xref: /freebsd/sys/kern/kern_thread.c (revision 3e0efd2ec4fcb4cd68fb8ccf8aea6fc6151c454b)
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_kdtrace.h"
31 #include "opt_hwpmc_hooks.h"
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/proc.h>
42 #include <sys/rangelock.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sdt.h>
45 #include <sys/smp.h>
46 #include <sys/sched.h>
47 #include <sys/sleepqueue.h>
48 #include <sys/selinfo.h>
49 #include <sys/turnstile.h>
50 #include <sys/ktr.h>
51 #include <sys/rwlock.h>
52 #include <sys/umtx.h>
53 #include <sys/cpuset.h>
54 #ifdef	HWPMC_HOOKS
55 #include <sys/pmckern.h>
56 #endif
57 
58 #include <security/audit/audit.h>
59 
60 #include <vm/vm.h>
61 #include <vm/vm_extern.h>
62 #include <vm/uma.h>
63 #include <sys/eventhandler.h>
64 
65 SDT_PROVIDER_DECLARE(proc);
66 SDT_PROBE_DEFINE(proc, , , lwp_exit, lwp-exit);
67 
68 
69 /*
70  * thread related storage.
71  */
72 static uma_zone_t thread_zone;
73 
74 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
75 static struct mtx zombie_lock;
76 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
77 
78 static void thread_zombie(struct thread *);
79 
80 #define TID_BUFFER_SIZE	1024
81 
82 struct mtx tid_lock;
83 static struct unrhdr *tid_unrhdr;
84 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
85 static int tid_head, tid_tail;
86 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
87 
88 struct	tidhashhead *tidhashtbl;
89 u_long	tidhash;
90 struct	rwlock tidhash_lock;
91 
92 static lwpid_t
93 tid_alloc(void)
94 {
95 	lwpid_t	tid;
96 
97 	tid = alloc_unr(tid_unrhdr);
98 	if (tid != -1)
99 		return (tid);
100 	mtx_lock(&tid_lock);
101 	if (tid_head == tid_tail) {
102 		mtx_unlock(&tid_lock);
103 		return (-1);
104 	}
105 	tid = tid_buffer[tid_head++];
106 	tid_head %= TID_BUFFER_SIZE;
107 	mtx_unlock(&tid_lock);
108 	return (tid);
109 }
110 
111 static void
112 tid_free(lwpid_t tid)
113 {
114 	lwpid_t tmp_tid = -1;
115 
116 	mtx_lock(&tid_lock);
117 	if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
118 		tmp_tid = tid_buffer[tid_head++];
119 		tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
120 	}
121 	tid_buffer[tid_tail++] = tid;
122 	tid_tail %= TID_BUFFER_SIZE;
123 	mtx_unlock(&tid_lock);
124 	if (tmp_tid != -1)
125 		free_unr(tid_unrhdr, tmp_tid);
126 }
127 
128 /*
129  * Prepare a thread for use.
130  */
131 static int
132 thread_ctor(void *mem, int size, void *arg, int flags)
133 {
134 	struct thread	*td;
135 
136 	td = (struct thread *)mem;
137 	td->td_state = TDS_INACTIVE;
138 	td->td_oncpu = NOCPU;
139 
140 	td->td_tid = tid_alloc();
141 
142 	/*
143 	 * Note that td_critnest begins life as 1 because the thread is not
144 	 * running and is thereby implicitly waiting to be on the receiving
145 	 * end of a context switch.
146 	 */
147 	td->td_critnest = 1;
148 	td->td_lend_user_pri = PRI_MAX;
149 	EVENTHANDLER_INVOKE(thread_ctor, td);
150 #ifdef AUDIT
151 	audit_thread_alloc(td);
152 #endif
153 	umtx_thread_alloc(td);
154 	return (0);
155 }
156 
157 /*
158  * Reclaim a thread after use.
