xref: /freebsd/sys/kern/kern_thread.c (revision 5608fd23c27fa1e8ee595d7b678cbfd35d657fbe)
1 /*-
2  * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
3  *  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice(s), this list of conditions and the following disclaimer as
10  *    the first lines of this file unmodified other than the possible
11  *    addition of one or more copyright notices.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice(s), this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
26  * DAMAGE.
27  */
28 
29 #include "opt_witness.h"
30 #include "opt_hwpmc_hooks.h"
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/rangelock.h>
42 #include <sys/resourcevar.h>
43 #include <sys/sdt.h>
44 #include <sys/smp.h>
45 #include <sys/sched.h>
46 #include <sys/sleepqueue.h>
47 #include <sys/selinfo.h>
48 #include <sys/turnstile.h>
49 #include <sys/ktr.h>
50 #include <sys/rwlock.h>
51 #include <sys/umtx.h>
52 #include <sys/cpuset.h>
53 #ifdef	HWPMC_HOOKS
54 #include <sys/pmckern.h>
55 #endif
56 
57 #include <security/audit/audit.h>
58 
59 #include <vm/vm.h>
60 #include <vm/vm_extern.h>
61 #include <vm/uma.h>
62 #include <sys/eventhandler.h>
63 
64 SDT_PROVIDER_DECLARE(proc);
65 SDT_PROBE_DEFINE(proc, , , lwp__exit);
66 
67 
68 /*
69  * thread related storage.
70  */
71 static uma_zone_t thread_zone;
72 
73 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
74 static struct mtx zombie_lock;
75 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
76 
77 static void thread_zombie(struct thread *);
78 
79 #define TID_BUFFER_SIZE	1024
80 
81 struct mtx tid_lock;
82 static struct unrhdr *tid_unrhdr;
83 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
84 static int tid_head, tid_tail;
85 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
86 
87 struct	tidhashhead *tidhashtbl;
88 u_long	tidhash;
89 struct	rwlock tidhash_lock;
90 
91 static lwpid_t
92 tid_alloc(void)
93 {
94 	lwpid_t	tid;
95 
96 	tid = alloc_unr(tid_unrhdr);
97 	if (tid != -1)
98 		return (tid);
99 	mtx_lock(&tid_lock);
100 	if (tid_head == tid_tail) {
101 		mtx_unlock(&tid_lock);
102 		return (-1);
103 	}
104 	tid = tid_buffer[tid_head];
105 	tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
106 	mtx_unlock(&tid_lock);
107 	return (tid);
108 }
109 
110 static void
111 tid_free(lwpid_t tid)
112 {
113 	lwpid_t tmp_tid = -1;
114 
115 	mtx_lock(&tid_lock);
116 	if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
117 		tmp_tid = tid_buffer[tid_head];
118 		tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
119 	}
120 	tid_buffer[tid_tail] = tid;
121 	tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE;
122 	mtx_unlock(&tid_lock);
123 	if (tmp_tid != -1)
124 		free_unr(tid_unrhdr, tmp_tid);
125 }
126 
127 /*
128  * Prepare a thread for use.
129  */
130 static int
131 thread_ctor(void *mem, int size, void *arg, int flags)
132 {
133 	struct thread	*td;
134 
135 	td = (struct thread *)mem;
136 	td->td_state = TDS_INACTIVE;
137 	td->td_oncpu = NOCPU;
138 
139 	td->td_tid = tid_alloc();
140 
141 	/*
142 	 * Note that td_critnest begins life as 1 because the thread is not
143 	 * running and is thereby implicitly waiting to be on the receiving
144 	 * end of a context switch.
145 	 */
146 	td->td_critnest = 1;
147 	td->td_lend_user_pri = PRI_MAX;
148 	EVENTHANDLER_INVOKE(thread_ctor, td);
149 #ifdef AUDIT
150 	audit_thread_alloc(td);
151 #endif
152 	umtx_thread_alloc(td);
153 	return (0);
154 }
155 
156 /*
157  * Reclaim a thread after use.
