xref: /freebsd/sys/kern/kern_thread.c (revision f05cddf9)
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_head + 1) % 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_tail + 1) % 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 
273 	/*
274 	 * pid_max cannot be greater than PID_MAX.
275 	 * leave one number for thread0.
276 	 */
277 	tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
278 
279 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
280 	    thread_ctor, thread_dtor, thread_init, thread_fini,
281 	    16 - 1, 0);
282 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
283 	rw_init(&tidhash_lock, "tidhash");
284 }
285 
286 /*
287  * Place an unused thread on the zombie list.
288  * Use the slpq as that must be unused by now.
289  */
290 void
291 thread_zombie(struct thread *td)
292 {
293 	mtx_lock_spin(&zombie_lock);
294 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
295 	mtx_unlock_spin(&zombie_lock);
296 }
297 
298 /*
299  * Release a thread that has exited after cpu_throw().
300  */
301 void
302 thread_stash(struct thread *td)
303 {
304 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
305 	thread_zombie(td);
306 }
307 
308 /*
309  * Reap zombie resources.
310  */
311 void
312 thread_reap(void)
313 {
314 	struct thread *td_first, *td_next;
315 
316 	/*
317 	 * Don't even bother to lock if none at this instant,
318 	 * we really don't care about the next instant..
319 	 */
320 	if (!TAILQ_EMPTY(&zombie_threads)) {
321 		mtx_lock_spin(&zombie_lock);
322 		td_first = TAILQ_FIRST(&zombie_threads);
323 		if (td_first)
324 			TAILQ_INIT(&zombie_threads);
325 		mtx_unlock_spin(&zombie_lock);
326 		while (td_first) {
327 			td_next = TAILQ_NEXT(td_first, td_slpq);
328 			if (td_first->td_ucred)
329 				crfree(td_first->td_ucred);
330 			thread_free(td_first);
331 			td_first = td_next;
332 		}
333 	}
334 }
335 
336 /*
337  * Allocate a thread.
338  */
339 struct thread *
340 thread_alloc(int pages)
341 {
342 	struct thread *td;
343 
344 	thread_reap(); /* check if any zombies to get */
345 
346 	td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
347 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
348 	if (!vm_thread_new(td, pages)) {
349 		uma_zfree(thread_zone, td);
350 		return (NULL);
351 	}
352 	cpu_thread_alloc(td);
353 	return (td);
354 }
355 
356 int
357 thread_alloc_stack(struct thread *td, int pages)
358 {
359 
360 	KASSERT(td->td_kstack == 0,
361 	    ("thread_alloc_stack called on a thread with kstack"));
362 	if (!vm_thread_new(td, pages))
363 		return (0);
364 	cpu_thread_alloc(td);
365 	return (1);
366 }
367 
368 /*
369  * Deallocate a thread.
370  */
371 void
372 thread_free(struct thread *td)
373 {
374 
375 	lock_profile_thread_exit(td);
376 	if (td->td_cpuset)
377 		cpuset_rel(td->td_cpuset);
378 	td->td_cpuset = NULL;
379 	cpu_thread_free(td);
380 	if (td->td_kstack != 0)
381 		vm_thread_dispose(td);
382 	uma_zfree(thread_zone, td);
383 }
384 
385 /*
386  * Discard the current thread and exit from its context.
387  * Always called with scheduler locked.
388  *
389  * Because we can't free a thread while we're operating under its context,
390  * push the current thread into our CPU's deadthread holder. This means
391  * we needn't worry about someone else grabbing our context before we
392  * do a cpu_throw().
393  */
394 void
395 thread_exit(void)
396 {
397 	uint64_t runtime, new_switchtime;
398 	struct thread *td;
399 	struct thread *td2;
400 	struct proc *p;
401 	int wakeup_swapper;
402 
403 	td = curthread;
404 	p = td->td_proc;
405 
406 	PROC_SLOCK_ASSERT(p, MA_OWNED);
407 	mtx_assert(&Giant, MA_NOTOWNED);
408 
409 	PROC_LOCK_ASSERT(p, MA_OWNED);
410 	KASSERT(p != NULL, ("thread exiting without a process"));
411 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
412 	    (long)p->p_pid, td->td_name);
413 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
414 
415 #ifdef AUDIT
416 	AUDIT_SYSCALL_EXIT(0, td);
417 #endif
418 	umtx_thread_exit(td);
419 	/*
420 	 * drop FPU & debug register state storage, or any other
421 	 * architecture specific resources that
422 	 * would not be on a new untouched process.
