xref: /freebsd/sys/kern/kern_thread.c (revision 0bb263df82e129f5f8c82da6deb55dfe10daa677)
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 <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/kernel.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/proc.h>
38 #include <sys/resourcevar.h>
39 #include <sys/smp.h>
40 #include <sys/sysctl.h>
41 #include <sys/sched.h>
42 #include <sys/sleepqueue.h>
43 #include <sys/turnstile.h>
44 #include <sys/ktr.h>
45 #include <sys/umtx.h>
46 
47 #include <security/audit/audit.h>
48 
49 #include <vm/vm.h>
50 #include <vm/vm_extern.h>
51 #include <vm/uma.h>
52 
53 /*
54  * thread related storage.
55  */
56 static uma_zone_t thread_zone;
57 
58 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
59 
60 int max_threads_per_proc = 1500;
61 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
62 	&max_threads_per_proc, 0, "Limit on threads per proc");
63 
64 int max_threads_hits;
65 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
66 	&max_threads_hits, 0, "");
67 
68 #ifdef KSE
69 int virtual_cpu;
70 
71 #endif
72 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
73 struct mtx zombie_lock;
74 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
75 
76 static void thread_zombie(struct thread *);
77 
78 #ifdef KSE
79 static int
80 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
81 {
82 	int error, new_val;
83 	int def_val;
84 
85 	def_val = mp_ncpus;
86 	if (virtual_cpu == 0)
87 		new_val = def_val;
88 	else
89 		new_val = virtual_cpu;
90 	error = sysctl_handle_int(oidp, &new_val, 0, req);
91 	if (error != 0 || req->newptr == NULL)
92 		return (error);
93 	if (new_val < 0)
94 		return (EINVAL);
95 	virtual_cpu = new_val;
96 	return (0);
97 }
98 
99 /* DEBUG ONLY */
100 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
101 	0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
102 	"debug virtual cpus");
103 #endif
104 
105 struct mtx tid_lock;
106 static struct unrhdr *tid_unrhdr;
107 
108 /*
109  * Prepare a thread for use.
110  */
111 static int
112 thread_ctor(void *mem, int size, void *arg, int flags)
113 {
114 	struct thread	*td;
115 
116 	td = (struct thread *)mem;
117 	td->td_state = TDS_INACTIVE;
118 	td->td_oncpu = NOCPU;
119 
120 	td->td_tid = alloc_unr(tid_unrhdr);
121 	td->td_syscalls = 0;
122 
123 	/*
124 	 * Note that td_critnest begins life as 1 because the thread is not
125 	 * running and is thereby implicitly waiting to be on the receiving
126 	 * end of a context switch.
127 	 */
128 	td->td_critnest = 1;
129 
130 #ifdef AUDIT
131 	audit_thread_alloc(td);
132 #endif
133 	umtx_thread_alloc(td);
134 	return (0);
135 }
136 
137 /*
138  * Reclaim a thread after use.
139  */
140 static void
141 thread_dtor(void *mem, int size, void *arg)
142 {
143 	struct thread *td;
144 
145 	td = (struct thread *)mem;
146 
147 #ifdef INVARIANTS
148 	/* Verify that this thread is in a safe state to free. */
149 	switch (td->td_state) {
150 	case TDS_INHIBITED:
151 	case TDS_RUNNING:
152 	case TDS_CAN_RUN:
153 	case TDS_RUNQ:
154 		/*
155 		 * We must never unlink a thread that is in one of
156 		 * these states, because it is currently active.
157 		 */
158 		panic("bad state for thread unlinking");
159 		/* NOTREACHED */
160 	case TDS_INACTIVE:
161 		break;
162 	default:
163 		panic("bad thread state");
164 		/* NOTREACHED */
165 	}
166 #endif
167 #ifdef AUDIT
168 	audit_thread_free(td);
169 #endif
170 	free_unr(tid_unrhdr, td->td_tid);
171 	sched_newthread(td);
172 }
173 
174 /*
175  * Initialize type-stable parts of a thread (when newly created).
