xref: /illumos-gate/usr/src/uts/common/syscall/lwpsys.c (revision cd3e933325e68e23516a196a8fea7f49b1e497c3)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
28 
29 #include <sys/param.h>
30 #include <sys/types.h>
31 #include <sys/sysmacros.h>
32 #include <sys/systm.h>
33 #include <sys/prsystm.h>
34 #include <sys/cred.h>
35 #include <sys/errno.h>
36 #include <sys/proc.h>
37 #include <sys/signal.h>
38 #include <sys/kmem.h>
39 #include <sys/unistd.h>
40 #include <sys/cmn_err.h>
41 #include <sys/schedctl.h>
42 #include <sys/debug.h>
43 #include <sys/contract/process_impl.h>
44 
45 kthread_t *
46 idtot(proc_t *p, id_t lwpid)
47 {
48 	lwpdir_t *ldp;
49 
50 	if ((ldp = lwp_hash_lookup(p, lwpid)) != NULL)
51 		return (ldp->ld_entry->le_thread);
52 	return (NULL);
53 }
54 
55 /*
56  * Same as idtot(), but acquire and return
57  * the tid hash table entry lock on success.
58  * This allows lwp_unpark() to do its job without acquiring
59  * p->p_lock (and thereby causing congestion problems when
60  * the application calls lwp_unpark() too often).
61  */
62 static kthread_t *
63 idtot_and_lock(proc_t *p, id_t lwpid, kmutex_t **mpp)
64 {
65 	lwpdir_t *ldp;
66 	kthread_t *t;
67 
68 	if ((ldp = lwp_hash_lookup_and_lock(p, lwpid, mpp)) != NULL) {
69 		if ((t = ldp->ld_entry->le_thread) == NULL)
70 			mutex_exit(*mpp);
71 		return (t);
72 	}
73 	return (NULL);
74 }
75 
76 /*
77  * Stop an lwp of the current process
78  */
79 int
80 syslwp_suspend(id_t lwpid)
81 {
82 	kthread_t *t;
83 	int error;
84 	proc_t *p = ttoproc(curthread);
85 
86 	mutex_enter(&p->p_lock);
87 	if ((t = idtot(p, lwpid)) == NULL)
88 		error = ESRCH;
89 	else
90 		error = lwp_suspend(t);
91 	mutex_exit(&p->p_lock);
92 	if (error)
93 		return (set_errno(error));
94 	return (0);
95 }
96 
97 int
98 syslwp_continue(id_t lwpid)
99 {
100 	kthread_t *t;
101 	proc_t *p = ttoproc(curthread);
102 
103 	mutex_enter(&p->p_lock);
104 	if ((t = idtot(p, lwpid)) == NULL) {
105 		mutex_exit(&p->p_lock);
106 		return (set_errno(ESRCH));
107 	}
108 	lwp_continue(t);
109 	mutex_exit(&p->p_lock);
110 	return (0);
111 }
112 
113 int
114 lwp_kill(id_t lwpid, int sig)
115 {
116 	sigqueue_t *sqp;
117 	kthread_t *t;
118 	proc_t *p = ttoproc(curthread);
119 
120 	if (sig < 0 || sig >= NSIG)
121 		return (set_errno(EINVAL));
122 	if (sig != 0)
123 		sqp = kmem_zalloc(sizeof (sigqueue_t), KM_SLEEP);
124 	mutex_enter(&p->p_lock);
125 	if ((t = idtot(p, lwpid)) == NULL) {
126 		mutex_exit(&p->p_lock);
127 		if (sig != 0)
128 			kmem_free(sqp, sizeof (sigqueue_t));
129 		return (set_errno(ESRCH));
130 	}
131 	if (sig == 0) {
132 		mutex_exit(&p->p_lock);
133 		return (0);
134 	}
135 	sqp->sq_info.si_signo = sig;
136 	sqp->sq_info.si_code = SI_LWP;
137 	sqp->sq_info.si_pid = p->p_pid;
138 	sqp->sq_info.si_ctid = PRCTID(p);
139 	sqp->sq_info.si_zoneid = getzoneid();
140 	sqp->sq_info.si_uid = crgetruid(CRED());
141 	sigaddqa(p, t, sqp);
142 	mutex_exit(&p->p_lock);
143 	return (0);
144 }
145 
146 /*
147  * This is the specification of lwp_wait() from the _lwp_wait(2) manual page:
148  *
149  * The lwp_wait() function blocks the current lwp until the lwp specified
150  * by 'lwpid' terminates.  If the specified lwp terminated prior to the call
151  * to lwp_wait(), then lwp_wait() returns immediately.  If 'lwpid' is zero,
152  * then lwp_wait() waits for any undetached lwp in the current process.