423 	 */
424 	cpu_thread_exit(td);	/* XXXSMP */
425 
426 	/*
427 	 * The last thread is left attached to the process
428 	 * So that the whole bundle gets recycled. Skip
429 	 * all this stuff if we never had threads.
430 	 * EXIT clears all sign of other threads when
431 	 * it goes to single threading, so the last thread always
432 	 * takes the short path.
433 	 */
434 	if (p->p_flag & P_HADTHREADS) {
435 		if (p->p_numthreads > 1) {
436 			thread_unlink(td);
437 			td2 = FIRST_THREAD_IN_PROC(p);
438 			sched_exit_thread(td2, td);
439 
440 			/*
441 			 * The test below is NOT true if we are the
442 			 * sole exiting thread. P_STOPPED_SINGLE is unset
443 			 * in exit1() after it is the only survivor.
444 			 */
445 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
446 				if (p->p_numthreads == p->p_suspcount) {
447 					thread_lock(p->p_singlethread);
448 					wakeup_swapper = thread_unsuspend_one(
449 						p->p_singlethread);
450 					thread_unlock(p->p_singlethread);
451 					if (wakeup_swapper)
452 						kick_proc0();
453 				}
454 			}
455 
456 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
457 			PCPU_SET(deadthread, td);
458 		} else {
459 			/*
460 			 * The last thread is exiting.. but not through exit()
461 			 */
462 			panic ("thread_exit: Last thread exiting on its own");
463 		}
464 	}
465 #ifdef	HWPMC_HOOKS
466 	/*
467 	 * If this thread is part of a process that is being tracked by hwpmc(4),
468 	 * inform the module of the thread's impending exit.
469 	 */
470 	if (PMC_PROC_IS_USING_PMCS(td->td_proc))
471 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
472 #endif
473 	PROC_UNLOCK(p);
474 
475 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
476 	new_switchtime = cpu_ticks();
477 	runtime = new_switchtime - PCPU_GET(switchtime);
478 	td->td_runtime += runtime;
479 	td->td_incruntime += runtime;
480 	PCPU_SET(switchtime, new_switchtime);
481 	PCPU_SET(switchticks, ticks);
482 	PCPU_INC(cnt.v_swtch);
483 
484 	/* Save our resource usage in our process. */
485 	td->td_ru.ru_nvcsw++;
486 	ruxagg(p, td);
487 	rucollect(&p->p_ru, &td->td_ru);
488 
489 	thread_lock(td);
490 	PROC_SUNLOCK(p);
491 	td->td_state = TDS_INACTIVE;
492 #ifdef WITNESS
493 	witness_thread_exit(td);
494 #endif
495 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
496 	sched_throw(td);
497 	panic("I'm a teapot!");
498 	/* NOTREACHED */
499 }
500 
501 /*
502  * Do any thread specific cleanups that may be needed in wait()
503  * called with Giant, proc and schedlock not held.
504  */
505 void
506 thread_wait(struct proc *p)
507 {
508 	struct thread *td;
509 
510 	mtx_assert(&Giant, MA_NOTOWNED);
511 	KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
512 	td = FIRST_THREAD_IN_PROC(p);
513 	/* Lock the last thread so we spin until it exits cpu_throw(). */
514 	thread_lock(td);
515 	thread_unlock(td);
516 	/* Wait for any remaining threads to exit cpu_throw(). */
517 	while (p->p_exitthreads)
518 		sched_relinquish(curthread);
519 	lock_profile_thread_exit(td);
520 	cpuset_rel(td->td_cpuset);
521 	td->td_cpuset = NULL;
522 	cpu_thread_clean(td);
523 	crfree(td->td_ucred);
524 	thread_reap();	/* check for zombie threads etc. */
525 }
526 
527 /*
528  * Link a thread to a process.
529  * set up anything that needs to be initialized for it to
530  * be used by the process.
531  */
532 void
533 thread_link(struct thread *td, struct proc *p)
534 {
535 
536 	/*
537 	 * XXX This can't be enabled because it's called for proc0 before
538 	 * its lock has been created.