176  */
177 static int
178 thread_init(void *mem, int size, int flags)
179 {
180 	struct thread *td;
181 
182 	td = (struct thread *)mem;
183 
184 	vm_thread_new(td, 0);
185 	cpu_thread_setup(td);
186 	td->td_sleepqueue = sleepq_alloc();
187 	td->td_turnstile = turnstile_alloc();
188 	td->td_sched = (struct td_sched *)&td[1];
189 	sched_newthread(td);
190 	umtx_thread_init(td);
191 	return (0);
192 }
193 
194 /*
195  * Tear down type-stable parts of a thread (just before being discarded).
196  */
197 static void
198 thread_fini(void *mem, int size)
199 {
200 	struct thread *td;
201 
202 	td = (struct thread *)mem;
203 	turnstile_free(td->td_turnstile);
204 	sleepq_free(td->td_sleepqueue);
205 	umtx_thread_fini(td);
206 	vm_thread_dispose(td);
207 }
208 
209 /*
210  * For a newly created process,
211  * link up all the structures and its initial threads etc.
212  * called from:
213  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
214  * proc_dtor() (should go away)
215  * proc_init()
216  */
217 void
218 proc_linkup(struct proc *p, struct thread *td)
219 {
220 
221 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
222 #ifdef KSE
223 	TAILQ_INIT(&p->p_upcalls);	     /* upcall list */
224 #endif
225 	sigqueue_init(&p->p_sigqueue, p);
226 	p->p_ksi = ksiginfo_alloc(1);
227 	if (p->p_ksi != NULL) {
228 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
229 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
230 	}
231 	LIST_INIT(&p->p_mqnotifier);
232 	p->p_numthreads = 0;
233 	thread_link(td, p);
234 }
235 
236 /*
237  * Initialize global thread allocation resources.
238  */
239 void
240 threadinit(void)
241 {
242 
243 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
244 	tid_unrhdr = new_unrhdr(PID_MAX + 1, INT_MAX, &tid_lock);
245 
246 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
247 	    thread_ctor, thread_dtor, thread_init, thread_fini,
248 	    16 - 1, 0);
249 #ifdef KSE
250 	kseinit();	/* set up kse specific stuff  e.g. upcall zone*/
251 #endif
252 }
253 
254 /*
255  * Place an unused thread on the zombie list.
256  * Use the slpq as that must be unused by now.
257  */
258 void
259 thread_zombie(struct thread *td)
260 {
261 	mtx_lock_spin(&zombie_lock);
262 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
263 	mtx_unlock_spin(&zombie_lock);
264 }
265 
266 /*
267  * Release a thread that has exited after cpu_throw().
268  */
269 void
270 thread_stash(struct thread *td)
271 {
272 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
273 	thread_zombie(td);
274 }
275 
276 /*
277  * Reap zombie kse resource.
278  */
279 void
280 thread_reap(void)
281 {
282 	struct thread *td_first, *td_next;
283 
284 	/*
285 	 * Don't even bother to lock if none at this instant,
286 	 * we really don't care about the next instant..
287 	 */
288 	if (!TAILQ_EMPTY(&zombie_threads)) {
289 		mtx_lock_spin(&zombie_lock);
290 		td_first = TAILQ_FIRST(&zombie_threads);
291 		if (td_first)
292 			TAILQ_INIT(&zombie_threads);
293 		mtx_unlock_spin(&zombie_lock);
294 		while (td_first) {
295 			td_next = TAILQ_NEXT(td_first, td_slpq);
296 			if (td_first->td_ucred)
297 				crfree(td_first->td_ucred);
298 			thread_free(td_first);
299 			td_first = td_next;
300 		}
301 	}
302 }
303 
304 /*
305  * Allocate a thread.
306  */
307 struct thread *
308 thread_alloc(void)
309 {
310 
311 	thread_reap(); /* check if any zombies to get */
312 	return (uma_zalloc(thread_zone, M_WAITOK));
313 }
314 
315 
316 /*
317  * Deallocate a thread.
318  */
319 void
320 thread_free(struct thread *td)
321 {
322 
323 	cpu_thread_clean(td);
324 	uma_zfree(thread_zone, td);
325 }
326 
327 /*
328  * Discard the current thread and exit from its context.
329  * Always called with scheduler locked.
330  *
331  * Because we can't free a thread while we're operating under its context,
332  * push the current thread into our CPU's deadthread holder. This means
333  * we needn't worry about someone else grabbing our context before we
334  * do a cpu_throw().  This may not be needed now as we are under schedlock.
