xref: /freebsd/sys/kern/kern_thread.c (revision f9218d3d4fd34f082473b3a021c6d4d109fb47cf)
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  * $FreeBSD$
29  */
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/kernel.h>
34 #include <sys/lock.h>
35 #include <sys/malloc.h>
36 #include <sys/mutex.h>
37 #include <sys/proc.h>
38 #include <sys/smp.h>
39 #include <sys/sysctl.h>
40 #include <sys/sysproto.h>
41 #include <sys/filedesc.h>
42 #include <sys/sched.h>
43 #include <sys/signalvar.h>
44 #include <sys/sx.h>
45 #include <sys/tty.h>
46 #include <sys/user.h>
47 #include <sys/jail.h>
48 #include <sys/kse.h>
49 #include <sys/ktr.h>
50 #include <sys/ucontext.h>
51 
52 #include <vm/vm.h>
53 #include <vm/vm_object.h>
54 #include <vm/pmap.h>
55 #include <vm/uma.h>
56 #include <vm/vm_map.h>
57 
58 #include <machine/frame.h>
59 
60 /*
61  * KSEGRP related storage.
62  */
63 static uma_zone_t ksegrp_zone;
64 static uma_zone_t kse_zone;
65 static uma_zone_t thread_zone;
66 static uma_zone_t upcall_zone;
67 
68 /* DEBUG ONLY */
69 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
70 static int thread_debug = 0;
71 SYSCTL_INT(_kern_threads, OID_AUTO, debug, CTLFLAG_RW,
72 	&thread_debug, 0, "thread debug");
73 
74 static int max_threads_per_proc = 30;
75 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
76 	&max_threads_per_proc, 0, "Limit on threads per proc");
77 
78 static int max_groups_per_proc = 5;
79 SYSCTL_INT(_kern_threads, OID_AUTO, max_groups_per_proc, CTLFLAG_RW,
80 	&max_groups_per_proc, 0, "Limit on thread groups per proc");
81 
82 static int max_threads_hits;
83 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
84 	&max_threads_hits, 0, "");
85 
86 static int virtual_cpu;
87 
88 #define RANGEOF(type, start, end) (offsetof(type, end) - offsetof(type, start))
89 
90 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
91 TAILQ_HEAD(, kse) zombie_kses = TAILQ_HEAD_INITIALIZER(zombie_kses);
92 TAILQ_HEAD(, ksegrp) zombie_ksegrps = TAILQ_HEAD_INITIALIZER(zombie_ksegrps);
93 TAILQ_HEAD(, kse_upcall) zombie_upcalls =
94 	TAILQ_HEAD_INITIALIZER(zombie_upcalls);
95 struct mtx kse_zombie_lock;
96 MTX_SYSINIT(kse_zombie_lock, &kse_zombie_lock, "kse zombie lock", MTX_SPIN);
97 
98 static void kse_purge(struct proc *p, struct thread *td);
99 static void kse_purge_group(struct thread *td);
100 static int thread_update_usr_ticks(struct thread *td, int user);
101 static void thread_alloc_spare(struct thread *td, struct thread *spare);
102 
103 static int
104 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
105 {
106 	int error, new_val;
107 	int def_val;
108 
109 #ifdef SMP
110 	def_val = mp_ncpus;
111 #else
112 	def_val = 1;
113 #endif
114 	if (virtual_cpu == 0)
115 		new_val = def_val;
116 	else
117 		new_val = virtual_cpu;
118 	error = sysctl_handle_int(oidp, &new_val, 0, req);
119         if (error != 0 || req->newptr == NULL)
120 		return (error);
121 	if (new_val < 0)
122 		return (EINVAL);
123 	virtual_cpu = new_val;
124 	return (0);
125 }
126 
127 /* DEBUG ONLY */
128 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
129 	0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
130 	"debug virtual cpus");
131 
132 /*
133  * Prepare a thread for use.
134  */
135 static void
136 thread_ctor(void *mem, int size, void *arg)
137 {
138 	struct thread	*td;
139 
140 	td = (struct thread *)mem;
141 	td->td_state = TDS_INACTIVE;
142 }
143 
144 /*
145  * Reclaim a thread after use.
146  */
147 static void
148 thread_dtor(void *mem, int size, void *arg)
149 {
150 	struct thread	*td;
151 
152 	td = (struct thread *)mem;
153 
154 #ifdef INVARIANTS
155 	/* Verify that this thread is in a safe state to free. */
156 	switch (td->td_state) {
157 	case TDS_INHIBITED:
158 	case TDS_RUNNING:
159 	case TDS_CAN_RUN:
160 	case TDS_RUNQ:
161 		/*
162 		 * We must never unlink a thread that is in one of
163 		 * these states, because it is currently active.
164 		 */
165 		panic("bad state for thread unlinking");
166 		/* NOTREACHED */
167 	case TDS_INACTIVE:
168 		break;
169 	default:
170 		panic("bad thread state");
171 		/* NOTREACHED */
172 	}
173 #endif
174 }
175 
176 /*
177  * Initialize type-stable parts of a thread (when newly created).
178  */
179 static void
180 thread_init(void *mem, int size)
181 {
182 	struct thread	*td;
183 
184 	td = (struct thread *)mem;
185 	mtx_lock(&Giant);
186 	pmap_new_thread(td, 0);
187 	mtx_unlock(&Giant);
188 	cpu_thread_setup(td);
189 	td->td_sched = (struct td_sched *)&td[1];
190 }
191 
192 /*
193  * Tear down type-stable parts of a thread (just before being discarded).
194  */
195 static void
196 thread_fini(void *mem, int size)
197 {
198 	struct thread	*td;
199 
200 	td = (struct thread *)mem;
201 	pmap_dispose_thread(td);
202 }
203 
204 /*
205  * Initialize type-stable parts of a kse (when newly created).
206  */
207 static void
208 kse_init(void *mem, int size)
209 {
210 	struct kse	*ke;
211 
212 	ke = (struct kse *)mem;
213 	ke->ke_sched = (struct ke_sched *)&ke[1];
214 }
215 
216 /*
217  * Initialize type-stable parts of a ksegrp (when newly created).