159  */
160 static void
161 thread_dtor(void *mem, int size, void *arg)
162 {
163 	struct thread *td;
164 
165 	td = (struct thread *)mem;
166 
167 #ifdef INVARIANTS
168 	/* Verify that this thread is in a safe state to free. */
169 	switch (td->td_state) {
170 	case TDS_INHIBITED:
171 	case TDS_RUNNING:
172 	case TDS_CAN_RUN:
173 	case TDS_RUNQ:
174 		/*
175 		 * We must never unlink a thread that is in one of
176 		 * these states, because it is currently active.
177 		 */
178 		panic("bad state for thread unlinking");
179 		/* NOTREACHED */
180 	case TDS_INACTIVE:
181 		break;
182 	default:
183 		panic("bad thread state");
184 		/* NOTREACHED */
185 	}
186 #endif
187 #ifdef AUDIT
188 	audit_thread_free(td);
189 #endif
190 	/* Free all OSD associated to this thread. */
191 	osd_thread_exit(td);
192 
193 	EVENTHANDLER_INVOKE(thread_dtor, td);
194 	tid_free(td->td_tid);
195 }
196 
197 /*
198  * Initialize type-stable parts of a thread (when newly created).
199  */
200 static int
201 thread_init(void *mem, int size, int flags)
202 {
203 	struct thread *td;
204 
205 	td = (struct thread *)mem;
206 
207 	td->td_sleepqueue = sleepq_alloc();
208 	td->td_turnstile = turnstile_alloc();
209 	td->td_rlqe = NULL;
210 	EVENTHANDLER_INVOKE(thread_init, td);
211 	td->td_sched = (struct td_sched *)&td[1];
212 	umtx_thread_init(td);
213 	td->td_kstack = 0;
214 	return (0);
215 }
216 
217 /*
218  * Tear down type-stable parts of a thread (just before being discarded).
219  */
220 static void
221 thread_fini(void *mem, int size)
222 {
223 	struct thread *td;
224 
225 	td = (struct thread *)mem;
226 	EVENTHANDLER_INVOKE(thread_fini, td);
227 	rlqentry_free(td->td_rlqe);
228 	turnstile_free(td->td_turnstile);
229 	sleepq_free(td->td_sleepqueue);
230 	umtx_thread_fini(td);
231 	seltdfini(td);
232 }
233 
234 /*
235  * For a newly created process,
236  * link up all the structures and its initial threads etc.
237  * called from:
238  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
239  * proc_dtor() (should go away)
240  * proc_init()
241  */
242 void
243 proc_linkup0(struct proc *p, struct thread *td)
244 {
245 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
246 	proc_linkup(p, td);
247 }
248 
249 void
250 proc_linkup(struct proc *p, struct thread *td)
251 {
252 
253 	sigqueue_init(&p->p_sigqueue, p);
254 	p->p_ksi = ksiginfo_alloc(1);
255 	if (p->p_ksi != NULL) {
256 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
257 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
258 	}
259 	LIST_INIT(&p->p_mqnotifier);
260 	p->p_numthreads = 0;
261 	thread_link(td, p);
262 }
263 
264 /*
265  * Initialize global thread allocation resources.
266  */
267 void
268 threadinit(void)
269 {
270 
271 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
272 	/* leave one number for thread0 */
273 	tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
274 
275 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
276 	    thread_ctor, thread_dtor, thread_init, thread_fini,
277 	    16 - 1, 0);
278 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
279 	rw_init(&tidhash_lock, "tidhash");
280 }
281 
282 /*
283  * Place an unused thread on the zombie list.
284  * Use the slpq as that must be unused by now.
285  */
286 void
287 thread_zombie(struct thread *td)
288 {
289 	mtx_lock_spin(&zombie_lock);
290 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
291 	mtx_unlock_spin(&zombie_lock);
292 }
293 
294 /*
295  * Release a thread that has exited after cpu_throw().
296  */
297 void
298 thread_stash(struct thread *td)
299 {
300 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
301 	thread_zombie(td);
302 }
303 
304 /*
305  * Reap zombie resources.
306  */
307 void
308 thread_reap(void)
309 {
310 	struct thread *td_first, *td_next;
311 
312 	/*
313 	 * Don't even bother to lock if none at this instant,
314 	 * we really don't care about the next instant..