153  * If 'lwpid' is not zero, then it must specify an undetached lwp in the
154  * current process.  If 'departed' is not NULL, then it points to a location
155  * where the id of the exited lwp is stored.
156  *
157  * When an lwp exits and there are one or more lwps in the process waiting
158  * for this specific lwp to exit, then one of the waiting lwps is unblocked
159  * and it returns from lwp_wait() successfully.  Any other lwps waiting for
160  * this same lwp to exit are also unblocked, however, they return from
161  * lwp_wait() with the error ESRCH.  If there are no lwps in the process
162  * waiting for this specific lwp to exit but there are one or more lwps
163  * waiting for any lwp to exit, then one of the waiting lwps is unblocked
164  * and it returns from lwp_wait() successfully.
165  *
166  * If an lwp is waiting for any lwp to exit, it blocks until an undetached
167  * lwp for which no other lwp is waiting terminates, at which time it returns
168  * successfully, or until all other lwps in the process are either daemon
169  * lwps or lwps waiting in lwp_wait(), in which case it returns EDEADLK.
170  */
171 int
172 lwp_wait(id_t lwpid, id_t *departed)
173 {
174 	proc_t *p = ttoproc(curthread);
175 	int error = 0;
176 	int daemon = (curthread->t_proc_flag & TP_DAEMON)? 1 : 0;
177 	lwpent_t *target_lep;
178 	lwpdir_t *ldp;
179 	lwpent_t *lep;
180 
181 	/*
182 	 * lwp_wait() is not supported for the /proc agent lwp.
183 	 */
184 	if (curthread == p->p_agenttp)
185 		return (set_errno(ENOTSUP));
186 
187 	mutex_enter(&p->p_lock);
188 	prbarrier(p);
189 
190 	curthread->t_waitfor = lwpid;
191 	p->p_lwpwait++;
192 	p->p_lwpdwait += daemon;
193 
194 	if (lwpid != 0) {
195 		if ((ldp = lwp_hash_lookup(p, lwpid)) == NULL)
196 			target_lep = NULL;
197 		else {
198 			target_lep = ldp->ld_entry;
199 			target_lep->le_waiters++;
200 			target_lep->le_dwaiters += daemon;
201 		}
202 	}
203 
204 	while (error == 0) {
205 		kthread_t *t;
206 		id_t tid;
207 		int i;
208 
209 		if (lwpid != 0) {
210 			/*
211 			 * Look for a specific zombie lwp.
212 			 */
213 			if (target_lep == NULL)
214 				error = ESRCH;
215 			else if ((t = target_lep->le_thread) != NULL) {
216 				if (!(t->t_proc_flag & TP_TWAIT))
217 					error = EINVAL;
218 			} else {
219 				/*
220 				 * We found the zombie we are waiting for.
221 				 */
222 				ASSERT(p->p_zombcnt > 0);
223 				p->p_zombcnt--;
224 				p->p_lwpwait--;
225 				p->p_lwpdwait -= daemon;
226 				curthread->t_waitfor = -1;
227 				lwp_hash_out(p, lwpid);
228 				mutex_exit(&p->p_lock);
229 				if (departed != NULL &&
230 				    copyout(&lwpid, departed, sizeof (id_t)))
231 					return (set_errno(EFAULT));
232 				return (0);
233 			}
234 		} else {
235 			/*
236 			 * Look for any zombie lwp.
237 			 */
238 			int some_non_daemon_will_return = 0;
239 
240 			/* for each entry in the lwp directory... */
241 			ldp = p->p_lwpdir;
242 			for (i = 0; i < p->p_lwpdir_sz; i++, ldp++) {
243 
244 				if ((lep = ldp->ld_entry) == NULL ||
245 				    lep->le_thread != NULL)
246 					continue;
247 
248 				/*
249 				 * We found a zombie lwp.  If there is some
250 				 * other thread waiting specifically for the
251 				 * zombie we just found, then defer to the other
252 				 * waiting thread and continue searching for
253 				 * another zombie.  Also check to see if there
254 				 * is some non-daemon thread sleeping here in
255 				 * lwp_wait() that will succeed and return when
256 				 * we drop p->p_lock.  This is tested below.