218  */
219 static void
220 ksegrp_init(void *mem, int size)
221 {
222 	struct ksegrp	*kg;
223 
224 	kg = (struct ksegrp *)mem;
225 	kg->kg_sched = (struct kg_sched *)&kg[1];
226 }
227 
228 /*
229  * KSE is linked into kse group.
230  */
231 void
232 kse_link(struct kse *ke, struct ksegrp *kg)
233 {
234 	struct proc *p = kg->kg_proc;
235 
236 	TAILQ_INSERT_HEAD(&kg->kg_kseq, ke, ke_kglist);
237 	kg->kg_kses++;
238 	ke->ke_state	= KES_UNQUEUED;
239 	ke->ke_proc	= p;
240 	ke->ke_ksegrp	= kg;
241 	ke->ke_thread	= NULL;
242 	ke->ke_oncpu	= NOCPU;
243 	ke->ke_flags	= 0;
244 }
245 
246 void
247 kse_unlink(struct kse *ke)
248 {
249 	struct ksegrp *kg;
250 
251 	mtx_assert(&sched_lock, MA_OWNED);
252 	kg = ke->ke_ksegrp;
253 	TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist);
254 	if (ke->ke_state == KES_IDLE) {
255 		TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist);
256 		kg->kg_idle_kses--;
257 	}
258 	if (--kg->kg_kses == 0)
259 		ksegrp_unlink(kg);
260 	/*
261 	 * Aggregate stats from the KSE
262 	 */
263 	kse_stash(ke);
264 }
265 
266 void
267 ksegrp_link(struct ksegrp *kg, struct proc *p)
268 {
269 
270 	TAILQ_INIT(&kg->kg_threads);
271 	TAILQ_INIT(&kg->kg_runq);	/* links with td_runq */
272 	TAILQ_INIT(&kg->kg_slpq);	/* links with td_runq */
273 	TAILQ_INIT(&kg->kg_kseq);	/* all kses in ksegrp */
274 	TAILQ_INIT(&kg->kg_iq);		/* all idle kses in ksegrp */
275 	TAILQ_INIT(&kg->kg_upcalls);	/* all upcall structure in ksegrp */
276 	kg->kg_proc = p;
277 	/*
278 	 * the following counters are in the -zero- section
279 	 * and may not need clearing
280 	 */
281 	kg->kg_numthreads = 0;
282 	kg->kg_runnable   = 0;
283 	kg->kg_kses       = 0;
284 	kg->kg_runq_kses  = 0; /* XXXKSE change name */
285 	kg->kg_idle_kses  = 0;
286 	kg->kg_numupcalls = 0;
287 	/* link it in now that it's consistent */
288 	p->p_numksegrps++;
289 	TAILQ_INSERT_HEAD(&p->p_ksegrps, kg, kg_ksegrp);
290 }
291 
292 void
293 ksegrp_unlink(struct ksegrp *kg)
294 {
295 	struct proc *p;
296 
297 	mtx_assert(&sched_lock, MA_OWNED);
298 	KASSERT((kg->kg_numthreads == 0), ("ksegrp_unlink: residual threads"));
299 	KASSERT((kg->kg_kses == 0), ("ksegrp_unlink: residual kses"));
300 	KASSERT((kg->kg_numupcalls == 0), ("ksegrp_unlink: residual upcalls"));
301 
302 	p = kg->kg_proc;
303 	TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
304 	p->p_numksegrps--;
305 	/*
306 	 * Aggregate stats from the KSE
307 	 */
308 	ksegrp_stash(kg);
309 }
310 
311 struct kse_upcall *
312 upcall_alloc(void)
313 {
314 	struct kse_upcall *ku;
315 
316 	ku = uma_zalloc(upcall_zone, M_WAITOK);
317 	bzero(ku, sizeof(*ku));
318 	return (ku);
319 }
320 
321 void
322 upcall_free(struct kse_upcall *ku)
323 {
324 
325 	uma_zfree(upcall_zone, ku);
326 }
327 
328 void
329 upcall_link(struct kse_upcall *ku, struct ksegrp *kg)
330 {
331 
332 	mtx_assert(&sched_lock, MA_OWNED);
333 	TAILQ_INSERT_TAIL(&kg->kg_upcalls, ku, ku_link);
334 	ku->ku_ksegrp = kg;
335 	kg->kg_numupcalls++;
336 }
337 
338 void
339 upcall_unlink(struct kse_upcall *ku)
340 {
341 	struct ksegrp *kg = ku->ku_ksegrp;
342 
343 	mtx_assert(&sched_lock, MA_OWNED);
344 	KASSERT(ku->ku_owner == NULL, ("%s: have owner", __func__));
345 	TAILQ_REMOVE(&kg->kg_upcalls, ku, ku_link);
346 	kg->kg_numupcalls--;
347 	upcall_stash(ku);
348 }
349 
350 void
351 upcall_remove(struct thread *td)
352 {
353 
354 	if (td->td_upcall) {
355 		td->td_upcall->ku_owner = NULL;
356 		upcall_unlink(td->td_upcall);
357 		td->td_upcall = 0;
358 	}
359 }
360 
361 /*
362  * For a newly created process,
363  * link up all the structures and its initial threads etc.
364  */
365 void
366 proc_linkup(struct proc *p, struct ksegrp *kg,
367 	    struct kse *ke, struct thread *td)
368 {
369 
370 	TAILQ_INIT(&p->p_ksegrps);	     /* all ksegrps in proc */
371 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
372 	TAILQ_INIT(&p->p_suspended);	     /* Threads suspended */
373 	p->p_numksegrps = 0;
374 	p->p_numthreads = 0;
375 
376 	ksegrp_link(kg, p);
377 	kse_link(ke, kg);
378 	thread_link(td, kg);
379 }
380 
381 /*
382 struct kse_thr_interrupt_args {
383 	struct kse_thr_mailbox * tmbx;
384 };
385 */
386 int
387 kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap)
388 {
389 	struct proc *p;
390 	struct thread *td2;
391 
392 	p = td->td_proc;
393 	if (!(p->p_flag & P_THREADED) || (uap->tmbx == NULL))
394 		return (EINVAL);
395 	mtx_lock_spin(&sched_lock);
396 	FOREACH_THREAD_IN_PROC(p, td2) {
397 		if (td2->td_mailbox == uap->tmbx) {
398 			td2->td_flags |= TDF_INTERRUPT;
399 			if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR)) {
400 				if (td2->td_flags & TDF_CVWAITQ)
401 					cv_abort(td2);
402 				else
403 					abortsleep(td2);
404 			}
405 			mtx_unlock_spin(&sched_lock);
406 			return (0);
407 		}
408 	}
409 	mtx_unlock_spin(&sched_lock);
410 	return (ESRCH);
411 }
412 
413 /*
414 struct kse_exit_args {
415 	register_t dummy;
416 };
417 */
418 int
419 kse_exit(struct thread *td, struct kse_exit_args *uap)
420 {
421 	struct proc *p;
422 	struct ksegrp *kg;
423 	struct kse *ke;
424 
425 	p = td->td_proc;
426 	/*
427 	 * Only UTS can call the syscall and current group
428 	 * should be a threaded group.