315 	 */
316 	if (!TAILQ_EMPTY(&zombie_threads)) {
317 		mtx_lock_spin(&zombie_lock);
318 		td_first = TAILQ_FIRST(&zombie_threads);
319 		if (td_first)
320 			TAILQ_INIT(&zombie_threads);
321 		mtx_unlock_spin(&zombie_lock);
322 		while (td_first) {
323 			td_next = TAILQ_NEXT(td_first, td_slpq);
324 			if (td_first->td_ucred)
325 				crfree(td_first->td_ucred);
326 			thread_free(td_first);
327 			td_first = td_next;
328 		}
329 	}
330 }
331 
332 /*
333  * Allocate a thread.
334  */
335 struct thread *
336 thread_alloc(int pages)
337 {
338 	struct thread *td;
339 
340 	thread_reap(); /* check if any zombies to get */
341 
342 	td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
343 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
344 	if (!vm_thread_new(td, pages)) {
345 		uma_zfree(thread_zone, td);
346 		return (NULL);
347 	}
348 	cpu_thread_alloc(td);
349 	return (td);
350 }
351 
352 int
353 thread_alloc_stack(struct thread *td, int pages)
354 {
355 
356 	KASSERT(td->td_kstack == 0,
357 	    ("thread_alloc_stack called on a thread with kstack"));
358 	if (!vm_thread_new(td, pages))
359 		return (0);
360 	cpu_thread_alloc(td);
361 	return (1);
362 }
363 
364 /*
365  * Deallocate a thread.
366  */
367 void
368 thread_free(struct thread *td)
369 {
370 
371 	lock_profile_thread_exit(td);
372 	if (td->td_cpuset)
373 		cpuset_rel(td->td_cpuset);
374 	td->td_cpuset = NULL;
375 	cpu_thread_free(td);
376 	if (td->td_kstack != 0)
377 		vm_thread_dispose(td);
378 	uma_zfree(thread_zone, td);
379 }
380 
381 /*
382  * Discard the current thread and exit from its context.
383  * Always called with scheduler locked.
384  *
385  * Because we can't free a thread while we're operating under its context,
386  * push the current thread into our CPU's deadthread holder. This means
387  * we needn't worry about someone else grabbing our context before we
388  * do a cpu_throw().
389  */
390 void
391 thread_exit(void)
392 {
393 	uint64_t runtime, new_switchtime;
394 	struct thread *td;
395 	struct thread *td2;
396 	struct proc *p;
397 	int wakeup_swapper;
398 
399 	td = curthread;
400 	p = td->td_proc;
401 
402 	PROC_SLOCK_ASSERT(p, MA_OWNED);
403 	mtx_assert(&Giant, MA_NOTOWNED);
404 
405 	PROC_LOCK_ASSERT(p, MA_OWNED);
406 	KASSERT(p != NULL, ("thread exiting without a process"));
407 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
408 	    (long)p->p_pid, td->td_name);
409 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
410 
411 #ifdef AUDIT
412 	AUDIT_SYSCALL_EXIT(0, td);
413 #endif
414 	umtx_thread_exit(td);
415 	/*
416 	 * drop FPU & debug register state storage, or any other
417 	 * architecture specific resources that
418 	 * would not be on a new untouched process.
419 	 */
420 	cpu_thread_exit(td);	/* XXXSMP */
421 
422 	/*
423 	 * The last thread is left attached to the process
424 	 * So that the whole bundle gets recycled. Skip
425 	 * all this stuff if we never had threads.
426 	 * EXIT clears all sign of other threads when
427 	 * it goes to single threading, so the last thread always
428 	 * takes the short path.
429 	 */
430 	if (p->p_flag & P_HADTHREADS) {
431 		if (p->p_numthreads > 1) {
432 			thread_unlink(td);
433 			td2 = FIRST_THREAD_IN_PROC(p);
434 			sched_exit_thread(td2, td);
435 
436 			/*
437 			 * The test below is NOT true if we are the
438 			 * sole exiting thread. P_STOPPED_SINGLE is unset
439 			 * in exit1() after it is the only survivor.