158  */
159 static void
160 thread_dtor(void *mem, int size, void *arg)
161 {
162 	struct thread *td;
163 
164 	td = (struct thread *)mem;
165 
166 #ifdef INVARIANTS
167 	/* Verify that this thread is in a safe state to free. */
168 	switch (td->td_state) {
169 	case TDS_INHIBITED:
170 	case TDS_RUNNING:
171 	case TDS_CAN_RUN:
172 	case TDS_RUNQ:
173 		/*
174 		 * We must never unlink a thread that is in one of
175 		 * these states, because it is currently active.
176 		 */
177 		panic("bad state for thread unlinking");
178 		/* NOTREACHED */
179 	case TDS_INACTIVE:
180 		break;
181 	default:
182 		panic("bad thread state");
183 		/* NOTREACHED */
184 	}
185 #endif
186 #ifdef AUDIT
187 	audit_thread_free(td);
188 #endif
189 	/* Free all OSD associated to this thread. */
190 	osd_thread_exit(td);
191 
192 	EVENTHANDLER_INVOKE(thread_dtor, td);
193 	tid_free(td->td_tid);
194 }
195 
196 /*
197  * Initialize type-stable parts of a thread (when newly created).
198  */
199 static int
200 thread_init(void *mem, int size, int flags)
201 {
202 	struct thread *td;
203 
204 	td = (struct thread *)mem;
205 
206 	td->td_sleepqueue = sleepq_alloc();
207 	td->td_turnstile = turnstile_alloc();
208 	td->td_rlqe = NULL;
209 	EVENTHANDLER_INVOKE(thread_init, td);
210 	td->td_sched = (struct td_sched *)&td[1];
211 	umtx_thread_init(td);
212 	td->td_kstack = 0;
213 	return (0);
214 }
215 
216 /*
217  * Tear down type-stable parts of a thread (just before being discarded).
218  */
219 static void
220 thread_fini(void *mem, int size)
221 {
222 	struct thread *td;
223 
224 	td = (struct thread *)mem;
225 	EVENTHANDLER_INVOKE(thread_fini, td);
226 	rlqentry_free(td->td_rlqe);
227 	turnstile_free(td->td_turnstile);
228 	sleepq_free(td->td_sleepqueue);
229 	umtx_thread_fini(td);
230 	seltdfini(td);
231 }
232 
233 /*
234  * For a newly created process,
235  * link up all the structures and its initial threads etc.
236  * called from:
237  * {arch}/{arch}/machdep.c   {arch}_init(), init386() etc.
238  * proc_dtor() (should go away)
239  * proc_init()
240  */
241 void
242 proc_linkup0(struct proc *p, struct thread *td)
243 {
244 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
245 	proc_linkup(p, td);
246 }
247 
248 void
249 proc_linkup(struct proc *p, struct thread *td)
250 {
251 
252 	sigqueue_init(&p->p_sigqueue, p);
253 	p->p_ksi = ksiginfo_alloc(1);
254 	if (p->p_ksi != NULL) {
255 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
256 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
257 	}
258 	LIST_INIT(&p->p_mqnotifier);
259 	p->p_numthreads = 0;
260 	thread_link(td, p);
261 }
262 
263 /*
264  * Initialize global thread allocation resources.
265  */
266 void
267 threadinit(void)
268 {
269 
270 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
271 
272 	/*
273 	 * pid_max cannot be greater than PID_MAX.
274 	 * leave one number for thread0.
275 	 */
276 	tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
277 
278 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
279 	    thread_ctor, thread_dtor, thread_init, thread_fini,
280 	    16 - 1, 0);
281 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
282 	rw_init(&tidhash_lock, "tidhash");
283 }
284 
285 /*
286  * Place an unused thread on the zombie list.
287  * Use the slpq as that must be unused by now.
288  */
289 void
290 thread_zombie(struct thread *td)
291 {
292 	mtx_lock_spin(&zombie_lock);
293 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
294 	mtx_unlock_spin(&zombie_lock);
295 }
296 
297 /*
298  * Release a thread that has exited after cpu_throw().
299  */
300 void
301 thread_stash(struct thread *td)
302 {
303 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
304 	thread_zombie(td);
305 }
306 
307 /*
308  * Reap zombie resources.