429 	 */
430 	if ((td->td_mailbox != NULL) || (td->td_ksegrp->kg_numupcalls == 0))
431 		return (EINVAL);
432 	KASSERT((td->td_upcall != NULL), ("%s: not own an upcall", __func__));
433 
434 	kg = td->td_ksegrp;
435 	/* Serialize removing upcall */
436 	PROC_LOCK(p);
437 	mtx_lock_spin(&sched_lock);
438 	if ((kg->kg_numupcalls == 1) && (kg->kg_numthreads > 1)) {
439 		mtx_unlock_spin(&sched_lock);
440 		PROC_UNLOCK(p);
441 		return (EDEADLK);
442 	}
443 	ke = td->td_kse;
444 	upcall_remove(td);
445 	if (p->p_numthreads == 1) {
446 		kse_purge(p, td);
447 		p->p_flag &= ~P_THREADED;
448 		mtx_unlock_spin(&sched_lock);
449 		PROC_UNLOCK(p);
450 	} else {
451 		if (kg->kg_numthreads == 1) { /* Shutdown a group */
452 			kse_purge_group(td);
453 			ke->ke_flags |= KEF_EXIT;
454 		}
455 		thread_exit();
456 		/* NOTREACHED */
457 	}
458 	return (0);
459 }
460 
461 /*
462  * Either becomes an upcall or waits for an awakening event and
463  * then becomes an upcall. Only error cases return.
464  */
465 /*
466 struct kse_release_args {
467 	struct timespec *timeout;
468 };
469 */
470 int
471 kse_release(struct thread *td, struct kse_release_args *uap)
472 {
473 	struct proc *p;
474 	struct ksegrp *kg;
475 	struct timespec ts, ts2, ts3, timeout;
476 	struct timeval tv;
477 	int error;
478 
479 	p = td->td_proc;
480 	kg = td->td_ksegrp;
481 	/*
482 	 * Only UTS can call the syscall and current group
483 	 * should be a threaded group.
484 	 */
485 	if ((td->td_mailbox != NULL) || (td->td_ksegrp->kg_numupcalls == 0))
486 		return (EINVAL);
487 	KASSERT((td->td_upcall != NULL), ("%s: not own an upcall", __func__));
488 	if (uap->timeout != NULL) {
489 		if ((error = copyin(uap->timeout, &timeout, sizeof(timeout))))
490 			return (error);
491 		getnanouptime(&ts);
492 		timespecadd(&ts, &timeout);
493 		TIMESPEC_TO_TIMEVAL(&tv, &timeout);
494 	}
495 	mtx_lock_spin(&sched_lock);
496 	/* Change OURSELF to become an upcall. */
497 	td->td_flags = TDF_UPCALLING;
498 	if (p->p_sflag & PS_NEEDSIGCHK)
499 		td->td_flags |= TDF_ASTPENDING;
500 	mtx_unlock_spin(&sched_lock);
501 	PROC_LOCK(p);
502 	while ((td->td_upcall->ku_flags & KUF_DOUPCALL) == 0 &&
503 	       (kg->kg_completed == NULL)) {
504 		kg->kg_upsleeps++;
505 		error = msleep(&kg->kg_completed, &p->p_mtx, PPAUSE|PCATCH,
506 			"kse_rel", (uap->timeout ? tvtohz(&tv) : 0));
507 		kg->kg_upsleeps--;
508 		PROC_UNLOCK(p);
509 		if (uap->timeout == NULL || error != EWOULDBLOCK)
510 			return (0);
511 		getnanouptime(&ts2);
512 		if (timespeccmp(&ts2, &ts, >=))
513 			return (0);
514 		ts3 = ts;
515 		timespecsub(&ts3, &ts2);
516 		TIMESPEC_TO_TIMEVAL(&tv, &ts3);
517 		PROC_LOCK(p);
518 	}
519 	PROC_UNLOCK(p);
520 	return (0);
521 }
522 
523 /* struct kse_wakeup_args {
524 	struct kse_mailbox *mbx;
525 }; */
526 int
527 kse_wakeup(struct thread *td, struct kse_wakeup_args *uap)
528 {
529 	struct proc *p;
530 	struct ksegrp *kg;
531 	struct kse_upcall *ku;
532 	struct thread *td2;
533 
534 	p = td->td_proc;
535 	td2 = NULL;
536 	ku = NULL;
537 	/* KSE-enabled processes only, please. */
538 	if (!(p->p_flag & P_THREADED))
539 		return (EINVAL);
540 	PROC_LOCK(p);
541 	mtx_lock_spin(&sched_lock);
542 	if (uap->mbx) {
543 		FOREACH_KSEGRP_IN_PROC(p, kg) {
544 			FOREACH_UPCALL_IN_GROUP(kg, ku) {
545 				if (ku->ku_mailbox == uap->mbx)
546 					break;
547 			}
548 			if (ku)
549 				break;
550 		}
551 	} else {
552 		kg = td->td_ksegrp;
553 		if (kg->kg_upsleeps) {
554 			wakeup_one(&kg->kg_completed);
555 			mtx_unlock_spin(&sched_lock);
556 			PROC_UNLOCK(p);
557 			return (0);
558 		}
559 		ku = TAILQ_FIRST(&kg->kg_upcalls);
560 	}
561 	if (ku) {
562 		if ((td2 = ku->ku_owner) == NULL) {
563 			panic("%s: no owner", __func__);
564 		} else if (TD_ON_SLEEPQ(td2) &&
565 		           (td2->td_wchan == &kg->kg_completed)) {
566 			abortsleep(td2);
567 		} else {
568 			ku->ku_flags |= KUF_DOUPCALL;
569 		}
570 		mtx_unlock_spin(&sched_lock);
571 		PROC_UNLOCK(p);
572 		return (0);
573 	}
574 	mtx_unlock_spin(&sched_lock);
575 	PROC_UNLOCK(p);
576 	return (ESRCH);
577 }
578 
579 /*
580  * No new KSEG: first call: use current KSE, don't schedule an upcall
581  * All other situations, do allocate max new KSEs and schedule an upcall.