257 				 */
258 				tid = lep->le_lwpid;
259 				if (lep->le_waiters != 0) {
260 					if (lep->le_waiters - lep->le_dwaiters)
261 						some_non_daemon_will_return = 1;
262 					continue;
263 				}
264 
265 				/*
266 				 * We found a zombie that no one else
267 				 * is specifically waiting for.
268 				 */
269 				ASSERT(p->p_zombcnt > 0);
270 				p->p_zombcnt--;
271 				p->p_lwpwait--;
272 				p->p_lwpdwait -= daemon;
273 				curthread->t_waitfor = -1;
274 				lwp_hash_out(p, tid);
275 				mutex_exit(&p->p_lock);
276 				if (departed != NULL &&
277 				    copyout(&tid, departed, sizeof (id_t)))
278 					return (set_errno(EFAULT));
279 				return (0);
280 			}
281 
282 			/*
283 			 * We are waiting for anyone.  If all non-daemon lwps
284 			 * are waiting here, and if we determined above that
285 			 * no non-daemon lwp will return, we have deadlock.
286 			 */
287 			if (!some_non_daemon_will_return &&
288 			    p->p_lwpcnt == p->p_lwpdaemon +
289 			    (p->p_lwpwait - p->p_lwpdwait))
290 				error = EDEADLK;
291 		}
292 
293 		if (error == 0 && lwpid != 0) {
294 			/*
295 			 * We are waiting for a specific non-zombie lwp.
296 			 * Fail if there is a deadlock loop.
297 			 */
298 			for (;;) {
299 				if (t == curthread) {
300 					error = EDEADLK;
301 					break;
302 				}
303 				/* who is he waiting for? */
304 				if ((tid = t->t_waitfor) == -1)
305 					break;
306 				if (tid == 0) {
307 					/*
308 					 * The lwp we are waiting for is
309 					 * waiting for anyone (transitively).
310 					 * If there are no zombies right now
311 					 * and if we would have deadlock due
312 					 * to all non-daemon lwps waiting here,
313 					 * wake up the lwp that is waiting for
314 					 * anyone so it can return EDEADLK.
315 					 */
316 					if (p->p_zombcnt == 0 &&
317 					    p->p_lwpcnt == p->p_lwpdaemon +
318 					    p->p_lwpwait - p->p_lwpdwait)
319 						cv_broadcast(&p->p_lwpexit);
320 					break;
321 				}
322 				if ((ldp = lwp_hash_lookup(p, tid)) == NULL ||
323 				    (t = ldp->ld_entry->le_thread) == NULL)
324 					break;
325 			}
326 		}
327 
328 		if (error)
329 			break;
330 
331 		/*
332 		 * Wait for some lwp to terminate.
333 		 */
334 		if (!cv_wait_sig(&p->p_lwpexit, &p->p_lock))
335 			error = EINTR;
336 		prbarrier(p);
337 
338 		if (lwpid != 0) {
339 			if ((ldp = lwp_hash_lookup(p, lwpid)) == NULL)
340 				target_lep = NULL;
341 			else
342 				target_lep = ldp->ld_entry;
343 		}
344 	}
345 
346 	if (lwpid != 0 && target_lep != NULL) {
347 		target_lep->le_waiters--;
348 		target_lep->le_dwaiters -= daemon;
349 	}
350 	p->p_lwpwait--;
351 	p->p_lwpdwait -= daemon;
352 	curthread->t_waitfor = -1;
353 	mutex_exit(&p->p_lock);
354 	return (set_errno(error));
355 }
356 
357 int
358 lwp_detach(id_t lwpid)
359 {
360 	kthread_t *t;
361 	proc_t *p = ttoproc(curthread);
362 	lwpdir_t *ldp;
363 	int error = 0;
364 
365 	mutex_enter(&p->p_lock);
366 	prbarrier(p);
367 	if ((ldp = lwp_hash_lookup(p, lwpid)) == NULL)
368 		error = ESRCH;
369 	else if ((t = ldp->ld_entry->le_thread) != NULL) {
370 		if (!(t->t_proc_flag & TP_TWAIT))
371 			error = EINVAL;
372 		else {
373 			t->t_proc_flag &= ~TP_TWAIT;
374 			cv_broadcast(&p->p_lwpexit);
375 		}
376 	} else {
377 		ASSERT(p->p_zombcnt > 0);
378 		p->p_zombcnt--;
379 		lwp_hash_out(p, lwpid);
380 	}
381 	mutex_exit(&p->p_lock);
382 
383 	if (error)
384 		return (set_errno(error));
385 	return (0);
386 }
387 
388 /*
389  * Unpark the specified lwp.