539 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
540 	 */
541 	td->td_state    = TDS_INACTIVE;
542 	td->td_proc     = p;
543 	td->td_flags    = TDF_INMEM;
544 
545 	LIST_INIT(&td->td_contested);
546 	LIST_INIT(&td->td_lprof[0]);
547 	LIST_INIT(&td->td_lprof[1]);
548 	sigqueue_init(&td->td_sigqueue, p);
549 	callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
550 	TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
551 	p->p_numthreads++;
552 }
553 
554 /*
555  * Convert a process with one thread to an unthreaded process.
556  */
557 void
558 thread_unthread(struct thread *td)
559 {
560 	struct proc *p = td->td_proc;
561 
562 	KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
563 	p->p_flag &= ~P_HADTHREADS;
564 }
565 
566 /*
567  * Called from:
568  *  thread_exit()
569  */
570 void
571 thread_unlink(struct thread *td)
572 {
573 	struct proc *p = td->td_proc;
574 
575 	PROC_LOCK_ASSERT(p, MA_OWNED);
576 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
577 	p->p_numthreads--;
578 	/* could clear a few other things here */
579 	/* Must  NOT clear links to proc! */
580 }
581 
582 static int
583 calc_remaining(struct proc *p, int mode)
584 {
585 	int remaining;
586 
587 	PROC_LOCK_ASSERT(p, MA_OWNED);
588 	PROC_SLOCK_ASSERT(p, MA_OWNED);
589 	if (mode == SINGLE_EXIT)
590 		remaining = p->p_numthreads;
591 	else if (mode == SINGLE_BOUNDARY)
592 		remaining = p->p_numthreads - p->p_boundary_count;
593 	else if (mode == SINGLE_NO_EXIT)
594 		remaining = p->p_numthreads - p->p_suspcount;
595 	else
596 		panic("calc_remaining: wrong mode %d", mode);
597 	return (remaining);
598 }
599 
600 /*
601  * Enforce single-threading.
602  *
603  * Returns 1 if the caller must abort (another thread is waiting to
604  * exit the process or similar). Process is locked!
605  * Returns 0 when you are successfully the only thread running.
606  * A process has successfully single threaded in the suspend mode when
607  * There are no threads in user mode. Threads in the kernel must be
608  * allowed to continue until they get to the user boundary. They may even
609  * copy out their return values and data before suspending. They may however be
610  * accelerated in reaching the user boundary as we will wake up
611  * any sleeping threads that are interruptable. (PCATCH).
612  */
613 int
614 thread_single(int mode)
615 {
616 	struct thread *td;
617 	struct thread *td2;
618 	struct proc *p;
619 	int remaining, wakeup_swapper;
620 
621 	td = curthread;
622 	p = td->td_proc;
623 	mtx_assert(&Giant, MA_NOTOWNED);
624 	PROC_LOCK_ASSERT(p, MA_OWNED);
625 
626 	if ((p->p_flag & P_HADTHREADS) == 0)
627 		return (0);
628 
629 	/* Is someone already single threading? */
630 	if (p->p_singlethread != NULL && p->p_singlethread != td)
631 		return (1);
632 
633 	if (mode == SINGLE_EXIT) {
634 		p->p_flag |= P_SINGLE_EXIT;
635 		p->p_flag &= ~P_SINGLE_BOUNDARY;
636 	} else {
637 		p->p_flag &= ~P_SINGLE_EXIT;
638 		if (mode == SINGLE_BOUNDARY)
639 			p->p_flag |= P_SINGLE_BOUNDARY;
640 		else
641 			p->p_flag &= ~P_SINGLE_BOUNDARY;
642 	}
643 	p->p_flag |= P_STOPPED_SINGLE;
644 	PROC_SLOCK(p);
645 	p->p_singlethread = td;
646 	remaining = calc_remaining(p, mode);
647 	while (remaining != 1) {
648 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
649 			goto stopme;
650 		wakeup_swapper = 0;
651 		FOREACH_THREAD_IN_PROC(p, td2) {
652 			if (td2 == td)
653 				continue;
654 			thread_lock(td2);
655 			td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
656 			if (TD_IS_INHIBITED(td2)) {
657 				switch (mode) {
658 				case SINGLE_EXIT:
659 					if (TD_IS_SUSPENDED(td2))
660 						wakeup_swapper |=
661 						    thread_unsuspend_one(td2);
662 					if (TD_ON_SLEEPQ(td2) &&
663 					    (td2->td_flags & TDF_SINTR))
664 						wakeup_swapper |=
665 						    sleepq_abort(td2, EINTR);
666 					break;
667 				case SINGLE_BOUNDARY:
668 					if (TD_IS_SUSPENDED(td2) &&
669 					    !(td2->td_flags & TDF_BOUNDARY))
670 						wakeup_swapper |=
671 						    thread_unsuspend_one(td2);
672 					if (TD_ON_SLEEPQ(td2) &&
673 					    (td2->td_flags & TDF_SINTR))
674 						wakeup_swapper |=
675 						    sleepq_abort(td2, ERESTART);
676 					break;
677 				case SINGLE_NO_EXIT:
678 					if (TD_IS_SUSPENDED(td2) &&
679 					    !(td2->td_flags & TDF_BOUNDARY))
680 						wakeup_swapper |=
681 						    thread_unsuspend_one(td2);
682 					if (TD_ON_SLEEPQ(td2) &&
683 					    (td2->td_flags & TDF_SINTR))
684 						wakeup_swapper |=
685 						    sleepq_abort(td2, ERESTART);
686 					break;
687 				default:
688 					break;
689 				}
690 			}
691 #ifdef SMP
692 			else if (TD_IS_RUNNING(td2) && td != td2) {
693 				forward_signal(td2);
694 			}
695 #endif
696 			thread_unlock(td2);
697 		}
698 		if (wakeup_swapper)
699 			kick_proc0();
700 		remaining = calc_remaining(p, mode);
701 
702 		/*
703 		 * Maybe we suspended some threads.. was it enough?
704 		 */
705 		if (remaining == 1)
706 			break;
707 
708 stopme:
709 		/*
710 		 * Wake us up when everyone else has suspended.
711 		 * In the mean time we suspend as well.
712 		 */
713 		thread_suspend_switch(td);
714 		remaining = calc_remaining(p, mode);
715 	}
716 	if (mode == SINGLE_EXIT) {
717 		/*
718 		 * We have gotten rid of all the other threads and we
719 		 * are about to either exit or exec. In either case,
720 		 * we try our utmost to revert to being a non-threaded
721 		 * process.
722 		 */
723 		p->p_singlethread = NULL;
724 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
725 		thread_unthread(td);
726 	}
727 	PROC_SUNLOCK(p);
728 	return (0);
729 }
730 
731 /*
732  * Called in from locations that can safely check to see
733  * whether we have to suspend or at least throttle for a
734  * single-thread event (e.g. fork).
735  *
736  * Such locations include userret().
737  * If the "return_instead" argument is non zero, the thread must be able to
738  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
739  *
740  * The 'return_instead' argument tells the function if it may do a
741  * thread_exit() or suspend, or whether the caller must abort and back
742  * out instead.
743  *
744  * If the thread that set the single_threading request has set the
745  * P_SINGLE_EXIT bit in the process flags then this call will never return
746  * if 'return_instead' is false, but will exit.
747  *
748  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
749  *---------------+--------------------+---------------------
750  *       0       | returns 0          |   returns 0 or 1
751  *               | when ST ends       |   immediately
752  *---------------+--------------------+---------------------
753  *       1       | thread exits       |   returns 1
754  *               |                    |  immediately
755  * 0 = thread_exit() or suspension ok,
756  * other = return error instead of stopping the thread.
757  *
758  * While a full suspension is under effect, even a single threading
759  * thread would be suspended if it made this call (but it shouldn't).
760  * This call should only be made from places where
761  * thread_exit() would be safe as that may be the outcome unless
762  * return_instead is set.
763  */
764 int
765 thread_suspend_check(int return_instead)
766 {
767 	struct thread *td;
768 	struct proc *p;
769 	int wakeup_swapper;
770 
771 	td = curthread;
772 	p = td->td_proc;
773 	mtx_assert(&Giant, MA_NOTOWNED);
774 	PROC_LOCK_ASSERT(p, MA_OWNED);
775 	while (P_SHOULDSTOP(p) ||
776 	      ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
777 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
778 			KASSERT(p->p_singlethread != NULL,
779 			    ("singlethread not set"));
780 			/*
781 			 * The only suspension in action is a
782 			 * single-threading. Single threader need not stop.