309  */
310 void
311 thread_reap(void)
312 {
313 	struct thread *td_first, *td_next;
314 
315 	/*
316 	 * Don't even bother to lock if none at this instant,
317 	 * we really don't care about the next instant..
318 	 */
319 	if (!TAILQ_EMPTY(&zombie_threads)) {
320 		mtx_lock_spin(&zombie_lock);
321 		td_first = TAILQ_FIRST(&zombie_threads);
322 		if (td_first)
323 			TAILQ_INIT(&zombie_threads);
324 		mtx_unlock_spin(&zombie_lock);
325 		while (td_first) {
326 			td_next = TAILQ_NEXT(td_first, td_slpq);
327 			if (td_first->td_ucred)
328 				crfree(td_first->td_ucred);
329 			thread_free(td_first);
330 			td_first = td_next;
331 		}
332 	}
333 }
334 
335 /*
336  * Allocate a thread.
337  */
338 struct thread *
339 thread_alloc(int pages)
340 {
341 	struct thread *td;
342 
343 	thread_reap(); /* check if any zombies to get */
344 
345 	td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
346 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
347 	if (!vm_thread_new(td, pages)) {
348 		uma_zfree(thread_zone, td);
349 		return (NULL);
350 	}
351 	cpu_thread_alloc(td);
352 	return (td);
353 }
354 
355 int
356 thread_alloc_stack(struct thread *td, int pages)
357 {
358 
359 	KASSERT(td->td_kstack == 0,
360 	    ("thread_alloc_stack called on a thread with kstack"));
361 	if (!vm_thread_new(td, pages))
362 		return (0);
363 	cpu_thread_alloc(td);
364 	return (1);
365 }
366 
367 /*
368  * Deallocate a thread.
369  */
370 void
371 thread_free(struct thread *td)
372 {
373 
374 	lock_profile_thread_exit(td);
375 	if (td->td_cpuset)
376 		cpuset_rel(td->td_cpuset);
377 	td->td_cpuset = NULL;
378 	cpu_thread_free(td);
379 	if (td->td_kstack != 0)
380 		vm_thread_dispose(td);
381 	uma_zfree(thread_zone, td);
382 }
383 
384 /*
385  * Discard the current thread and exit from its context.
386  * Always called with scheduler locked.
387  *
388  * Because we can't free a thread while we're operating under its context,
389  * push the current thread into our CPU's deadthread holder. This means
390  * we needn't worry about someone else grabbing our context before we
391  * do a cpu_throw().
392  */
393 void
394 thread_exit(void)
395 {
396 	uint64_t runtime, new_switchtime;
397 	struct thread *td;
398 	struct thread *td2;
399 	struct proc *p;
400 	int wakeup_swapper;
401 
402 	td = curthread;
403 	p = td->td_proc;
404 
405 	PROC_SLOCK_ASSERT(p, MA_OWNED);
406 	mtx_assert(&Giant, MA_NOTOWNED);
407 
408 	PROC_LOCK_ASSERT(p, MA_OWNED);
409 	KASSERT(p != NULL, ("thread exiting without a process"));
410 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
411 	    (long)p->p_pid, td->td_name);
412 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
413 
414 #ifdef AUDIT
415 	AUDIT_SYSCALL_EXIT(0, td);
416 #endif
417 	umtx_thread_exit(td);
418 	/*
419 	 * drop FPU & debug register state storage, or any other
420 	 * architecture specific resources that
421 	 * would not be on a new untouched process.
422 	 */
423 	cpu_thread_exit(td);	/* XXXSMP */
424 
425 	/*
426 	 * The last thread is left attached to the process
427 	 * So that the whole bundle gets recycled. Skip
428 	 * all this stuff if we never had threads.
429 	 * EXIT clears all sign of other threads when
430 	 * it goes to single threading, so the last thread always
431 	 * takes the short path.
432 	 */
433 	if (p->p_flag & P_HADTHREADS) {
434 		if (p->p_numthreads > 1) {
435 			thread_unlink(td);
436 			td2 = FIRST_THREAD_IN_PROC(p);
437 			sched_exit_thread(td2, td);
438 
439 			/*
440 			 * The test below is NOT true if we are the
441 			 * sole exiting thread. P_STOPPED_SINGLE is unset
442 			 * in exit1() after it is the only survivor.