440 			 */
441 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
442 				if (p->p_numthreads == p->p_suspcount) {
443 					thread_lock(p->p_singlethread);
444 					wakeup_swapper = thread_unsuspend_one(
445 						p->p_singlethread);
446 					thread_unlock(p->p_singlethread);
447 					if (wakeup_swapper)
448 						kick_proc0();
449 				}
450 			}
451 
452 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
453 			PCPU_SET(deadthread, td);
454 		} else {
455 			/*
456 			 * The last thread is exiting.. but not through exit()
457 			 */
458 			panic ("thread_exit: Last thread exiting on its own");
459 		}
460 	}
461 #ifdef	HWPMC_HOOKS
462 	/*
463 	 * If this thread is part of a process that is being tracked by hwpmc(4),
464 	 * inform the module of the thread's impending exit.
465 	 */
466 	if (PMC_PROC_IS_USING_PMCS(td->td_proc))
467 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
468 #endif
469 	PROC_UNLOCK(p);
470 
471 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
472 	new_switchtime = cpu_ticks();
473 	runtime = new_switchtime - PCPU_GET(switchtime);
474 	td->td_runtime += runtime;
475 	td->td_incruntime += runtime;
476 	PCPU_SET(switchtime, new_switchtime);
477 	PCPU_SET(switchticks, ticks);
478 	PCPU_INC(cnt.v_swtch);
479 
480 	/* Save our resource usage in our process. */
481 	td->td_ru.ru_nvcsw++;
482 	ruxagg(p, td);
483 	rucollect(&p->p_ru, &td->td_ru);
484 
485 	thread_lock(td);
486 	PROC_SUNLOCK(p);
487 	td->td_state = TDS_INACTIVE;
488 #ifdef WITNESS
489 	witness_thread_exit(td);
490 #endif
491 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
492 	sched_throw(td);
493 	panic("I'm a teapot!");
494 	/* NOTREACHED */
495 }
496 
497 /*
498  * Do any thread specific cleanups that may be needed in wait()
499  * called with Giant, proc and schedlock not held.
500  */
501 void
502 thread_wait(struct proc *p)
503 {
504 	struct thread *td;
505 
506 	mtx_assert(&Giant, MA_NOTOWNED);
507 	KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
508 	td = FIRST_THREAD_IN_PROC(p);
509 	/* Lock the last thread so we spin until it exits cpu_throw(). */
510 	thread_lock(td);
511 	thread_unlock(td);
512 	/* Wait for any remaining threads to exit cpu_throw(). */
513 	while (p->p_exitthreads)
514 		sched_relinquish(curthread);
515 	lock_profile_thread_exit(td);
516 	cpuset_rel(td->td_cpuset);
517 	td->td_cpuset = NULL;
518 	cpu_thread_clean(td);
519 	crfree(td->td_ucred);
520 	thread_reap();	/* check for zombie threads etc. */
521 }
522 
523 /*
524  * Link a thread to a process.
525  * set up anything that needs to be initialized for it to
526  * be used by the process.
527  */
528 void
529 thread_link(struct thread *td, struct proc *p)
530 {
531 
532 	/*
533 	 * XXX This can't be enabled because it's called for proc0 before
534 	 * its lock has been created.
535 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
536 	 */
537 	td->td_state    = TDS_INACTIVE;
538 	td->td_proc     = p;
539 	td->td_flags    = TDF_INMEM;
540 
541 	LIST_INIT(&td->td_contested);
542 	LIST_INIT(&td->td_lprof[0]);
543 	LIST_INIT(&td->td_lprof[1]);
544 	sigqueue_init(&td->td_sigqueue, p);
545 	callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
546 	TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
547 	p->p_numthreads++;
548 }
549 
550 /*
551  * Convert a process with one thread to an unthreaded process.