335  * Maybe we can just do a thread_stash() as thr_exit1 does.
336  */
337 /*  XXX
338  * libthr expects its thread exit to return for the last
339  * thread, meaning that the program is back to non-threaded
340  * mode I guess. Because we do this (cpu_throw) unconditionally
341  * here, they have their own version of it. (thr_exit1())
342  * that doesn't do it all if this was the last thread.
343  * It is also called from thread_suspend_check().
344  * Of course in the end, they end up coming here through exit1
345  * anyhow..  After fixing 'thr' to play by the rules we should be able
346  * to merge these two functions together.
347  *
348  * called from:
349  * exit1()
350  * kse_exit()
351  * thr_exit()
352  * ifdef KSE
353  * thread_user_enter()
354  * thread_userret()
355  * endif
356  * thread_suspend_check()
357  */
358 void
359 thread_exit(void)
360 {
361 	uint64_t new_switchtime;
362 	struct thread *td;
363 	struct thread *td2;
364 	struct proc *p;
365 
366 	td = curthread;
367 	p = td->td_proc;
368 
369 	PROC_SLOCK_ASSERT(p, MA_OWNED);
370 	mtx_assert(&Giant, MA_NOTOWNED);
371 
372 	PROC_LOCK_ASSERT(p, MA_OWNED);
373 	KASSERT(p != NULL, ("thread exiting without a process"));
374 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
375 	    (long)p->p_pid, p->p_comm);
376 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
377 
378 #ifdef AUDIT
379 	AUDIT_SYSCALL_EXIT(0, td);
380 #endif
381 
382 #ifdef KSE
383 	if (td->td_standin != NULL) {
384 		/*
385 		 * Note that we don't need to free the cred here as it
386 		 * is done in thread_reap().
387 		 */
388 		thread_zombie(td->td_standin);
389 		td->td_standin = NULL;
390 	}
391 #endif
392 
393 	umtx_thread_exit(td);
394 
395 	/*
396 	 * drop FPU & debug register state storage, or any other
397 	 * architecture specific resources that
398 	 * would not be on a new untouched process.
399 	 */
400 	cpu_thread_exit(td);	/* XXXSMP */
401 
402 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
403 	new_switchtime = cpu_ticks();
404 	p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
405 	PCPU_SET(switchtime, new_switchtime);
406 	PCPU_SET(switchticks, ticks);
407 	PCPU_INC(cnt.v_swtch);
408 	/* Save our resource usage in our process. */
409 	td->td_ru.ru_nvcsw++;
410 	rucollect(&p->p_ru, &td->td_ru);
411 	/*
412 	 * The last thread is left attached to the process
413 	 * So that the whole bundle gets recycled. Skip
414 	 * all this stuff if we never had threads.
415 	 * EXIT clears all sign of other threads when
416 	 * it goes to single threading, so the last thread always
417 	 * takes the short path.
418 	 */
419 	if (p->p_flag & P_HADTHREADS) {
420 		if (p->p_numthreads > 1) {
421 			thread_lock(td);
422 #ifdef KSE
423 			kse_unlink(td);
424 #else
425 			thread_unlink(td);
426 #endif
427 			thread_unlock(td);
428 			td2 = FIRST_THREAD_IN_PROC(p);
429 			sched_exit_thread(td2, td);
430 
431 			/*
432 			 * The test below is NOT true if we are the
433 			 * sole exiting thread. P_STOPPED_SNGL is unset
434 			 * in exit1() after it is the only survivor.
435 			 */
436 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
437 				if (p->p_numthreads == p->p_suspcount) {
438 					thread_lock(p->p_singlethread);
439 					thread_unsuspend_one(p->p_singlethread);
440 					thread_unlock(p->p_singlethread);
441 				}
442 			}
443 
444 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
445 			PCPU_SET(deadthread, td);
446 		} else {
447 			/*
448 			 * The last thread is exiting.. but not through exit()
449 			 * what should we do?