443 			 */
444 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
445 				if (p->p_numthreads == p->p_suspcount) {
446 					thread_lock(p->p_singlethread);
447 					wakeup_swapper = thread_unsuspend_one(
448 						p->p_singlethread);
449 					thread_unlock(p->p_singlethread);
450 					if (wakeup_swapper)
451 						kick_proc0();
452 				}
453 			}
454 
455 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
456 			PCPU_SET(deadthread, td);
457 		} else {
458 			/*
459 			 * The last thread is exiting.. but not through exit()
460 			 */
461 			panic ("thread_exit: Last thread exiting on its own");
462 		}
463 	}
464 #ifdef	HWPMC_HOOKS
465 	/*
466 	 * If this thread is part of a process that is being tracked by hwpmc(4),
467 	 * inform the module of the thread's impending exit.
468 	 */
469 	if (PMC_PROC_IS_USING_PMCS(td->td_proc))
470 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
471 #endif
472 	PROC_UNLOCK(p);
473 
474 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
475 	new_switchtime = cpu_ticks();
476 	runtime = new_switchtime - PCPU_GET(switchtime);
477 	td->td_runtime += runtime;
478 	td->td_incruntime += runtime;
479 	PCPU_SET(switchtime, new_switchtime);
480 	PCPU_SET(switchticks, ticks);
481 	PCPU_INC(cnt.v_swtch);
482 
483 	/* Save our resource usage in our process. */
484 	td->td_ru.ru_nvcsw++;
485 	ruxagg(p, td);
486 	rucollect(&p->p_ru, &td->td_ru);
487 
488 	thread_lock(td);
489 	PROC_SUNLOCK(p);
490 	td->td_state = TDS_INACTIVE;
491 #ifdef WITNESS
492 	witness_thread_exit(td);
493 #endif
494 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
495 	sched_throw(td);
496 	panic("I'm a teapot!");
497 	/* NOTREACHED */
498 }
499 
500 /*
501  * Do any thread specific cleanups that may be needed in wait()
502  * called with Giant, proc and schedlock not held.
503  */
504 void
505 thread_wait(struct proc *p)
506 {
507 	struct thread *td;
508 
509 	mtx_assert(&Giant, MA_NOTOWNED);
510 	KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
511 	td = FIRST_THREAD_IN_PROC(p);
512 	/* Lock the last thread so we spin until it exits cpu_throw(). */
513 	thread_lock(td);
514 	thread_unlock(td);
515 	/* Wait for any remaining threads to exit cpu_throw(). */
516 	while (p->p_exitthreads)
517 		sched_relinquish(curthread);
518 	lock_profile_thread_exit(td);
519 	cpuset_rel(td->td_cpuset);
520 	td->td_cpuset = NULL;
521 	cpu_thread_clean(td);
522 	crfree(td->td_ucred);
523 	thread_reap();	/* check for zombie threads etc. */
524 }
525 
526 /*
527  * Link a thread to a process.
528  * set up anything that needs to be initialized for it to
529  * be used by the process.
530  */
531 void
532 thread_link(struct thread *td, struct proc *p)
533 {
534 
535 	/*
536 	 * XXX This can't be enabled because it's called for proc0 before
537 	 * its lock has been created.
538 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
539 	 */
540 	td->td_state    = TDS_INACTIVE;
541 	td->td_proc     = p;
542 	td->td_flags    = TDF_INMEM;
543 
544 	LIST_INIT(&td->td_contested);
545 	LIST_INIT(&td->td_lprof[0]);
546 	LIST_INIT(&td->td_lprof[1]);
547 	sigqueue_init(&td->td_sigqueue, p);
548 	callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
549 	TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
550 	p->p_numthreads++;
551 }
552 
553 /*
554  * Convert a process with one thread to an unthreaded process.