552  */
553 void
554 thread_unthread(struct thread *td)
555 {
556 	struct proc *p = td->td_proc;
557 
558 	KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
559 	p->p_flag &= ~P_HADTHREADS;
560 }
561 
562 /*
563  * Called from:
564  *  thread_exit()
565  */
566 void
567 thread_unlink(struct thread *td)
568 {
569 	struct proc *p = td->td_proc;
570 
571 	PROC_LOCK_ASSERT(p, MA_OWNED);
572 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
573 	p->p_numthreads--;
574 	/* could clear a few other things here */
575 	/* Must  NOT clear links to proc! */
576 }
577 
578 static int
579 calc_remaining(struct proc *p, int mode)
580 {
581 	int remaining;
582 
583 	PROC_LOCK_ASSERT(p, MA_OWNED);
584 	PROC_SLOCK_ASSERT(p, MA_OWNED);
585 	if (mode == SINGLE_EXIT)
586 		remaining = p->p_numthreads;
587 	else if (mode == SINGLE_BOUNDARY)
588 		remaining = p->p_numthreads - p->p_boundary_count;
589 	else if (mode == SINGLE_NO_EXIT)
590 		remaining = p->p_numthreads - p->p_suspcount;
591 	else
592 		panic("calc_remaining: wrong mode %d", mode);
593 	return (remaining);
594 }
595 
596 /*
597  * Enforce single-threading.
598  *
599  * Returns 1 if the caller must abort (another thread is waiting to
600  * exit the process or similar). Process is locked!
601  * Returns 0 when you are successfully the only thread running.
602  * A process has successfully single threaded in the suspend mode when
603  * There are no threads in user mode. Threads in the kernel must be
604  * allowed to continue until they get to the user boundary. They may even
605  * copy out their return values and data before suspending. They may however be
606  * accelerated in reaching the user boundary as we will wake up
607  * any sleeping threads that are interruptable. (PCATCH).
608  */
609 int
610 thread_single(int mode)
611 {
612 	struct thread *td;
613 	struct thread *td2;
614 	struct proc *p;
615 	int remaining, wakeup_swapper;
616 
617 	td = curthread;
618 	p = td->td_proc;
619 	mtx_assert(&Giant, MA_NOTOWNED);
620 	PROC_LOCK_ASSERT(p, MA_OWNED);
621 	KASSERT((td != NULL), ("curthread is NULL"));
622 
623 	if ((p->p_flag & P_HADTHREADS) == 0)
624 		return (0);
625 
626 	/* Is someone already single threading? */
627 	if (p->p_singlethread != NULL && p->p_singlethread != td)
628 		return (1);
629 
630 	if (mode == SINGLE_EXIT) {
631 		p->p_flag |= P_SINGLE_EXIT;
632 		p->p_flag &= ~P_SINGLE_BOUNDARY;
633 	} else {
634 		p->p_flag &= ~P_SINGLE_EXIT;
635 		if (mode == SINGLE_BOUNDARY)
636 			p->p_flag |= P_SINGLE_BOUNDARY;
637 		else
638 			p->p_flag &= ~P_SINGLE_BOUNDARY;
639 	}
640 	p->p_flag |= P_STOPPED_SINGLE;
641 	PROC_SLOCK(p);
642 	p->p_singlethread = td;
643 	remaining = calc_remaining(p, mode);
644 	while (remaining != 1) {
645 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
646 			goto stopme;
647 		wakeup_swapper = 0;
648 		FOREACH_THREAD_IN_PROC(p, td2) {
649 			if (td2 == td)
650 				continue;
651 			thread_lock(td2);
652 			td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
653 			if (TD_IS_INHIBITED(td2)) {
654 				switch (mode) {
655 				case SINGLE_EXIT:
656 					if (TD_IS_SUSPENDED(td2))
657 						wakeup_swapper |=
658 						    thread_unsuspend_one(td2);
659 					if (TD_ON_SLEEPQ(td2) &&
660 					    (td2->td_flags & TDF_SINTR))
661 						wakeup_swapper |=
662 						    sleepq_abort(td2, EINTR);
663 					break;
664 				case SINGLE_BOUNDARY:
665 					if (TD_IS_SUSPENDED(td2) &&
666 					    !(td2->td_flags & TDF_BOUNDARY))
667 						wakeup_swapper |=
668 						    thread_unsuspend_one(td2);
669 					if (TD_ON_SLEEPQ(td2) &&
670 					    (td2->td_flags & TDF_SINTR))
671 						wakeup_swapper |=
672 						    sleepq_abort(td2, ERESTART);
673 					break;
674 				case SINGLE_NO_EXIT:
675 					if (TD_IS_SUSPENDED(td2) &&
676 					    !(td2->td_flags & TDF_BOUNDARY))
677 						wakeup_swapper |=
678 						    thread_unsuspend_one(td2);
679 					if (TD_ON_SLEEPQ(td2) &&
680 					    (td2->td_flags & TDF_SINTR))
681 						wakeup_swapper |=
682 						    sleepq_abort(td2, ERESTART);
683 					break;
684 				default:
685 					break;
686 				}
687 			}
688 #ifdef SMP
689 			else if (TD_IS_RUNNING(td2) && td != td2) {
690 				forward_signal(td2);
691 			}
692 #endif
693 			thread_unlock(td2);
694 		}
695 		if (wakeup_swapper)
696 			kick_proc0();
697 		remaining = calc_remaining(p, mode);
698 
699 		/*
700 		 * Maybe we suspended some threads.. was it enough?