390  */
391 static int
392 lwp_unpark(id_t lwpid)
393 {
394 	proc_t *p = ttoproc(curthread);
395 	kthread_t *t;
396 	kmutex_t *mp;
397 	int error = 0;
398 
399 	if ((t = idtot_and_lock(p, lwpid, &mp)) == NULL) {
400 		error = ESRCH;
401 	} else {
402 		mutex_enter(&t->t_delay_lock);
403 		t->t_unpark = 1;
404 		cv_signal(&t->t_delay_cv);
405 		mutex_exit(&t->t_delay_lock);
406 		mutex_exit(mp);
407 	}
408 	return (error);
409 }
410 
411 /*
412  * Cancel a previous unpark for the specified lwp.
413  *
414  * This interface exists ONLY to support older versions of libthread, which
415  * called lwp_unpark(self) to force calls to lwp_park(self) to return
416  * immediately.  These older libthreads required a mechanism to cancel the
417  * lwp_unpark(self).
418  *
419  * libc does not call this interface.  Instead, the sc_park flag in the
420  * schedctl page is cleared to force calls to lwp_park() to return
421  * immediately.
422  */
423 static int
424 lwp_unpark_cancel(id_t lwpid)
425 {
426 	proc_t *p = ttoproc(curthread);
427 	kthread_t *t;
428 	kmutex_t *mp;
429 	int error = 0;
430 
431 	if ((t = idtot_and_lock(p, lwpid, &mp)) == NULL) {
432 		error = ESRCH;
433 	} else {
434 		mutex_enter(&t->t_delay_lock);
435 		t->t_unpark = 0;
436 		mutex_exit(&t->t_delay_lock);
437 		mutex_exit(mp);
438 	}
439 	return (error);
440 }
441 
442 /*
443  * Sleep until we are set running by lwp_unpark() or until we are
444  * interrupted by a signal or until we exhaust our timeout.
445  * timeoutp is an in/out parameter.  On entry, it contains the relative
446  * time until timeout.  On exit, we copyout the residual time left to it.
447  */
448 static int
449 lwp_park(timespec_t *timeoutp, id_t lwpid)
450 {
451 	timespec_t rqtime;
452 	timespec_t rmtime;
453 	timespec_t now;
454 	timespec_t *rqtp = NULL;
455 	kthread_t *t = curthread;
456 	int timecheck = 0;
457 	int error = 0;
458 	model_t datamodel = ttoproc(t)->p_model;
459 
460 	if (lwpid != 0)		/* unpark the other lwp, if any */
461 		(void) lwp_unpark(lwpid);
462 
463 	if (timeoutp) {
464 		timecheck = timechanged;
465 		gethrestime(&now);
466 		if (datamodel == DATAMODEL_NATIVE) {
467 			if (copyin(timeoutp, &rqtime, sizeof (timespec_t))) {
468 				error = EFAULT;
469 				goto out;
470 			}
471 		} else {
472 			timespec32_t timeout32;
473 
474 			if (copyin(timeoutp, &timeout32, sizeof (timeout32))) {
475 				error = EFAULT;
476 				goto out;
477 			}
478 			TIMESPEC32_TO_TIMESPEC(&rqtime, &timeout32)
479 		}
480 
481 		if (itimerspecfix(&rqtime)) {
482 			error = EINVAL;
483 			goto out;
484 		}
485 		/*
486 		 * Convert the timespec value into absolute time.