555  */
556 void
557 thread_unthread(struct thread *td)
558 {
559 	struct proc *p = td->td_proc;
560 
561 	KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
562 	p->p_flag &= ~P_HADTHREADS;
563 }
564 
565 /*
566  * Called from:
567  *  thread_exit()
568  */
569 void
570 thread_unlink(struct thread *td)
571 {
572 	struct proc *p = td->td_proc;
573 
574 	PROC_LOCK_ASSERT(p, MA_OWNED);
575 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
576 	p->p_numthreads--;
577 	/* could clear a few other things here */
578 	/* Must  NOT clear links to proc! */
579 }
580 
581 static int
582 calc_remaining(struct proc *p, int mode)
583 {
584 	int remaining;
585 
586 	PROC_LOCK_ASSERT(p, MA_OWNED);
587 	PROC_SLOCK_ASSERT(p, MA_OWNED);
588 	if (mode == SINGLE_EXIT)
589 		remaining = p->p_numthreads;
590 	else if (mode == SINGLE_BOUNDARY)
591 		remaining = p->p_numthreads - p->p_boundary_count;
592 	else if (mode == SINGLE_NO_EXIT)
593 		remaining = p->p_numthreads - p->p_suspcount;
594 	else
595 		panic("calc_remaining: wrong mode %d", mode);
596 	return (remaining);
597 }
598 
599 /*
600  * Enforce single-threading.
601  *
602  * Returns 1 if the caller must abort (another thread is waiting to
603  * exit the process or similar). Process is locked!
604  * Returns 0 when you are successfully the only thread running.
605  * A process has successfully single threaded in the suspend mode when
606  * There are no threads in user mode. Threads in the kernel must be
607  * allowed to continue until they get to the user boundary. They may even
608  * copy out their return values and data before suspending. They may however be
609  * accelerated in reaching the user boundary as we will wake up
610  * any sleeping threads that are interruptable. (PCATCH).
611  */
612 int
613 thread_single(int mode)
614 {
615 	struct thread *td;
616 	struct thread *td2;
617 	struct proc *p;
618 	int remaining, wakeup_swapper;
619 
620 	td = curthread;
621 	p = td->td_proc;
622 	mtx_assert(&Giant, MA_NOTOWNED);
623 	PROC_LOCK_ASSERT(p, MA_OWNED);
624 
625 	if ((p->p_flag & P_HADTHREADS) == 0)
626 		return (0);
627 
628 	/* Is someone already single threading? */
629 	if (p->p_singlethread != NULL && p->p_singlethread != td)
630 		return (1);
631 
632 	if (mode == SINGLE_EXIT) {
633 		p->p_flag |= P_SINGLE_EXIT;
634 		p->p_flag &= ~P_SINGLE_BOUNDARY;
635 	} else {
636 		p->p_flag &= ~P_SINGLE_EXIT;
637 		if (mode == SINGLE_BOUNDARY)
638 			p->p_flag |= P_SINGLE_BOUNDARY;
639 		else
640 			p->p_flag &= ~P_SINGLE_BOUNDARY;
641 	}
642 	p->p_flag |= P_STOPPED_SINGLE;
643 	PROC_SLOCK(p);
644 	p->p_singlethread = td;
645 	remaining = calc_remaining(p, mode);
646 	while (remaining != 1) {
647 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
648 			goto stopme;
649 		wakeup_swapper = 0;
650 		FOREACH_THREAD_IN_PROC(p, td2) {
651 			if (td2 == td)
652 				continue;
653 			thread_lock(td2);
654 			td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
655 			if (TD_IS_INHIBITED(td2)) {
656 				switch (mode) {
657 				case SINGLE_EXIT:
658 					if (TD_IS_SUSPENDED(td2))
659 						wakeup_swapper |=
660 						    thread_unsuspend_one(td2);
661 					if (TD_ON_SLEEPQ(td2) &&
662 					    (td2->td_flags & TDF_SINTR))
663 						wakeup_swapper |=
664 						    sleepq_abort(td2, EINTR);
665 					break;
666 				case SINGLE_BOUNDARY:
667 					if (TD_IS_SUSPENDED(td2) &&
668 					    !(td2->td_flags & TDF_BOUNDARY))
669 						wakeup_swapper |=
670 						    thread_unsuspend_one(td2);
671 					if (TD_ON_SLEEPQ(td2) &&
672 					    (td2->td_flags & TDF_SINTR))
673 						wakeup_swapper |=
674 						    sleepq_abort(td2, ERESTART);
675 					break;
676 				case SINGLE_NO_EXIT:
677 					if (TD_IS_SUSPENDED(td2) &&
678 					    !(td2->td_flags & TDF_BOUNDARY))
679 						wakeup_swapper |=
680 						    thread_unsuspend_one(td2);
681 					if (TD_ON_SLEEPQ(td2) &&
682 					    (td2->td_flags & TDF_SINTR))
683 						wakeup_swapper |=
684 						    sleepq_abort(td2, ERESTART);
685 					break;
686 				default:
687 					break;
688 				}
689 			}
690 #ifdef SMP
691 			else if (TD_IS_RUNNING(td2) && td != td2) {
692 				forward_signal(td2);
693 			}
694 #endif
695 			thread_unlock(td2);
696 		}
697 		if (wakeup_swapper)
698 			kick_proc0();
699 		remaining = calc_remaining(p, mode);
700 
701 		/*
702 		 * Maybe we suspended some threads.. was it enough?