701 		 */
702 		if (remaining == 1)
703 			break;
704 
705 stopme:
706 		/*
707 		 * Wake us up when everyone else has suspended.
708 		 * In the mean time we suspend as well.
709 		 */
710 		thread_suspend_switch(td);
711 		remaining = calc_remaining(p, mode);
712 	}
713 	if (mode == SINGLE_EXIT) {
714 		/*
715 		 * We have gotten rid of all the other threads and we
716 		 * are about to either exit or exec. In either case,
717 		 * we try our utmost  to revert to being a non-threaded
718 		 * process.
719 		 */
720 		p->p_singlethread = NULL;
721 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
722 		thread_unthread(td);
723 	}
724 	PROC_SUNLOCK(p);
725 	return (0);
726 }
727 
728 /*
729  * Called in from locations that can safely check to see
730  * whether we have to suspend or at least throttle for a
731  * single-thread event (e.g. fork).
732  *
733  * Such locations include userret().
734  * If the "return_instead" argument is non zero, the thread must be able to
735  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
736  *
737  * The 'return_instead' argument tells the function if it may do a
738  * thread_exit() or suspend, or whether the caller must abort and back
739  * out instead.
740  *
741  * If the thread that set the single_threading request has set the
742  * P_SINGLE_EXIT bit in the process flags then this call will never return
743  * if 'return_instead' is false, but will exit.
744  *
745  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
746  *---------------+--------------------+---------------------
747  *       0       | returns 0          |   returns 0 or 1
748  *               | when ST ends       |   immediatly
749  *---------------+--------------------+---------------------
750  *       1       | thread exits       |   returns 1
751  *               |                    |  immediatly
752  * 0 = thread_exit() or suspension ok,
753  * other = return error instead of stopping the thread.
754  *
755  * While a full suspension is under effect, even a single threading
756  * thread would be suspended if it made this call (but it shouldn't).
757  * This call should only be made from places where
758  * thread_exit() would be safe as that may be the outcome unless
759  * return_instead is set.
760  */
761 int
762 thread_suspend_check(int return_instead)
763 {
764 	struct thread *td;
765 	struct proc *p;
766 	int wakeup_swapper;
767 
768 	td = curthread;
769 	p = td->td_proc;
770 	mtx_assert(&Giant, MA_NOTOWNED);
771 	PROC_LOCK_ASSERT(p, MA_OWNED);
772 	while (P_SHOULDSTOP(p) ||
773 	      ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
774 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
775 			KASSERT(p->p_singlethread != NULL,
776 			    ("singlethread not set"));
777 			/*
778 			 * The only suspension in action is a
779 			 * single-threading. Single threader need not stop.
780 			 * XXX Should be safe to access unlocked
781 			 * as it can only be set to be true by us.
782 			 */
783 			if (p->p_singlethread == td)
784 				return (0);	/* Exempt from stopping. */
785 		}
786 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
787 			return (EINTR);
788 
789 		/* Should we goto user boundary if we didn't come from there? */
790 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
791 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
792 			return (ERESTART);
793 
794 		/*
795 		 * If the process is waiting for us to exit,
796 		 * this thread should just suicide.