582  */
583 /* struct kse_create_args {
584 	struct kse_mailbox *mbx;
585 	int newgroup;
586 }; */
587 int
588 kse_create(struct thread *td, struct kse_create_args *uap)
589 {
590 	struct kse *newke;
591 	struct ksegrp *newkg;
592 	struct ksegrp *kg;
593 	struct proc *p;
594 	struct kse_mailbox mbx;
595 	struct kse_upcall *newku;
596 	int err, ncpus;
597 
598 	p = td->td_proc;
599 	if ((err = copyin(uap->mbx, &mbx, sizeof(mbx))))
600 		return (err);
601 
602 	/* Too bad, why hasn't kernel always a cpu counter !? */
603 #ifdef SMP
604 	ncpus = mp_ncpus;
605 #else
606 	ncpus = 1;
607 #endif
608 	if (thread_debug && virtual_cpu != 0)
609 		ncpus = virtual_cpu;
610 
611 	/* Easier to just set it than to test and set */
612 	p->p_flag |= P_THREADED;
613 	kg = td->td_ksegrp;
614 	if (uap->newgroup) {
615 		/* Have race condition but it is cheap */
616 		if (p->p_numksegrps >= max_groups_per_proc)
617 			return (EPROCLIM);
618 		/*
619 		 * If we want a new KSEGRP it doesn't matter whether
620 		 * we have already fired up KSE mode before or not.
621 		 * We put the process in KSE mode and create a new KSEGRP.
622 		 */
623 		newkg = ksegrp_alloc();
624 		bzero(&newkg->kg_startzero, RANGEOF(struct ksegrp,
625 		      kg_startzero, kg_endzero));
626 		bcopy(&kg->kg_startcopy, &newkg->kg_startcopy,
627 		      RANGEOF(struct ksegrp, kg_startcopy, kg_endcopy));
628 		mtx_lock_spin(&sched_lock);
629 		if (p->p_numksegrps >= max_groups_per_proc) {
630 			mtx_unlock_spin(&sched_lock);
631 			ksegrp_free(newkg);
632 			return (EPROCLIM);
633 		}
634 		ksegrp_link(newkg, p);
635 		mtx_unlock_spin(&sched_lock);
636 	} else {
637 		newkg = kg;
638 	}
639 
640 	/*
641 	 * Creating upcalls more than number of physical cpu does
642 	 * not help performance.
643 	 */
644 	if (newkg->kg_numupcalls >= ncpus)
645 		return (EPROCLIM);
646 
647 	if (newkg->kg_numupcalls == 0) {
648 		/*
649 		 * Initialize KSE group, optimized for MP.
650 		 * Create KSEs as many as physical cpus, this increases
651 		 * concurrent even if userland is not MP safe and can only run
652 		 * on single CPU (for early version of libpthread, it is true).
653 		 * In ideal world, every physical cpu should execute a thread.
654 		 * If there is enough KSEs, threads in kernel can be
655 		 * executed parallel on different cpus with full speed,
656 		 * Concurrent in kernel shouldn't be restricted by number of
657 		 * upcalls userland provides.
658 		 * Adding more upcall structures only increases concurrent
659 		 * in userland.
660 		 * Highest performance configuration is:
661 		 * N kses = N upcalls = N phyiscal cpus
662 		 */
663 		while (newkg->kg_kses < ncpus) {
664 			newke = kse_alloc();
665 			bzero(&newke->ke_startzero, RANGEOF(struct kse,
666 			      ke_startzero, ke_endzero));
667 #if 0
668 			mtx_lock_spin(&sched_lock);
669 			bcopy(&ke->ke_startcopy, &newke->ke_startcopy,
670 			      RANGEOF(struct kse, ke_startcopy, ke_endcopy));
671 			mtx_unlock_spin(&sched_lock);
672 #endif
673 			mtx_lock_spin(&sched_lock);
674 			kse_link(newke, newkg);
675 			/* Add engine */
676 			kse_reassign(newke);
677 			mtx_unlock_spin(&sched_lock);
678 		}
679 	}
680 	newku = upcall_alloc();
681 	newku->ku_mailbox = uap->mbx;
682 	newku->ku_func = mbx.km_func;
683 	bcopy(&mbx.km_stack, &newku->ku_stack, sizeof(stack_t));
684 
685 	/* For the first call this may not have been set */
686 	if (td->td_standin == NULL)
687 		thread_alloc_spare(td, NULL);
688 
689 	mtx_lock_spin(&sched_lock);
690 	if (newkg->kg_numupcalls >= ncpus) {
691 		mtx_unlock_spin(&sched_lock);
692 		upcall_free(newku);
693 		return (EPROCLIM);
694 	}
695 	upcall_link(newku, newkg);
696 
697 	/*
698 	 * Each upcall structure has an owner thread, find which
699 	 * one owns it.
700 	 */
701 	if (uap->newgroup) {
702 		/*
703 		 * Because new ksegrp hasn't thread,
704 		 * create an initial upcall thread to own it.
705 		 */
706 		thread_schedule_upcall(td, newku);
707 	} else {
708 		/*
709 		 * If current thread hasn't an upcall structure,
710 		 * just assign the upcall to it.
711 		 */
712 		if (td->td_upcall == NULL) {
713 			newku->ku_owner = td;
714 			td->td_upcall = newku;
715 		} else {
716 			/*
717 			 * Create a new upcall thread to own it.