783 			 * XXX Should be safe to access unlocked
784 			 * as it can only be set to be true by us.
785 			 */
786 			if (p->p_singlethread == td)
787 				return (0);	/* Exempt from stopping. */
788 		}
789 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
790 			return (EINTR);
791 
792 		/* Should we goto user boundary if we didn't come from there? */
793 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
794 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
795 			return (ERESTART);
796 
797 		/*
798 		 * Ignore suspend requests for stop signals if they
799 		 * are deferred.
800 		 */
801 		if (P_SHOULDSTOP(p) == P_STOPPED_SIG &&
802 		    td->td_flags & TDF_SBDRY) {
803 			KASSERT(return_instead,
804 			    ("TDF_SBDRY set for unsafe thread_suspend_check"));
805 			return (0);
806 		}
807 
808 		/*
809 		 * If the process is waiting for us to exit,
810 		 * this thread should just suicide.
811 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
812 		 */
813 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
814 			PROC_UNLOCK(p);
815 			tidhash_remove(td);
816 			PROC_LOCK(p);
817 			tdsigcleanup(td);
818 			PROC_SLOCK(p);
819 			thread_stopped(p);
820 			thread_exit();
821 		}
822 
823 		PROC_SLOCK(p);
824 		thread_stopped(p);
825 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
826 			if (p->p_numthreads == p->p_suspcount + 1) {
827 				thread_lock(p->p_singlethread);
828 				wakeup_swapper =
829 				    thread_unsuspend_one(p->p_singlethread);
830 				thread_unlock(p->p_singlethread);
831 				if (wakeup_swapper)
832 					kick_proc0();
833 			}
834 		}
835 		PROC_UNLOCK(p);
836 		thread_lock(td);
837 		/*
838 		 * When a thread suspends, it just
839 		 * gets taken off all queues.
840 		 */
841 		thread_suspend_one(td);
842 		if (return_instead == 0) {
843 			p->p_boundary_count++;
844 			td->td_flags |= TDF_BOUNDARY;
845 		}
846 		PROC_SUNLOCK(p);
847 		mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
848 		if (return_instead == 0)
849 			td->td_flags &= ~TDF_BOUNDARY;
850 		thread_unlock(td);
851 		PROC_LOCK(p);
852 		if (return_instead == 0) {
853 			PROC_SLOCK(p);
854 			p->p_boundary_count--;
855 			PROC_SUNLOCK(p);
856 		}
857 	}
858 	return (0);
859 }
860 
861 void
862 thread_suspend_switch(struct thread *td)
863 {
864 	struct proc *p;
865 
866 	p = td->td_proc;
867 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
868 	PROC_LOCK_ASSERT(p, MA_OWNED);
869 	PROC_SLOCK_ASSERT(p, MA_OWNED);
870 	/*
871 	 * We implement thread_suspend_one in stages here to avoid
872 	 * dropping the proc lock while the thread lock is owned.
873 	 */
874 	thread_stopped(p);
875 	p->p_suspcount++;
876 	PROC_UNLOCK(p);
877 	thread_lock(td);
878 	td->td_flags &= ~TDF_NEEDSUSPCHK;
879 	TD_SET_SUSPENDED(td);
880 	sched_sleep(td, 0);
881 	PROC_SUNLOCK(p);
882 	DROP_GIANT();
883 	mi_switch(SW_VOL | SWT_SUSPEND, NULL);
884 	thread_unlock(td);
885 	PICKUP_GIANT();
886 	PROC_LOCK(p);
887 	PROC_SLOCK(p);
888 }
889 
890 void
891 thread_suspend_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), ("already suspended"));
898 	p->p_suspcount++;
899 	td->td_flags &= ~TDF_NEEDSUSPCHK;
900 	TD_SET_SUSPENDED(td);
901 	sched_sleep(td, 0);
902 }
903 
904 int
905 thread_unsuspend_one(struct thread *td)
906 {
907 	struct proc *p = td->td_proc;
908 
909 	PROC_SLOCK_ASSERT(p, MA_OWNED);
910 	THREAD_LOCK_ASSERT(td, MA_OWNED);
911 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
912 	TD_CLR_SUSPENDED(td);
913 	p->p_suspcount--;
914 	return (setrunnable(td));
915 }
916 
917 /*
918  * Allow all threads blocked by single threading to continue running.