797 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
798 		 */
799 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
800 			PROC_UNLOCK(p);
801 			tidhash_remove(td);
802 			PROC_LOCK(p);
803 			tdsigcleanup(td);
804 			PROC_SLOCK(p);
805 			thread_stopped(p);
806 			thread_exit();
807 		}
808 
809 		PROC_SLOCK(p);
810 		thread_stopped(p);
811 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
812 			if (p->p_numthreads == p->p_suspcount + 1) {
813 				thread_lock(p->p_singlethread);
814 				wakeup_swapper =
815 				    thread_unsuspend_one(p->p_singlethread);
816 				thread_unlock(p->p_singlethread);
817 				if (wakeup_swapper)
818 					kick_proc0();
819 			}
820 		}
821 		PROC_UNLOCK(p);
822 		thread_lock(td);
823 		/*
824 		 * When a thread suspends, it just
825 		 * gets taken off all queues.
826 		 */
827 		thread_suspend_one(td);
828 		if (return_instead == 0) {
829 			p->p_boundary_count++;
830 			td->td_flags |= TDF_BOUNDARY;
831 		}
832 		PROC_SUNLOCK(p);
833 		mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
834 		if (return_instead == 0)
835 			td->td_flags &= ~TDF_BOUNDARY;
836 		thread_unlock(td);
837 		PROC_LOCK(p);
838 		if (return_instead == 0) {
839 			PROC_SLOCK(p);
840 			p->p_boundary_count--;
841 			PROC_SUNLOCK(p);
842 		}
843 	}
844 	return (0);
845 }
846 
847 void
848 thread_suspend_switch(struct thread *td)
849 {
850 	struct proc *p;
851 
852 	p = td->td_proc;
853 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
854 	PROC_LOCK_ASSERT(p, MA_OWNED);
855 	PROC_SLOCK_ASSERT(p, MA_OWNED);
856 	/*
857 	 * We implement thread_suspend_one in stages here to avoid
858 	 * dropping the proc lock while the thread lock is owned.
859 	 */
860 	thread_stopped(p);
861 	p->p_suspcount++;
862 	PROC_UNLOCK(p);
863 	thread_lock(td);
864 	td->td_flags &= ~TDF_NEEDSUSPCHK;
865 	TD_SET_SUSPENDED(td);
866 	sched_sleep(td, 0);
867 	PROC_SUNLOCK(p);
868 	DROP_GIANT();
869 	mi_switch(SW_VOL | SWT_SUSPEND, NULL);
870 	thread_unlock(td);
871 	PICKUP_GIANT();
872 	PROC_LOCK(p);
873 	PROC_SLOCK(p);
874 }
875 
876 void
877 thread_suspend_one(struct thread *td)
878 {
879 	struct proc *p = td->td_proc;
880 
881 	PROC_SLOCK_ASSERT(p, MA_OWNED);
882 	THREAD_LOCK_ASSERT(td, MA_OWNED);
883 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
884 	p->p_suspcount++;
885 	td->td_flags &= ~TDF_NEEDSUSPCHK;
886 	TD_SET_SUSPENDED(td);
887 	sched_sleep(td, 0);
888 }
889 
890 int
891 thread_unsuspend_one(struct thread *td)
892 {
893 	struct proc *p = td->td_proc;
894 
895 	PROC_SLOCK_ASSERT(p, MA_OWNED);
896 	THREAD_LOCK_ASSERT(td, MA_OWNED);
897 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
898 	TD_CLR_SUSPENDED(td);
899 	p->p_suspcount--;
900 	return (setrunnable(td));
901 }
902 
903 /*
904  * Allow all threads blocked by single threading to continue running.