718 			 */
719 			thread_schedule_upcall(td, newku);
720 		}
721 	}
722 	mtx_unlock_spin(&sched_lock);
723 	return (0);
724 }
725 
726 /*
727  * Fill a ucontext_t with a thread's context information.
728  *
729  * This is an analogue to getcontext(3).
730  */
731 void
732 thread_getcontext(struct thread *td, ucontext_t *uc)
733 {
734 
735 /*
736  * XXX this is declared in a MD include file, i386/include/ucontext.h but
737  * is used in MI code.
738  */
739 #ifdef __i386__
740 	get_mcontext(td, &uc->uc_mcontext);
741 #endif
742 	uc->uc_sigmask = td->td_proc->p_sigmask;
743 }
744 
745 /*
746  * Set a thread's context from a ucontext_t.
747  *
748  * This is an analogue to setcontext(3).
749  */
750 int
751 thread_setcontext(struct thread *td, ucontext_t *uc)
752 {
753 	int ret;
754 
755 /*
756  * XXX this is declared in a MD include file, i386/include/ucontext.h but
757  * is used in MI code.
758  */
759 #ifdef __i386__
760 	ret = set_mcontext(td, &uc->uc_mcontext);
761 #else
762 	ret = ENOSYS;
763 #endif
764 	if (ret == 0) {
765 		SIG_CANTMASK(uc->uc_sigmask);
766 		PROC_LOCK(td->td_proc);
767 		td->td_proc->p_sigmask = uc->uc_sigmask;
768 		PROC_UNLOCK(td->td_proc);
769 	}
770 	return (ret);
771 }
772 
773 /*
774  * Initialize global thread allocation resources.
775  */
776 void
777 threadinit(void)
778 {
779 
780 #ifndef __ia64__
781 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
782 	    thread_ctor, thread_dtor, thread_init, thread_fini,
783 	    UMA_ALIGN_CACHE, 0);
784 #else
785 	/*
786 	 * XXX the ia64 kstack allocator is really lame and is at the mercy
787 	 * of contigmallloc().  This hackery is to pre-construct a whole
788 	 * pile of thread structures with associated kernel stacks early
789 	 * in the system startup while contigmalloc() still works. Once we
790 	 * have them, keep them.  Sigh.
791 	 */
792 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
793 	    thread_ctor, thread_dtor, thread_init, thread_fini,
794 	    UMA_ALIGN_CACHE, UMA_ZONE_NOFREE);
795 	uma_prealloc(thread_zone, 512);		/* XXX arbitary */
796 #endif
797 	ksegrp_zone = uma_zcreate("KSEGRP", sched_sizeof_ksegrp(),
798 	    NULL, NULL, ksegrp_init, NULL,
799 	    UMA_ALIGN_CACHE, 0);
800 	kse_zone = uma_zcreate("KSE", sched_sizeof_kse(),
801 	    NULL, NULL, kse_init, NULL,
802 	    UMA_ALIGN_CACHE, 0);
803 	upcall_zone = uma_zcreate("UPCALL", sizeof(struct kse_upcall),
804 	    NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
805 }
806 
807 /*
808  * Stash an embarasingly extra thread into the zombie thread queue.
809  */
810 void
811 thread_stash(struct thread *td)
812 {
813 	mtx_lock_spin(&kse_zombie_lock);
814 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq);
815 	mtx_unlock_spin(&kse_zombie_lock);
816 }
817 
818 /*
819  * Stash an embarasingly extra kse into the zombie kse queue.
820  */
821 void
822 kse_stash(struct kse *ke)
823 {
824 	mtx_lock_spin(&kse_zombie_lock);
825 	TAILQ_INSERT_HEAD(&zombie_kses, ke, ke_procq);
826 	mtx_unlock_spin(&kse_zombie_lock);
827 }
828 
829 /*
830  * Stash an embarasingly extra upcall into the zombie upcall queue.
831  */
832 
833 void
834 upcall_stash(struct kse_upcall *ku)
835 {
836 	mtx_lock_spin(&kse_zombie_lock);
837 	TAILQ_INSERT_HEAD(&zombie_upcalls, ku, ku_link);
838 	mtx_unlock_spin(&kse_zombie_lock);
839 }
840 
841 /*
842  * Stash an embarasingly extra ksegrp into the zombie ksegrp queue.
843  */
844 void
845 ksegrp_stash(struct ksegrp *kg)
846 {
847 	mtx_lock_spin(&kse_zombie_lock);
848 	TAILQ_INSERT_HEAD(&zombie_ksegrps, kg, kg_ksegrp);
849 	mtx_unlock_spin(&kse_zombie_lock);
850 }
851 
852 /*
853  * Reap zombie kse resource.
854  */
855 void
856 thread_reap(void)
857 {
858 	struct thread *td_first, *td_next;
859 	struct kse *ke_first, *ke_next;
860 	struct ksegrp *kg_first, * kg_next;
861 	struct kse_upcall *ku_first, *ku_next;
862 
863 	/*
864 	 * Don't even bother to lock if none at this instant,
865 	 * we really don't care about the next instant..
866 	 */
867 	if ((!TAILQ_EMPTY(&zombie_threads))
868 	    || (!TAILQ_EMPTY(&zombie_kses))
869 	    || (!TAILQ_EMPTY(&zombie_ksegrps))
870 	    || (!TAILQ_EMPTY(&zombie_upcalls))) {
871 		mtx_lock_spin(&kse_zombie_lock);
872 		td_first = TAILQ_FIRST(&zombie_threads);
873 		ke_first = TAILQ_FIRST(&zombie_kses);
874 		kg_first = TAILQ_FIRST(&zombie_ksegrps);
875 		ku_first = TAILQ_FIRST(&zombie_upcalls);
876 		if (td_first)
877 			TAILQ_INIT(&zombie_threads);
878 		if (ke_first)
879 			TAILQ_INIT(&zombie_kses);
880 		if (kg_first)
881 			TAILQ_INIT(&zombie_ksegrps);
882 		if (ku_first)
883 			TAILQ_INIT(&zombie_upcalls);
884 		mtx_unlock_spin(&kse_zombie_lock);
885 		while (td_first) {
886 			td_next = TAILQ_NEXT(td_first, td_runq);
887 			if (td_first->td_ucred)
888 				crfree(td_first->td_ucred);
889 			thread_free(td_first);
890 			td_first = td_next;
891 		}
892 		while (ke_first) {
893 			ke_next = TAILQ_NEXT(ke_first, ke_procq);
894 			kse_free(ke_first);
895 			ke_first = ke_next;
896 		}
897 		while (kg_first) {
898 			kg_next = TAILQ_NEXT(kg_first, kg_ksegrp);
899 			ksegrp_free(kg_first);
900 			kg_first = kg_next;
901 		}
902 		while (ku_first) {
903 			ku_next = TAILQ_NEXT(ku_first, ku_link);
904 			upcall_free(ku_first);
905 			ku_first = ku_next;
906 		}
907 	}
908 }
909 
910 /*
911  * Allocate a ksegrp.