450 			 * Theoretically this can't happen
451  			 * exit1() - clears threading flags before coming here
452  			 * kse_exit() - treats last thread specially
453  			 * thr_exit() - treats last thread specially
454 			 * ifdef KSE
455  			 * thread_user_enter() - only if more exist
456  			 * thread_userret() - only if more exist
457 			 * endif
458  			 * thread_suspend_check() - only if more exist
459 			 */
460 			panic ("thread_exit: Last thread exiting on its own");
461 		}
462 	}
463 	PROC_UNLOCK(p);
464 	thread_lock(td);
465 	/* Save our tick information with both the thread and proc locked */
466 	ruxagg(&p->p_rux, td);
467 	PROC_SUNLOCK(p);
468 	td->td_state = TDS_INACTIVE;
469 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
470 	sched_throw(td);
471 	panic("I'm a teapot!");
472 	/* NOTREACHED */
473 }
474 
475 /*
476  * Do any thread specific cleanups that may be needed in wait()
477  * called with Giant, proc and schedlock not held.
478  */
479 void
480 thread_wait(struct proc *p)
481 {
482 	struct thread *td;
483 
484 	mtx_assert(&Giant, MA_NOTOWNED);
485 	KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
486 	td = FIRST_THREAD_IN_PROC(p);
487 #ifdef KSE
488 	if (td->td_standin != NULL) {
489 		if (td->td_standin->td_ucred != NULL) {
490 			crfree(td->td_standin->td_ucred);
491 			td->td_standin->td_ucred = NULL;
492 		}
493 		thread_free(td->td_standin);
494 		td->td_standin = NULL;
495 	}
496 #endif
497 	/* Lock the last thread so we spin until it exits cpu_throw(). */
498 	thread_lock(td);
499 	thread_unlock(td);
500 	/* Wait for any remaining threads to exit cpu_throw(). */
501 	while (p->p_exitthreads)
502 		sched_relinquish(curthread);
503 	cpu_thread_clean(td);
504 	crfree(td->td_ucred);
505 	thread_reap();	/* check for zombie threads etc. */
506 }
507 
508 /*
509  * Link a thread to a process.
510  * set up anything that needs to be initialized for it to
511  * be used by the process.
512  *
513  * Note that we do not link to the proc's ucred here.
514  * The thread is linked as if running but no KSE assigned.
515  * Called from:
516  *  proc_linkup()
517  *  thread_schedule_upcall()
518  *  thr_create()
519  */
520 void
521 thread_link(struct thread *td, struct proc *p)
522 {
523 
524 	/*
525 	 * XXX This can't be enabled because it's called for proc0 before
526 	 * it's spinlock has been created.
527 	 * PROC_SLOCK_ASSERT(p, MA_OWNED);
528 	 */
529 	td->td_state    = TDS_INACTIVE;
530 	td->td_proc     = p;
531 	td->td_flags    = 0;
532 
533 	LIST_INIT(&td->td_contested);
534 	sigqueue_init(&td->td_sigqueue, p);
535 	callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
536 	TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
537 	p->p_numthreads++;
538 }
539 
540 /*
541  * Convert a process with one thread to an unthreaded process.
542  * Called from:
543  *  thread_single(exit)  (called from execve and exit)
544  *  kse_exit()		XXX may need cleaning up wrt KSE stuff
545  */
546 void
547 thread_unthread(struct thread *td)
548 {
549 	struct proc *p = td->td_proc;
550 
551 	KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
552 #ifdef KSE
553 	upcall_remove(td);
554 	p->p_flag &= ~(P_SA|P_HADTHREADS);
555 	td->td_mailbox = NULL;
556 	td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
557 	if (td->td_standin != NULL) {
558 		thread_zombie(td->td_standin);
559 		td->td_standin = NULL;
560 	}
561 #else
562 	p->p_flag &= ~P_HADTHREADS;
563 #endif
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_SLOCK_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 /*
583  * Enforce single-threading.
584  *
585  * Returns 1 if the caller must abort (another thread is waiting to
586  * exit the process or similar). Process is locked!
587  * Returns 0 when you are successfully the only thread running.
588  * A process has successfully single threaded in the suspend mode when
589  * There are no threads in user mode. Threads in the kernel must be
590  * allowed to continue until they get to the user boundary. They may even
591  * copy out their return values and data before suspending. They may however be
592  * accelerated in reaching the user boundary as we will wake up
593  * any sleeping threads that are interruptable. (PCATCH).