703 		 */
704 		if (remaining == 1)
705 			break;
706 
707 stopme:
708 		/*
709 		 * Wake us up when everyone else has suspended.
710 		 * In the mean time we suspend as well.
711 		 */
712 		thread_suspend_switch(td);
713 		remaining = calc_remaining(p, mode);
714 	}
715 	if (mode == SINGLE_EXIT) {
716 		/*
717 		 * We have gotten rid of all the other threads and we
718 		 * are about to either exit or exec. In either case,
719 		 * we try our utmost to revert to being a non-threaded
720 		 * process.
721 		 */
722 		p->p_singlethread = NULL;
723 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
724 		thread_unthread(td);
725 	}
726 	PROC_SUNLOCK(p);
727 	return (0);
728 }
729 
730 /*
731  * Called in from locations that can safely check to see
732  * whether we have to suspend or at least throttle for a
733  * single-thread event (e.g. fork).
734  *
735  * Such locations include userret().
736  * If the "return_instead" argument is non zero, the thread must be able to
737  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
738  *
739  * The 'return_instead' argument tells the function if it may do a
740  * thread_exit() or suspend, or whether the caller must abort and back
741  * out instead.
742  *
743  * If the thread that set the single_threading request has set the
744  * P_SINGLE_EXIT bit in the process flags then this call will never return
745  * if 'return_instead' is false, but will exit.
746  *
747  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
748  *---------------+--------------------+---------------------
749  *       0       | returns 0          |   returns 0 or 1
750  *               | when ST ends       |   immediately
751  *---------------+--------------------+---------------------
752  *       1       | thread exits       |   returns 1
753  *               |                    |  immediately
754  * 0 = thread_exit() or suspension ok,
755  * other = return error instead of stopping the thread.
756  *
757  * While a full suspension is under effect, even a single threading
758  * thread would be suspended if it made this call (but it shouldn't).
759  * This call should only be made from places where
760  * thread_exit() would be safe as that may be the outcome unless
761  * return_instead is set.
762  */
763 int
764 thread_suspend_check(int return_instead)
765 {
766 	struct thread *td;
767 	struct proc *p;
768 	int wakeup_swapper;
769 
770 	td = curthread;
771 	p = td->td_proc;
772 	mtx_assert(&Giant, MA_NOTOWNED);
773 	PROC_LOCK_ASSERT(p, MA_OWNED);
774 	while (P_SHOULDSTOP(p) ||
775 	      ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
776 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
777 			KASSERT(p->p_singlethread != NULL,
778 			    ("singlethread not set"));
779 			/*
780 			 * The only suspension in action is a
781 			 * single-threading. Single threader need not stop.
782 			 * XXX Should be safe to access unlocked
783 			 * as it can only be set to be true by us.
784 			 */
785 			if (p->p_singlethread == td)
786 				return (0);	/* Exempt from stopping. */
787 		}
788 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
789 			return (EINTR);
790 
791 		/* Should we goto user boundary if we didn't come from there? */
792 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
793 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
794 			return (ERESTART);
795 
796 		/*
797 		 * Ignore suspend requests for stop signals if they
798 		 * are deferred.