912  */
913 struct ksegrp *
914 ksegrp_alloc(void)
915 {
916 	return (uma_zalloc(ksegrp_zone, M_WAITOK));
917 }
918 
919 /*
920  * Allocate a kse.
921  */
922 struct kse *
923 kse_alloc(void)
924 {
925 	return (uma_zalloc(kse_zone, M_WAITOK));
926 }
927 
928 /*
929  * Allocate a thread.
930  */
931 struct thread *
932 thread_alloc(void)
933 {
934 	thread_reap(); /* check if any zombies to get */
935 	return (uma_zalloc(thread_zone, M_WAITOK));
936 }
937 
938 /*
939  * Deallocate a ksegrp.
940  */
941 void
942 ksegrp_free(struct ksegrp *td)
943 {
944 	uma_zfree(ksegrp_zone, td);
945 }
946 
947 /*
948  * Deallocate a kse.
949  */
950 void
951 kse_free(struct kse *td)
952 {
953 	uma_zfree(kse_zone, td);
954 }
955 
956 /*
957  * Deallocate a thread.
958  */
959 void
960 thread_free(struct thread *td)
961 {
962 
963 	cpu_thread_clean(td);
964 	uma_zfree(thread_zone, td);
965 }
966 
967 /*
968  * Store the thread context in the UTS's mailbox.
969  * then add the mailbox at the head of a list we are building in user space.
970  * The list is anchored in the ksegrp structure.
971  */
972 int
973 thread_export_context(struct thread *td)
974 {
975 	struct proc *p;
976 	struct ksegrp *kg;
977 	uintptr_t mbx;
978 	void *addr;
979 	int error,temp;
980 	ucontext_t uc;
981 
982 	p = td->td_proc;
983 	kg = td->td_ksegrp;
984 
985 	/* Export the user/machine context. */
986 	addr = (void *)(&td->td_mailbox->tm_context);
987 	error = copyin(addr, &uc, sizeof(ucontext_t));
988 	if (error)
989 		goto bad;
990 
991 	thread_getcontext(td, &uc);
992 	error = copyout(&uc, addr, sizeof(ucontext_t));
993 	if (error)
994 		goto bad;
995 
996 	/* Exports clock ticks in kernel mode */
997 	addr = (caddr_t)(&td->td_mailbox->tm_sticks);
998 	temp = fuword(addr) + td->td_usticks;
999 	if (suword(addr, temp))
1000 		goto bad;
1001 
1002 	addr = (caddr_t)(&td->td_mailbox->tm_slices);
1003 	temp = fuword(addr) - td->td_usticks;
1004 	if (suword(addr, temp))
1005 		goto bad;
1006 
1007 	/* Get address in latest mbox of list pointer */
1008 	addr = (void *)(&td->td_mailbox->tm_next);
1009 	/*
1010 	 * Put the saved address of the previous first
1011 	 * entry into this one
1012 	 */
1013 	for (;;) {
1014 		mbx = (uintptr_t)kg->kg_completed;
1015 		if (suword(addr, mbx)) {
1016 			error = EFAULT;
1017 			goto bad;
1018 		}
1019 		PROC_LOCK(p);
1020 		if (mbx == (uintptr_t)kg->kg_completed) {
1021 			kg->kg_completed = td->td_mailbox;
1022 			/*
1023 			 * The thread context may be taken away by
1024 			 * other upcall threads when we unlock
1025 			 * process lock. it's no longer valid to
1026 			 * use it again in any other places.
1027 			 */
1028 			td->td_mailbox = NULL;
1029 			PROC_UNLOCK(p);
1030 			break;
1031 		}
1032 		PROC_UNLOCK(p);
1033 	}
1034 	td->td_usticks = 0;
1035 	return (0);
1036 
1037 bad:
1038 	PROC_LOCK(p);
1039 	psignal(p, SIGSEGV);
1040 	PROC_UNLOCK(p);
1041 	/* The mailbox is bad, don't use it */
1042 	td->td_mailbox = NULL;
1043 	td->td_usticks = 0;
1044 	return (error);
1045 }
1046 
1047 /*
1048  * Take the list of completed mailboxes for this KSEGRP and put them on this
1049  * upcall's mailbox as it's the next one going up.
1050  */
1051 static int
1052 thread_link_mboxes(struct ksegrp *kg, struct kse_upcall *ku)
1053 {
1054 	struct proc *p = kg->kg_proc;
1055 	void *addr;
1056 	uintptr_t mbx;
1057 
1058 	addr = (void *)(&ku->ku_mailbox->km_completed);
1059 	for (;;) {
1060 		mbx = (uintptr_t)kg->kg_completed;
1061 		if (suword(addr, mbx)) {
1062 			PROC_LOCK(p);
1063 			psignal(p, SIGSEGV);
1064 			PROC_UNLOCK(p);
1065 			return (EFAULT);
1066 		}
1067 		PROC_LOCK(p);
1068 		if (mbx == (uintptr_t)kg->kg_completed) {
1069 			kg->kg_completed = NULL;
1070 			PROC_UNLOCK(p);
1071 			break;
1072 		}
1073 		PROC_UNLOCK(p);
1074 	}
1075 	return (0);
1076 }
1077 
1078 /*
1079  * This function should be called at statclock interrupt time
1080  */
1081 int
1082 thread_statclock(int user)
1083 {
1084 	struct thread *td = curthread;
1085 
1086 	if (td->td_ksegrp->kg_numupcalls == 0)
1087 		return (-1);
1088 	if (user) {
1089 		/* Current always do via ast() */
1090 		td->td_flags |= (TDF_USTATCLOCK|TDF_ASTPENDING);
1091 		td->td_uuticks++;
1092 	} else {
1093 		if (td->td_mailbox != NULL)
1094 			td->td_usticks++;
1095 		else {
1096 			/* XXXKSE
1097 		 	 * We will call thread_user_enter() for every
1098 			 * kernel entry in future, so if the thread mailbox
1099 			 * is NULL, it must be a UTS kernel, don't account
1100 			 * clock ticks for it.