487 		 */
488 		timespecadd(&rqtime, &now);
489 		rqtp = &rqtime;
490 	}
491 
492 	(void) new_mstate(t, LMS_USER_LOCK);
493 
494 	mutex_enter(&t->t_delay_lock);
495 	if (!schedctl_is_park())
496 		error = EINTR;
497 	while (error == 0 && t->t_unpark == 0) {
498 		switch (cv_waituntil_sig(&t->t_delay_cv,
499 		    &t->t_delay_lock, rqtp, timecheck)) {
500 		case 0:
501 			error = EINTR;
502 			break;
503 		case -1:
504 			error = ETIME;
505 			break;
506 		}
507 	}
508 	t->t_unpark = 0;
509 	mutex_exit(&t->t_delay_lock);
510 
511 	if (timeoutp != NULL) {
512 		rmtime.tv_sec = rmtime.tv_nsec = 0;
513 		if (error != ETIME) {
514 			gethrestime(&now);
515 			if ((now.tv_sec < rqtime.tv_sec) ||
516 			    ((now.tv_sec == rqtime.tv_sec) &&
517 			    (now.tv_nsec < rqtime.tv_nsec))) {
518 				rmtime = rqtime;
519 				timespecsub(&rmtime, &now);
520 			}
521 		}
522 		if (datamodel == DATAMODEL_NATIVE) {
523 			if (copyout(&rmtime, timeoutp, sizeof (rmtime)))
524 				error = EFAULT;
525 		} else {
526 			timespec32_t rmtime32;
527 
528 			TIMESPEC_TO_TIMESPEC32(&rmtime32, &rmtime);
529 			if (copyout(&rmtime32, timeoutp, sizeof (rmtime32)))
530 				error = EFAULT;
531 		}
532 	}
533 out:
534 	schedctl_unpark();
535 	if (t->t_mstate == LMS_USER_LOCK)
536 		(void) new_mstate(t, LMS_SYSTEM);
537 	return (error);
538 }
539 
540 #define	MAXLWPIDS	1024
541 
542 /*
543  * Unpark all of the specified lwps.
544  * Do it in chunks of MAXLWPIDS to avoid allocating too much memory.
545  */
546 static int
547 lwp_unpark_all(id_t *lwpidp, int nids)
548 {
549 	proc_t *p = ttoproc(curthread);
550 	kthread_t *t;
551 	kmutex_t *mp;
552 	int error = 0;
553 	id_t *lwpid;
554 	size_t lwpidsz;
555 	int n;
556 	int i;
557 
558 	if (nids <= 0)
559 		return (EINVAL);
560 
561 	lwpidsz = MIN(nids, MAXLWPIDS) * sizeof (id_t);
562 	lwpid = kmem_alloc(lwpidsz, KM_SLEEP);
563 	while (nids > 0) {
564 		n = MIN(nids, MAXLWPIDS);
565 		if (copyin(lwpidp, lwpid, n * sizeof (id_t))) {
566 			error = EFAULT;
567 			break;
568 		}
569 		for (i = 0; i < n; i++) {
570 			if ((t = idtot_and_lock(p, lwpid[i], &mp)) == NULL) {
571 				error = ESRCH;
572 			} else {
573 				mutex_enter(&t->t_delay_lock);
574 				t->t_unpark = 1;
575 				cv_signal(&t->t_delay_cv);
576 				mutex_exit(&t->t_delay_lock);
577 				mutex_exit(mp);
578 			}
579 		}
580 		lwpidp += n;
581 		nids -= n;
582 	}
583 	kmem_free(lwpid, lwpidsz);
584 	return (error);
585 }
586 
587 /*
588  * SYS_lwp_park() system call.
589  */
590 int
591 syslwp_park(int which, uintptr_t arg1, uintptr_t arg2)
592 {
593 	int error;
594 
595 	switch (which) {
596 	case 0:
597 		error = lwp_park((timespec_t *)arg1, (id_t)arg2);
598 		break;
599 	case 1:
600 		error = lwp_unpark((id_t)arg1);
601 		break;
602 	case 2:
603 		error = lwp_unpark_all((id_t *)arg1, (int)arg2);
604 		break;
605 	case 3:
606 		/*
607 		 * This subcode is not used by libc.  It exists ONLY to
608 		 * support older versions of libthread which do not use
609 		 * the sc_park flag in the schedctl page.
610 		 *
611 		 * These versions of libthread need to be modifed or emulated
612 		 * to change calls to syslwp_park(1, tid, 0) to
613 		 * syslwp_park(3, tid).
614 		 */
615 		error = lwp_unpark_cancel((id_t)arg1);
616 		break;
617 	case 4:
618 		/*
619 		 * This subcode is not used by libc.  It exists ONLY to
620 		 * support older versions of libthread which do not use
621 		 * the sc_park flag in the schedctl page.
622 		 *
623 		 * These versions of libthread need to be modified or emulated
624 		 * to change calls to syslwp_park(0, ts, tid) to
625 		 * syslwp_park(4, ts, tid).
626 		 */
627 		schedctl_set_park();
628 		error = lwp_park((timespec_t *)arg1, (id_t)arg2);
629 		break;
630 	default:
631 		error = EINVAL;
632 		break;
633 	}
634 
635 	if (error)
636 		return (set_errno(error));
637 	return (0);
638 }
639