799 		 */
800 		if (P_SHOULDSTOP(p) == P_STOPPED_SIG &&
801 		    td->td_flags & TDF_SBDRY) {
802 			KASSERT(return_instead,
803 			    ("TDF_SBDRY set for unsafe thread_suspend_check"));
804 			return (0);
805 		}
806 
807 		/*
808 		 * If the process is waiting for us to exit,
809 		 * this thread should just suicide.
810 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
811 		 */
812 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
813 			PROC_UNLOCK(p);
814 			tidhash_remove(td);
815 			PROC_LOCK(p);
816 			tdsigcleanup(td);
817 			PROC_SLOCK(p);
818 			thread_stopped(p);
819 			thread_exit();
820 		}
821 
822 		PROC_SLOCK(p);
823 		thread_stopped(p);
824 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
825 			if (p->p_numthreads == p->p_suspcount + 1) {
826 				thread_lock(p->p_singlethread);
827 				wakeup_swapper =
828 				    thread_unsuspend_one(p->p_singlethread);
829 				thread_unlock(p->p_singlethread);
830 				if (wakeup_swapper)
831 					kick_proc0();
832 			}
833 		}
834 		PROC_UNLOCK(p);
835 		thread_lock(td);
836 		/*
837 		 * When a thread suspends, it just
838 		 * gets taken off all queues.
839 		 */
840 		thread_suspend_one(td);
841 		if (return_instead == 0) {
842 			p->p_boundary_count++;
843 			td->td_flags |= TDF_BOUNDARY;
844 		}
845 		PROC_SUNLOCK(p);
846 		mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
847 		if (return_instead == 0)
848 			td->td_flags &= ~TDF_BOUNDARY;
849 		thread_unlock(td);
850 		PROC_LOCK(p);
851 		if (return_instead == 0) {
852 			PROC_SLOCK(p);
853 			p->p_boundary_count--;
854 			PROC_SUNLOCK(p);
855 		}
856 	}
857 	return (0);
858 }
859 
860 void
861 thread_suspend_switch(struct thread *td)
862 {
863 	struct proc *p;
864 
865 	p = td->td_proc;
866 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
867 	PROC_LOCK_ASSERT(p, MA_OWNED);
868 	PROC_SLOCK_ASSERT(p, MA_OWNED);
869 	/*
870 	 * We implement thread_suspend_one in stages here to avoid
871 	 * dropping the proc lock while the thread lock is owned.
872 	 */
873 	thread_stopped(p);
874 	p->p_suspcount++;
875 	PROC_UNLOCK(p);
876 	thread_lock(td);
877 	td->td_flags &= ~TDF_NEEDSUSPCHK;
878 	TD_SET_SUSPENDED(td);
879 	sched_sleep(td, 0);
880 	PROC_SUNLOCK(p);
881 	DROP_GIANT();
882 	mi_switch(SW_VOL | SWT_SUSPEND, NULL);
883 	thread_unlock(td);
884 	PICKUP_GIANT();
885 	PROC_LOCK(p);
886 	PROC_SLOCK(p);
887 }
888 
889 void
890 thread_suspend_one(struct thread *td)
891 {
892 	struct proc *p = td->td_proc;
893 
894 	PROC_SLOCK_ASSERT(p, MA_OWNED);
895 	THREAD_LOCK_ASSERT(td, MA_OWNED);
896 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
897 	p->p_suspcount++;
898 	td->td_flags &= ~TDF_NEEDSUSPCHK;
899 	TD_SET_SUSPENDED(td);
900 	sched_sleep(td, 0);
901 }
902 
903 int
904 thread_unsuspend_one(struct thread *td)
905 {
906 	struct proc *p = td->td_proc;
907 
908 	PROC_SLOCK_ASSERT(p, MA_OWNED);
909 	THREAD_LOCK_ASSERT(td, MA_OWNED);
910 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
911 	TD_CLR_SUSPENDED(td);
912 	p->p_suspcount--;
913 	return (setrunnable(td));
914 }
915 
916 /*
917  * Allow all threads blocked by single threading to continue running.