1101 			 */
1102 		}
1103 	}
1104 	return (0);
1105 }
1106 
1107 /*
1108  * Export state clock ticks for userland
1109  */
1110 static int
1111 thread_update_usr_ticks(struct thread *td, int user)
1112 {
1113 	struct proc *p = td->td_proc;
1114 	struct kse_thr_mailbox *tmbx;
1115 	struct kse_upcall *ku;
1116 	caddr_t addr;
1117 	uint uticks;
1118 	int slices;
1119 
1120 	if ((ku = td->td_upcall) == NULL)
1121 		return (-1);
1122 
1123 	tmbx = (void *)fuword((void *)&ku->ku_mailbox->km_curthread);
1124 	if ((tmbx == NULL) || (tmbx == (void *)-1))
1125 		return (-1);
1126 	if (user) {
1127 		uticks = td->td_uuticks;
1128 		td->td_uuticks = 0;
1129 		addr = (caddr_t)&tmbx->tm_uticks;
1130 	} else {
1131 		uticks = td->td_usticks;
1132 		td->td_usticks = 0;
1133 		addr = (caddr_t)&tmbx->tm_sticks;
1134 	}
1135 	if (uticks) {
1136 		if (suword(addr, uticks+fuword(addr))) {
1137 			PROC_LOCK(p);
1138 			psignal(p, SIGSEGV);
1139 			PROC_UNLOCK(p);
1140 			return (-2);
1141 		}
1142 		addr = (caddr_t)&tmbx->tm_slices;
1143 		slices = (int)fuword(addr);
1144 		if (slices > 0) {
1145 			slices -= (int)uticks;
1146 			if (suword(addr, slices)) {
1147 				PROC_LOCK(p);
1148 				psignal(p, SIGSEGV);
1149 				PROC_UNLOCK(p);
1150 				return (-2);
1151 			}
1152 			if (slices <= 0) {
1153 				mtx_lock_spin(&sched_lock);
1154 				td->td_upcall->ku_flags |= KUF_DOUPCALL;
1155 				mtx_unlock_spin(&sched_lock);
1156 			}
1157 		}
1158 	}
1159 	return (0);
1160 }
1161 
1162 /*
1163  * Discard the current thread and exit from its context.
1164  *
1165  * Because we can't free a thread while we're operating under its context,
1166  * push the current thread into our CPU's deadthread holder. This means
1167  * we needn't worry about someone else grabbing our context before we
1168  * do a cpu_throw().
1169  */
1170 void
1171 thread_exit(void)
1172 {
1173 	struct thread *td;
1174 	struct kse *ke;
1175 	struct proc *p;
1176 	struct ksegrp	*kg;
1177 
1178 	td = curthread;
1179 	kg = td->td_ksegrp;
1180 	p = td->td_proc;
1181 	ke = td->td_kse;
1182 
1183 	mtx_assert(&sched_lock, MA_OWNED);
1184 	KASSERT(p != NULL, ("thread exiting without a process"));
1185 	KASSERT(ke != NULL, ("thread exiting without a kse"));
1186 	KASSERT(kg != NULL, ("thread exiting without a kse group"));
1187 	PROC_LOCK_ASSERT(p, MA_OWNED);
1188 	CTR1(KTR_PROC, "thread_exit: thread %p", td);
1189 	KASSERT(!mtx_owned(&Giant), ("dying thread owns giant"));
1190 
1191 	if (td->td_standin != NULL) {
1192 		thread_stash(td->td_standin);
1193 		td->td_standin = NULL;
1194 	}
1195 
1196 	cpu_thread_exit(td);	/* XXXSMP */
1197 
1198 	/*
1199 	 * The last thread is left attached to the process
1200 	 * So that the whole bundle gets recycled. Skip
1201 	 * all this stuff.
1202 	 */
1203 	if (p->p_numthreads > 1) {
1204 		/*
1205 		 * Unlink this thread from its proc and the kseg.
1206 		 * In keeping with the other structs we probably should
1207 		 * have a thread_unlink() that does some of this but it
1208 		 * would only be called from here (I think) so it would
1209 		 * be a waste. (might be useful for proc_fini() as well.)
1210  		 */
1211 		TAILQ_REMOVE(&p->p_threads, td, td_plist);
1212 		p->p_numthreads--;
1213 		TAILQ_REMOVE(&kg->kg_threads, td, td_kglist);
1214 		kg->kg_numthreads--;
1215 		if (p->p_maxthrwaits)
1216 			wakeup(&p->p_numthreads);
1217 		/*
1218 		 * The test below is NOT true if we are the
1219 		 * sole exiting thread. P_STOPPED_SNGL is unset
1220 		 * in exit1() after it is the only survivor.
1221 		 */
1222 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1223 			if (p->p_numthreads == p->p_suspcount) {
1224 				thread_unsuspend_one(p->p_singlethread);
1225 			}
1226 		}
1227 
1228 		/*
1229 		 * Because each upcall structure has an owner thread,
1230 		 * owner thread exits only when process is in exiting
1231 		 * state, so upcall to userland is no longer needed,
1232 		 * deleting upcall structure is safe here.
1233 		 * So when all threads in a group is exited, all upcalls
1234 		 * in the group should be automatically freed.
1235 		 */
1236 		if (td->td_upcall)
1237 			upcall_remove(td);
1238 
1239 		ke->ke_state = KES_UNQUEUED;
1240 		ke->ke_thread = NULL;
1241 		/*
1242 		 * Decide what to do with the KSE attached to this thread.