905  */
906 void
907 thread_unsuspend(struct proc *p)
908 {
909 	struct thread *td;
910 	int wakeup_swapper;
911 
912 	PROC_LOCK_ASSERT(p, MA_OWNED);
913 	PROC_SLOCK_ASSERT(p, MA_OWNED);
914 	wakeup_swapper = 0;
915 	if (!P_SHOULDSTOP(p)) {
916                 FOREACH_THREAD_IN_PROC(p, td) {
917 			thread_lock(td);
918 			if (TD_IS_SUSPENDED(td)) {
919 				wakeup_swapper |= thread_unsuspend_one(td);
920 			}
921 			thread_unlock(td);
922 		}
923 	} else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
924 	    (p->p_numthreads == p->p_suspcount)) {
925 		/*
926 		 * Stopping everything also did the job for the single
927 		 * threading request. Now we've downgraded to single-threaded,
928 		 * let it continue.
929 		 */
930 		thread_lock(p->p_singlethread);
931 		wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
932 		thread_unlock(p->p_singlethread);
933 	}
934 	if (wakeup_swapper)
935 		kick_proc0();
936 }
937 
938 /*
939  * End the single threading mode..
940  */
941 void
942 thread_single_end(void)
943 {
944 	struct thread *td;
945 	struct proc *p;
946 	int wakeup_swapper;
947 
948 	td = curthread;
949 	p = td->td_proc;
950 	PROC_LOCK_ASSERT(p, MA_OWNED);
951 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
952 	PROC_SLOCK(p);
953 	p->p_singlethread = NULL;
954 	wakeup_swapper = 0;
955 	/*
956 	 * If there are other threads they may now run,
957 	 * unless of course there is a blanket 'stop order'
958 	 * on the process. The single threader must be allowed
959 	 * to continue however as this is a bad place to stop.
960 	 */
961 	if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
962                 FOREACH_THREAD_IN_PROC(p, td) {
963 			thread_lock(td);
964 			if (TD_IS_SUSPENDED(td)) {
965 				wakeup_swapper |= thread_unsuspend_one(td);
966 			}
967 			thread_unlock(td);
968 		}
969 	}
970 	PROC_SUNLOCK(p);
971 	if (wakeup_swapper)
972 		kick_proc0();
973 }
974 
975 struct thread *
976 thread_find(struct proc *p, lwpid_t tid)
977 {
978 	struct thread *td;
979 
980 	PROC_LOCK_ASSERT(p, MA_OWNED);
981 	FOREACH_THREAD_IN_PROC(p, td) {
982 		if (td->td_tid == tid)
983 			break;
984 	}
985 	return (td);
986 }
987 
988 /* Locate a thread by number; return with proc lock held. */
989 struct thread *
990 tdfind(lwpid_t tid, pid_t pid)
991 {
992 #define RUN_THRESH	16
993 	struct thread *td;
994 	int run = 0;
995 
996 	rw_rlock(&tidhash_lock);
997 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
998 		if (td->td_tid == tid) {
999 			if (pid != -1 && td->td_proc->p_pid != pid) {
1000 				td = NULL;
1001 				break;
1002 			}
1003 			PROC_LOCK(td->td_proc);
1004 			if (td->td_proc->p_state == PRS_NEW) {
1005 				PROC_UNLOCK(td->td_proc);
1006 				td = NULL;
1007 				break;
1008 			}
1009 			if (run > RUN_THRESH) {
1010 				if (rw_try_upgrade(&tidhash_lock)) {
1011 					LIST_REMOVE(td, td_hash);
1012 					LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1013 						td, td_hash);
1014 					rw_wunlock(&tidhash_lock);
1015 					return (td);
1016 				}
1017 			}
1018 			break;
1019 		}
1020 		run++;
1021 	}
1022 	rw_runlock(&tidhash_lock);
1023 	return (td);
1024 }
1025 
1026 void
1027 tidhash_add(struct thread *td)
1028 {
1029 	rw_wlock(&tidhash_lock);
1030 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1031 	rw_wunlock(&tidhash_lock);
1032 }
1033 
1034 void
1035 tidhash_remove(struct thread *td)
1036 {
1037 	rw_wlock(&tidhash_lock);
1038 	LIST_REMOVE(td, td_hash);
1039 	rw_wunlock(&tidhash_lock);
1040 }
1041