594  */
595 int
596 thread_single(int mode)
597 {
598 	struct thread *td;
599 	struct thread *td2;
600 	struct proc *p;
601 	int remaining;
602 
603 	td = curthread;
604 	p = td->td_proc;
605 	mtx_assert(&Giant, MA_NOTOWNED);
606 	PROC_LOCK_ASSERT(p, MA_OWNED);
607 	KASSERT((td != NULL), ("curthread is NULL"));
608 
609 	if ((p->p_flag & P_HADTHREADS) == 0)
610 		return (0);
611 
612 	/* Is someone already single threading? */
613 	if (p->p_singlethread != NULL && p->p_singlethread != td)
614 		return (1);
615 
616 	if (mode == SINGLE_EXIT) {
617 		p->p_flag |= P_SINGLE_EXIT;
618 		p->p_flag &= ~P_SINGLE_BOUNDARY;
619 	} else {
620 		p->p_flag &= ~P_SINGLE_EXIT;
621 		if (mode == SINGLE_BOUNDARY)
622 			p->p_flag |= P_SINGLE_BOUNDARY;
623 		else
624 			p->p_flag &= ~P_SINGLE_BOUNDARY;
625 	}
626 	p->p_flag |= P_STOPPED_SINGLE;
627 	PROC_SLOCK(p);
628 	p->p_singlethread = td;
629 	if (mode == SINGLE_EXIT)
630 		remaining = p->p_numthreads;
631 	else if (mode == SINGLE_BOUNDARY)
632 		remaining = p->p_numthreads - p->p_boundary_count;
633 	else
634 		remaining = p->p_numthreads - p->p_suspcount;
635 	while (remaining != 1) {
636 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
637 			goto stopme;
638 		FOREACH_THREAD_IN_PROC(p, td2) {
639 			if (td2 == td)
640 				continue;
641 			thread_lock(td2);
642 			td2->td_flags |= TDF_ASTPENDING;
643 			if (TD_IS_INHIBITED(td2)) {
644 				switch (mode) {
645 				case SINGLE_EXIT:
646 					if (td->td_flags & TDF_DBSUSPEND)
647 						td->td_flags &= ~TDF_DBSUSPEND;
648 					if (TD_IS_SUSPENDED(td2))
649 						thread_unsuspend_one(td2);
650 					if (TD_ON_SLEEPQ(td2) &&
651 					    (td2->td_flags & TDF_SINTR))
652 						sleepq_abort(td2, EINTR);
653 					break;
654 				case SINGLE_BOUNDARY:
655 					if (TD_IS_SUSPENDED(td2) &&
656 					    !(td2->td_flags & TDF_BOUNDARY))
657 						thread_unsuspend_one(td2);
658 					if (TD_ON_SLEEPQ(td2) &&
659 					    (td2->td_flags & TDF_SINTR))
660 						sleepq_abort(td2, ERESTART);
661 					break;
662 				default:
663 					if (TD_IS_SUSPENDED(td2)) {
664 						thread_unlock(td2);
665 						continue;
666 					}
667 					/*
668 					 * maybe other inhibited states too?
669 					 */
670 					if ((td2->td_flags & TDF_SINTR) &&
671 					    (td2->td_inhibitors &
672 					    (TDI_SLEEPING | TDI_SWAPPED)))
673 						thread_suspend_one(td2);
674 					break;
675 				}
676 			}
677 #ifdef SMP
678 			else if (TD_IS_RUNNING(td2) && td != td2) {
679 				forward_signal(td2);
680 			}
681 #endif
682 			thread_unlock(td2);
683 		}
684 		if (mode == SINGLE_EXIT)
685 			remaining = p->p_numthreads;
686 		else if (mode == SINGLE_BOUNDARY)
687 			remaining = p->p_numthreads - p->p_boundary_count;
688 		else
689 			remaining = p->p_numthreads - p->p_suspcount;
690 
691 		/*
692 		 * Maybe we suspended some threads.. was it enough?
693 		 */
694 		if (remaining == 1)
695 			break;
696 
697 stopme:
698 		/*
699 		 * Wake us up when everyone else has suspended.
700 		 * In the mean time we suspend as well.
