		/* Copy the argument, which is for selnotify(). */
665 		sip->sel_fdinfo = selector->l_selrec;
666 		/* Replace selinfo's lock with the chosen cluster's lock. */
667 		sip->sel_cluster = sc;
668 	} else {
669 		/* 3. Multiple waiters: record a collision. */
670 		sip->sel_collision |= sc->sc_mask;
671 		KASSERT(sip->sel_cluster != NULL);
672 	}
673 }
674 
675 /*
676  * Record a knote.
677  *
678  * The caller holds the same lock as for selrecord().
679  */
680 void
681 selrecord_knote(struct selinfo *sip, struct knote *kn)
682 {
683 	klist_insert(&sip->sel_klist, kn);
684 }
685 
686 /*
687  * Remove a knote.
688  *
689  * The caller holds the same lock as for selrecord().
690  *
691  * Returns true if the last knote was removed and the list
692  * is now empty.
693  */
694 bool
695 selremove_knote(struct selinfo *sip, struct knote *kn)
696 {
697 	return klist_remove(&sip->sel_klist, kn);
698 }
699 
700 /*
701  * sel_setevents: a helper function for selnotify(), to set the events
702  * for LWP sleeping in selcommon() or pollcommon().
703  */
704 static inline bool
705 sel_setevents(lwp_t *l, struct selinfo *sip, const int events)
706 {
707 	const int oflag = l->l_selflag;
708 	int ret = 0;
709 
710 	/*
711 	 * If we require re-scan or it was required by somebody else,
712 	 * then just (re)set SEL_RESET and return.
713 	 */
714 	if (__predict_false(events == 0 || oflag == SEL_RESET)) {
715 		l->l_selflag = SEL_RESET;
716 		return true;
717 	}
718 	/*
719 	 * Direct set.  Note: select state of LWP is locked.  First,
720 	 * determine whether it is selcommon() or pollcommon().
721 	 */
722 	if (l->l_selbits != NULL) {
723 		const size_t ni = l->l_selni;
724 		fd_mask *fds = (fd_mask *)l->l_selbits;
725 		fd_mask *ofds = (fd_mask *)((char *)fds + ni * 3);
726 		const int fd = sip->sel_fdinfo, fbit = 1 << (fd & __NFDMASK);
727 		const int idx = fd >> __NFDSHIFT;
728 		int n;
729 
730 		for (n = 0; n < 3; n++) {
731 			if ((fds[idx] & fbit) != 0 &&
732 			    (ofds[idx] & fbit) == 0 &&
733 			    (sel_flag[n] & events)) {
734 				ofds[idx] |= fbit;
735 				ret++;
736 			}
737 			fds = (fd_mask *)((char *)fds + ni);
738 			ofds = (fd_mask *)((char *)ofds + ni);
739 		}
740 	} else {
741 		struct pollfd *pfd = (void *)sip->sel_fdinfo;
742 		int revents = events & (pfd->events | POLLERR | POLLHUP);
743 
744 		if (revents) {
745 			if (pfd->revents == 0)
746 				ret = 1;
747 			pfd->revents |= revents;
748 		}
749 	}
750 	/* Check whether there are any events to return. */
751 	if (!ret) {
752 		return false;
753 	}
754 	/* Indicate direct set and note the event (cluster lock is held). */
755 	l->l_selflag = SEL_EVENT;
756 	l->l_selret += ret;
757 	return true;
758 }
759 
760 /*
761  * Do a wakeup when a selectable event occurs.  Concurrency issues:
762  *
763  * As per selrecord(), the caller's object lock is held.  If there
764  * is a named waiter, we must acquire the associated selcluster's lock
765  * in order to synchronize with selclear() and pollers going to sleep
766  * in sel_do_scan().
767  *
768  * sip->sel_cluser cannot change at this point, as it is only changed
769  * in selrecord(), and concurrent calls to selrecord() are locked
770  * out by the caller.
771  */
772 void
773 selnotify(struct selinfo *sip, int events, long knhint)
774 {
775 	selcluster_t *sc;
776 	uint64_t mask;
777 	int index, oflag;
778 	lwp_t *l;
779 	kmutex_t *lock;
780 
781 	KNOTE(&sip->sel_klist, knhint);
782 
783 	if (sip->sel_lwp != NULL) {
784 		/* One named LWP is waiting. */
785 		sc = sip->sel_cluster;
786 		lock = sc->sc_lock;
787 		mutex_spin_enter(lock);
788 		/* Still there? */
789 		if (sip->sel_lwp != NULL) {
790 			/*
791 			 * Set the events for our LWP and indicate that.
792 			 * Otherwise, request for a full re-scan.
793 			 */
794 			l = sip->sel_lwp;
795 			oflag = l->l_selflag;
796 
797 			if (!direct_select) {
798 				l->l_selflag = SEL_RESET;
799 			} else if (!sel_setevents(l, sip, events)) {
800 				/* No events to return. */
801 				mutex_spin_exit(lock);
802 				return;
803 			}
804 
805 			/*
806 			 * If thread is sleeping, wake it up.  If it's not
807 			 * yet asleep, it will notice the change in state
808 			 * and will re-poll the descriptors.
809 			 */
810 			if (oflag == SEL_BLOCKING && l->l_mutex == lock) {
811 				KASSERT(l->l_wchan == sc);
812 				sleepq_remove(l->l_sleepq, l, true);
813 			}
814 		}
815 		mutex_spin_exit(lock);
816 	}
817 
818 	if ((mask = sip->sel_collision) != 0) {
819 		/*
820 		 * There was a collision (multiple waiters): we must
821 		 * inform all potentially interested waiters.