919  */
920 void
921 thread_unsuspend(struct proc *p)
922 {
923 	struct thread *td;
924 	int wakeup_swapper;
925 
926 	PROC_LOCK_ASSERT(p, MA_OWNED);
927 	PROC_SLOCK_ASSERT(p, MA_OWNED);
928 	wakeup_swapper = 0;
929 	if (!P_SHOULDSTOP(p)) {
930                 FOREACH_THREAD_IN_PROC(p, td) {
931 			thread_lock(td);
932 			if (TD_IS_SUSPENDED(td)) {
933 				wakeup_swapper |= thread_unsuspend_one(td);
934 			}
935 			thread_unlock(td);
936 		}
937 	} else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
938 	    (p->p_numthreads == p->p_suspcount)) {
939 		/*
940 		 * Stopping everything also did the job for the single
941 		 * threading request. Now we've downgraded to single-threaded,
942 		 * let it continue.
943 		 */
944 		thread_lock(p->p_singlethread);
945 		wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
946 		thread_unlock(p->p_singlethread);
947 	}
948 	if (wakeup_swapper)
949 		kick_proc0();
950 }
951 
952 /*
953  * End the single threading mode..
954  */
955 void
956 thread_single_end(void)
957 {
958 	struct thread *td;
959 	struct proc *p;
960 	int wakeup_swapper;
961 
962 	td = curthread;
963 	p = td->td_proc;
964 	PROC_LOCK_ASSERT(p, MA_OWNED);
965 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
966 	PROC_SLOCK(p);
967 	p->p_singlethread = NULL;
968 	wakeup_swapper = 0;
969 	/*
970 	 * If there are other threads they may now run,
971 	 * unless of course there is a blanket 'stop order'
972 	 * on the process. The single threader must be allowed
973 	 * to continue however as this is a bad place to stop.
974 	 */
975 	if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
976                 FOREACH_THREAD_IN_PROC(p, td) {
977 			thread_lock(td);
978 			if (TD_IS_SUSPENDED(td)) {
979 				wakeup_swapper |= thread_unsuspend_one(td);
980 			}
981 			thread_unlock(td);
982 		}
983 	}
984 	PROC_SUNLOCK(p);
985 	if (wakeup_swapper)
986 		kick_proc0();
987 }
988 
989 struct thread *
990 thread_find(struct proc *p, lwpid_t tid)
991 {
992 	struct thread *td;
993 
994 	PROC_LOCK_ASSERT(p, MA_OWNED);
995 	FOREACH_THREAD_IN_PROC(p, td) {
996 		if (td->td_tid == tid)
997 			break;
998 	}
999 	return (td);
1000 }
1001 
1002 /* Locate a thread by number; return with proc lock held. */
1003 struct thread *
1004 tdfind(lwpid_t tid, pid_t pid)
1005 {
1006 #define RUN_THRESH	16
1007 	struct thread *td;
1008 	int run = 0;
1009 
1010 	rw_rlock(&tidhash_lock);
1011 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1012 		if (td->td_tid == tid) {
1013 			if (pid != -1 && td->td_proc->p_pid != pid) {
1014 				td = NULL;
1015 				break;
1016 			}
1017 			PROC_LOCK(td->td_proc);
1018 			if (td->td_proc->p_state == PRS_NEW) {
1019 				PROC_UNLOCK(td->td_proc);
1020 				td = NULL;
1021 				break;
1022 			}
1023 			if (run > RUN_THRESH) {
1024 				if (rw_try_upgrade(&tidhash_lock)) {
1025 					LIST_REMOVE(td, td_hash);
1026 					LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1027 						td, td_hash);
1028 					rw_wunlock(&tidhash_lock);
1029 					return (td);
1030 				}
1031 			}
1032 			break;
1033 		}
1034 		run++;
1035 	}
1036 	rw_runlock(&tidhash_lock);
1037 	return (td);
1038 }
1039 
1040 void
1041 tidhash_add(struct thread *td)
1042 {
1043 	rw_wlock(&tidhash_lock);
1044 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1045 	rw_wunlock(&tidhash_lock);
1046 }
1047 
1048 void
1049 tidhash_remove(struct thread *td)
1050 {
1051 	rw_wlock(&tidhash_lock);
1052 	LIST_REMOVE(td, td_hash);
1053 	rw_wunlock(&tidhash_lock);
1054 }
1055