701 		 */
702 		thread_suspend_switch(td);
703 		if (mode == SINGLE_EXIT)
704 			remaining = p->p_numthreads;
705 		else if (mode == SINGLE_BOUNDARY)
706 			remaining = p->p_numthreads - p->p_boundary_count;
707 		else
708 			remaining = p->p_numthreads - p->p_suspcount;
709 	}
710 	if (mode == SINGLE_EXIT) {
711 		/*
712 		 * We have gotten rid of all the other threads and we
713 		 * are about to either exit or exec. In either case,
714 		 * we try our utmost  to revert to being a non-threaded
715 		 * process.
716 		 */
717 		p->p_singlethread = NULL;
718 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
719 		thread_unthread(td);
720 	}
721 	PROC_SUNLOCK(p);
722 	return (0);
723 }
724 
725 /*
726  * Called in from locations that can safely check to see
727  * whether we have to suspend or at least throttle for a
728  * single-thread event (e.g. fork).
729  *
730  * Such locations include userret().
731  * If the "return_instead" argument is non zero, the thread must be able to
732  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
733  *
734  * The 'return_instead' argument tells the function if it may do a
735  * thread_exit() or suspend, or whether the caller must abort and back
736  * out instead.
737  *
738  * If the thread that set the single_threading request has set the
739  * P_SINGLE_EXIT bit in the process flags then this call will never return
740  * if 'return_instead' is false, but will exit.
741  *
742  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
743  *---------------+--------------------+---------------------
744  *       0       | returns 0          |   returns 0 or 1
745  *               | when ST ends       |   immediatly
746  *---------------+--------------------+---------------------
747  *       1       | thread exits       |   returns 1
748  *               |                    |  immediatly
749  * 0 = thread_exit() or suspension ok,
750  * other = return error instead of stopping the thread.
751  *
752  * While a full suspension is under effect, even a single threading
753  * thread would be suspended if it made this call (but it shouldn't).
754  * This call should only be made from places where
755  * thread_exit() would be safe as that may be the outcome unless
756  * return_instead is set.
757  */
758 int
759 thread_suspend_check(int return_instead)
760 {
761 	struct thread *td;
762 	struct proc *p;
763 
764 	td = curthread;
765 	p = td->td_proc;
766 	mtx_assert(&Giant, MA_NOTOWNED);
767 	PROC_LOCK_ASSERT(p, MA_OWNED);
768 	while (P_SHOULDSTOP(p) ||
769 	      ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
770 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
771 			KASSERT(p->p_singlethread != NULL,
772 			    ("singlethread not set"));
773 			/*
774 			 * The only suspension in action is a
775 			 * single-threading. Single threader need not stop.
776 			 * XXX Should be safe to access unlocked
777 			 * as it can only be set to be true by us.
778 			 */
779 			if (p->p_singlethread == td)
780 				return (0);	/* Exempt from stopping. */
781 		}
782 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
783 			return (EINTR);
784 
785 		/* Should we goto user boundary if we didn't come from there? */
786 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
787 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
788 			return (ERESTART);
789 
790 		/* If thread will exit, flush its pending signals */
791 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
792 			sigqueue_flush(&td->td_sigqueue);
793 
794 		PROC_SLOCK(p);
795 		thread_stopped(p);
796 		/*
797 		 * If the process is waiting for us to exit,
798 		 * this thread should just suicide.
799 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
800 		 */
801 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
802 			thread_exit();
803 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
804 			if (p->p_numthreads == p->p_suspcount + 1) {
805 				thread_lock(p->p_singlethread);
806 				thread_unsuspend_one(p->p_singlethread);
807 				thread_unlock(p->p_singlethread);
808 			}
809 		}
810 		PROC_UNLOCK(p);
811 		thread_lock(td);
812 		/*
813 		 * When a thread suspends, it just
814 		 * gets taken off all queues.
815 		 */
816 		thread_suspend_one(td);
817 		if (return_instead == 0) {
818 			p->p_boundary_count++;
819 			td->td_flags |= TDF_BOUNDARY;
820 		}
821 		PROC_SUNLOCK(p);
822 		mi_switch(SW_INVOL, NULL);
823 		if (return_instead == 0)
824 			td->td_flags &= ~TDF_BOUNDARY;
825 		thread_unlock(td);
826 		PROC_LOCK(p);
827 		if (return_instead == 0)
828 			p->p_boundary_count--;
829 	}
830 	return (0);
831 }
832 
833 void
834 thread_suspend_switch(struct thread *td)
835 {
836 	struct proc *p;
837 
838 	p = td->td_proc;
839 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
840 	PROC_LOCK_ASSERT(p, MA_OWNED);
841 	PROC_SLOCK_ASSERT(p, MA_OWNED);
842 	/*
843 	 * We implement thread_suspend_one in stages here to avoid
844 	 * dropping the proc lock while the thread lock is owned.