822 		 */
823 		sip->sel_collision = 0;
824 		do {
825 			index = ffs64(mask) - 1;
826 			mask ^= __BIT(index);
827 			sc = selcluster[index];
828 			lock = sc->sc_lock;
829 			mutex_spin_enter(lock);
830 			sc->sc_ncoll++;
831 			sleepq_wake(&sc->sc_sleepq, sc, (u_int)-1, lock);
832 		} while (__predict_false(mask != 0));
833 	}
834 }
835 
836 /*
837  * Remove an LWP from all objects that it is waiting for.  Concurrency
838  * issues:
839  *
840  * The object owner's (e.g. device driver) lock is not held here.  Calls
841  * can be made to selrecord() and we do not synchronize against those
842  * directly using locks.  However, we use `sel_lwp' to lock out changes.
843  * Before clearing it we must use memory barriers to ensure that we can
844  * safely traverse the list of selinfo records.
845  */
846 static void
847 selclear(void)
848 {
849 	struct selinfo *sip, *next;
850 	selcluster_t *sc;
851 	lwp_t *l;
852 	kmutex_t *lock;
853 
854 	l = curlwp;
855 	sc = l->l_selcluster;
856 	lock = sc->sc_lock;
857 
858 	/*
859 	 * If the request was non-blocking, or we found events on the first
860 	 * descriptor, there will be no need to clear anything - avoid
861 	 * taking the lock.
862 	 */
863 	if (SLIST_EMPTY(&l->l_selwait)) {
864 		return;
865 	}
866 
867 	mutex_spin_enter(lock);
868 	for (sip = SLIST_FIRST(&l->l_selwait); sip != NULL; sip = next) {
869 		KASSERT(sip->sel_lwp == l);
870 		KASSERT(sip->sel_cluster == l->l_selcluster);
871 
872 		/*
873 		 * Read link to next selinfo record, if any.
874 		 * It's no longer safe to touch `sip' after clearing
875 		 * `sel_lwp', so ensure that the read of `sel_chain'
876 		 * completes before the clearing of sel_lwp becomes
877 		 * globally visible.
878 		 */
879 		next = SLIST_NEXT(sip, sel_chain);
880 		/* Release the record for another named waiter to use. */
881 		atomic_store_release(&sip->sel_lwp, NULL);
882 	}
883 	mutex_spin_exit(lock);
884 }
885 
886 /*
887  * Initialize the select/poll system calls.  Called once for each
888  * CPU in the system, as they are attached.
889  */
890 void
891 selsysinit(struct cpu_info *ci)
892 {
893 	selcluster_t *sc;
894 	u_int index;
895 
896 	/* If already a cluster in place for this bit, re-use. */
897 	index = cpu_index(ci) & SELCLUSTERMASK;
898 	sc = selcluster[index];
899 	if (sc == NULL) {
900 		sc = kmem_alloc(roundup2(sizeof(selcluster_t),
901 		    coherency_unit) + coherency_unit, KM_SLEEP);
902 		sc = (void *)roundup2((uintptr_t)sc, coherency_unit);
903 		sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SCHED);
904 		sleepq_init(&sc->sc_sleepq);
905 		sc->sc_ncoll = 0;
906 		sc->sc_mask = __BIT(index);
907 		selcluster[index] = sc;
908 	}
909 	ci->ci_data.cpu_selcluster = sc;
910 }
911 
912 /*
913  * Initialize a selinfo record.
914  */
915 void
916 selinit(struct selinfo *sip)
917 {
918 
919 	memset(sip, 0, sizeof(*sip));
920 	klist_init(&sip->sel_klist);
921 }
922 
923 /*
924  * Destroy a selinfo record.  The owning object must not gain new
925  * references while this is in progress: all activity on the record
926  * must be stopped.
927  *
928  * Concurrency issues: we only need guard against a call to selclear()
929  * by a thread exiting sel_do_scan().  The caller has prevented further
930  * references being made to the selinfo record via selrecord(), and it
931  * will not call selnotify() again.
932  */
933 void
934 seldestroy(struct selinfo *sip)
935 {
936 	selcluster_t *sc;
937 	kmutex_t *lock;
938 	lwp_t *l;
939 
940 	klist_fini(&sip->sel_klist);
941 
942 	if (sip->sel_lwp == NULL)
943 		return;
944 
945 	/*
946 	 * Lock out selclear().  The selcluster pointer can't change while
947 	 * we are here since it is only ever changed in selrecord(),
948 	 * and that will not be entered again for this record because
949 	 * it is dying.
950 	 */
951 	KASSERT(sip->sel_cluster != NULL);
952 	sc = sip->sel_cluster;
953 	lock = sc->sc_lock;
954 	mutex_spin_enter(lock);
955 	if ((l = sip->sel_lwp) != NULL) {
956 		/*
957 		 * This should rarely happen, so although SLIST_REMOVE()
958 		 * is slow, using it here is not a problem.
959 		 */
960 		KASSERT(l->l_selcluster == sc);
961 		SLIST_REMOVE(&l->l_selwait, sip, selinfo, sel_chain);
962 		sip->sel_lwp = NULL;
963 	}
964 	mutex_spin_exit(lock);
965 }
966 
967 /*
968  * System control nodes.
969  */
970 SYSCTL_SETUP(sysctl_select_setup, "sysctl select setup")
971 {
972 
973 	sysctl_createv(clog, 0, NULL, NULL,
974 		CTLFLAG_PERMANENT | CTLFLAG_READWRITE,
975 		CTLTYPE_INT, "direct_select",
976 		SYSCTL_DESCR("Enable/disable direct select (for testing)"),
977 		NULL, 0, &direct_select, 0,
978 		CTL_KERN, CTL_CREATE, CTL_EOL);
979 }
980