918  */
919 void
920 thread_unsuspend(struct proc *p)
921 {
922 	struct thread *td;
923 	int wakeup_swapper;
924 
925 	PROC_LOCK_ASSERT(p, MA_OWNED);
926 	PROC_SLOCK_ASSERT(p, MA_OWNED);
927 	wakeup_swapper = 0;
928 	if (!P_SHOULDSTOP(p)) {
929                 FOREACH_THREAD_IN_PROC(p, td) {
930 			thread_lock(td);
931 			if (TD_IS_SUSPENDED(td)) {
932 				wakeup_swapper |= thread_unsuspend_one(td);
933 			}
934 			thread_unlock(td);
935 		}
936 	} else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
937 	    (p->p_numthreads == p->p_suspcount)) {
938 		/*
939 		 * Stopping everything also did the job for the single
940 		 * threading request. Now we've downgraded to single-threaded,
941 		 * let it continue.
942 		 */
943 		thread_lock(p->p_singlethread);
944 		wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
945 		thread_unlock(p->p_singlethread);
946 	}
947 	if (wakeup_swapper)
948 		kick_proc0();
949 }
950 
951 /*
952  * End the single threading mode..
953  */
954 void
955 thread_single_end(void)
956 {
957 	struct thread *td;
958 	struct proc *p;
959 	int wakeup_swapper;
960 
961 	td = curthread;
962 	p = td->td_proc;
963 	PROC_LOCK_ASSERT(p, MA_OWNED);
964 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
965 	PROC_SLOCK(p);
966 	p->p_singlethread = NULL;
967 	wakeup_swapper = 0;
968 	/*
969 	 * If there are other threads they may now run,
970 	 * unless of course there is a blanket 'stop order'
971 	 * on the process. The single threader must be allowed
972 	 * to continue however as this is a bad place to stop.
973 	 */
974 	if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
975                 FOREACH_THREAD_IN_PROC(p, td) {
976 			thread_lock(td);
977 			if (TD_IS_SUSPENDED(td)) {
978 				wakeup_swapper |= thread_unsuspend_one(td);
979 			}
980 			thread_unlock(td);
981 		}
982 	}
983 	PROC_SUNLOCK(p);
984 	if (wakeup_swapper)
985 		kick_proc0();
986 }
987 
988 struct thread *
989 thread_find(struct proc *p, lwpid_t tid)
990 {
991 	struct thread *td;
992 
993 	PROC_LOCK_ASSERT(p, MA_OWNED);
994 	FOREACH_THREAD_IN_PROC(p, td) {
995 		if (td->td_tid == tid)
996 			break;
997 	}
998 	return (td);
999 }
1000 
1001 /* Locate a thread by number; return with proc lock held. */
1002 struct thread *
1003 tdfind(lwpid_t tid, pid_t pid)
1004 {
1005 #define RUN_THRESH	16
1006 	struct thread *td;
1007 	int run = 0;
1008 
1009 	rw_rlock(&tidhash_lock);
1010 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1011 		if (td->td_tid == tid) {
1012 			if (pid != -1 && td->td_proc->p_pid != pid) {
1013 				td = NULL;
1014 				break;
1015 			}
1016 			PROC_LOCK(td->td_proc);
1017 			if (td->td_proc->p_state == PRS_NEW) {
1018 				PROC_UNLOCK(td->td_proc);
1019 				td = NULL;
1020 				break;
1021 			}
1022 			if (run > RUN_THRESH) {
1023 				if (rw_try_upgrade(&tidhash_lock)) {
1024 					LIST_REMOVE(td, td_hash);
1025 					LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1026 						td, td_hash);
1027 					rw_wunlock(&tidhash_lock);
1028 					return (td);
1029 				}
1030 			}
1031 			break;
1032 		}
1033 		run++;
1034 	}
1035 	rw_runlock(&tidhash_lock);
1036 	return (td);
1037 }
1038 
1039 void
1040 tidhash_add(struct thread *td)
1041 {
1042 	rw_wlock(&tidhash_lock);
1043 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1044 	rw_wunlock(&tidhash_lock);
1045 }
1046 
1047 void
1048 tidhash_remove(struct thread *td)
1049 {
1050 	rw_wlock(&tidhash_lock);
1051 	LIST_REMOVE(td, td_hash);
1052 	rw_wunlock(&tidhash_lock);
1053 }
1054