845 	 */
846 	thread_stopped(p);
847 	p->p_suspcount++;
848 	PROC_UNLOCK(p);
849 	thread_lock(td);
850 	TD_SET_SUSPENDED(td);
851 	PROC_SUNLOCK(p);
852 	DROP_GIANT();
853 	mi_switch(SW_VOL, NULL);
854 	thread_unlock(td);
855 	PICKUP_GIANT();
856 	PROC_LOCK(p);
857 	PROC_SLOCK(p);
858 }
859 
860 void
861 thread_suspend_one(struct thread *td)
862 {
863 	struct proc *p = td->td_proc;
864 
865 	PROC_SLOCK_ASSERT(p, MA_OWNED);
866 	THREAD_LOCK_ASSERT(td, MA_OWNED);
867 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
868 	p->p_suspcount++;
869 	TD_SET_SUSPENDED(td);
870 }
871 
872 void
873 thread_unsuspend_one(struct thread *td)
874 {
875 	struct proc *p = td->td_proc;
876 
877 	PROC_SLOCK_ASSERT(p, MA_OWNED);
878 	THREAD_LOCK_ASSERT(td, MA_OWNED);
879 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
880 	TD_CLR_SUSPENDED(td);
881 	p->p_suspcount--;
882 	setrunnable(td);
883 }
884 
885 /*
886  * Allow all threads blocked by single threading to continue running.
887  */
888 void
889 thread_unsuspend(struct proc *p)
890 {
891 	struct thread *td;
892 
893 	PROC_LOCK_ASSERT(p, MA_OWNED);
894 	PROC_SLOCK_ASSERT(p, MA_OWNED);
895 	if (!P_SHOULDSTOP(p)) {
896                 FOREACH_THREAD_IN_PROC(p, td) {
897 			thread_lock(td);
898 			if (TD_IS_SUSPENDED(td)) {
899 				thread_unsuspend_one(td);
900 			}
901 			thread_unlock(td);
902 		}
903 	} else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
904 	    (p->p_numthreads == p->p_suspcount)) {
905 		/*
906 		 * Stopping everything also did the job for the single
907 		 * threading request. Now we've downgraded to single-threaded,
908 		 * let it continue.
909 		 */
910 		thread_lock(p->p_singlethread);
911 		thread_unsuspend_one(p->p_singlethread);
912 		thread_unlock(p->p_singlethread);
913 	}
914 }
915 
916 /*
917  * End the single threading mode..
918  */
919 void
920 thread_single_end(void)
921 {
922 	struct thread *td;
923 	struct proc *p;
924 
925 	td = curthread;
926 	p = td->td_proc;
927 	PROC_LOCK_ASSERT(p, MA_OWNED);
928 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
929 	PROC_SLOCK(p);
930 	p->p_singlethread = NULL;
931 	/*
932 	 * If there are other threads they mey now run,
933 	 * unless of course there is a blanket 'stop order'
934 	 * on the process. The single threader must be allowed
935 	 * to continue however as this is a bad place to stop.
936 	 */
937 	if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
938                 FOREACH_THREAD_IN_PROC(p, td) {
939 			thread_lock(td);
940 			if (TD_IS_SUSPENDED(td)) {
941 				thread_unsuspend_one(td);
942 			}
943 			thread_unlock(td);
944 		}
945 	}
946 	PROC_SUNLOCK(p);
947 }
948 
949 struct thread *
950 thread_find(struct proc *p, lwpid_t tid)
951 {
952 	struct thread *td;
953 
954 	PROC_LOCK_ASSERT(p, MA_OWNED);
955 	PROC_SLOCK(p);
956 	FOREACH_THREAD_IN_PROC(p, td) {
957 		if (td->td_tid == tid)
958 			break;
959 	}
960 	PROC_SUNLOCK(p);
961 	return (td);
962 }
963