1243 		 */
1244 		if (ke->ke_flags & KEF_EXIT)
1245 			kse_unlink(ke);
1246 		else
1247 			kse_reassign(ke);
1248 		PROC_UNLOCK(p);
1249 		td->td_kse	= NULL;
1250 		td->td_state	= TDS_INACTIVE;
1251 		td->td_proc	= NULL;
1252 		td->td_ksegrp	= NULL;
1253 		td->td_last_kse	= NULL;
1254 		PCPU_SET(deadthread, td);
1255 	} else {
1256 		PROC_UNLOCK(p);
1257 	}
1258 	cpu_throw();
1259 	/* NOTREACHED */
1260 }
1261 
1262 /*
1263  * Do any thread specific cleanups that may be needed in wait()
1264  * called with Giant held, proc and schedlock not held.
1265  */
1266 void
1267 thread_wait(struct proc *p)
1268 {
1269 	struct thread *td;
1270 
1271 	KASSERT((p->p_numthreads == 1), ("Muliple threads in wait1()"));
1272 	KASSERT((p->p_numksegrps == 1), ("Muliple ksegrps in wait1()"));
1273 	FOREACH_THREAD_IN_PROC(p, td) {
1274 		if (td->td_standin != NULL) {
1275 			thread_free(td->td_standin);
1276 			td->td_standin = NULL;
1277 		}
1278 		cpu_thread_clean(td);
1279 	}
1280 	thread_reap();	/* check for zombie threads etc. */
1281 }
1282 
1283 /*
1284  * Link a thread to a process.
1285  * set up anything that needs to be initialized for it to
1286  * be used by the process.
1287  *
1288  * Note that we do not link to the proc's ucred here.
1289  * The thread is linked as if running but no KSE assigned.
1290  */
1291 void
1292 thread_link(struct thread *td, struct ksegrp *kg)
1293 {
1294 	struct proc *p;
1295 
1296 	p = kg->kg_proc;
1297 	td->td_state    = TDS_INACTIVE;
1298 	td->td_proc     = p;
1299 	td->td_ksegrp   = kg;
1300 	td->td_last_kse = NULL;
1301 	td->td_flags    = 0;
1302 	td->td_kse      = NULL;
1303 
1304 	LIST_INIT(&td->td_contested);
1305 	callout_init(&td->td_slpcallout, 1);
1306 	TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
1307 	TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist);
1308 	p->p_numthreads++;
1309 	kg->kg_numthreads++;
1310 }
1311 
1312 /*
1313  * Purge a ksegrp resource. When a ksegrp is preparing to
1314  * exit, it calls this function.
1315  */
1316 void
1317 kse_purge_group(struct thread *td)
1318 {
1319 	struct ksegrp *kg;
1320 	struct kse *ke;
1321 
1322 	kg = td->td_ksegrp;
1323  	KASSERT(kg->kg_numthreads == 1, ("%s: bad thread number", __func__));
1324 	while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) {
1325 		KASSERT(ke->ke_state == KES_IDLE,
1326 			("%s: wrong idle KSE state", __func__));
1327 		kse_unlink(ke);
1328 	}
1329 	KASSERT((kg->kg_kses == 1),
1330 		("%s: ksegrp still has %d KSEs", __func__, kg->kg_kses));
1331 	KASSERT((kg->kg_numupcalls == 0),
1332 	        ("%s: ksegrp still has %d upcall datas",
1333 		__func__, kg->kg_numupcalls));
1334 }
1335 
1336 /*
1337  * Purge a process's KSE resource. When a process is preparing to
1338  * exit, it calls kse_purge to release any extra KSE resources in
1339  * the process.
1340  */
1341 void
1342 kse_purge(struct proc *p, struct thread *td)
1343 {
1344 	struct ksegrp *kg;
1345 	struct kse *ke;
1346 
1347  	KASSERT(p->p_numthreads == 1, ("bad thread number"));
1348 	mtx_lock_spin(&sched_lock);
1349 	while ((kg = TAILQ_FIRST(&p->p_ksegrps)) != NULL) {
1350 		TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
1351 		p->p_numksegrps--;
1352 		/*
1353 		 * There is no ownership for KSE, after all threads
1354 		 * in the group exited, it is possible that some KSEs
1355 		 * were left in idle queue, gc them now.
1356 		 */
1357 		while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) {
1358 			KASSERT(ke->ke_state == KES_IDLE,
1359 			   ("%s: wrong idle KSE state", __func__));
1360 			TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist);
1361 			kg->kg_idle_kses--;
1362 			TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist);
1363 			kg->kg_kses--;
1364 			kse_stash(ke);
1365 		}
1366 		KASSERT(((kg->kg_kses == 0) && (kg != td->td_ksegrp)) ||
1367 		        ((kg->kg_kses == 1) && (kg == td->td_ksegrp)),
1368 		        ("ksegrp has wrong kg_kses: %d", kg->kg_kses));
1369 		KASSERT((kg->kg_numupcalls == 0),
1370 		        ("%s: ksegrp still has %d upcall datas",
1371 			__func__, kg->kg_numupcalls));
1372 
1373 		if (kg != td->td_ksegrp)
1374 			ksegrp_stash(kg);
1375 	}
1376 	TAILQ_INSERT_HEAD(&p->p_ksegrps, td->td_ksegrp, kg_ksegrp);
1377 	p->p_numksegrps++;
1378 	mtx_unlock_spin(&sched_lock);
1379 }
1380 
1381 /*
1382  * This function is intended to be used to initialize a spare thread
1383  * for upcall. Initialize thread's large data area outside sched_lock
1384  * for thread_schedule_upcall().
1385  */
1386 void
1387 thread_alloc_spare(struct thread *td, struct thread *spare)
1388 {
1389 	if (td->td_standin)
1390 		return;
1391 	if (spare == NULL)
1392 		spare = thread_alloc();
1393 	td->td_standin = spare;
1394 	bzero(&spare->td_startzero,
1395 	    (unsigned)RANGEOF(struct thread, td_startzero, td_endzero));
1396 	spare->td_proc = td->td_proc;
1397 	/* Setup PCB and fork address */
1398 	cpu_set_upcall(spare, td->td_pcb);
1399 	/*
1400 	 * XXXKSE do we really need this? (default values for the
1401 	 * frame).
1402 	 */
1403 	bcopy(td->td_frame, spare->td_frame, sizeof(struct trapframe));
1404 	spare->td_ucred = crhold(td->td_ucred);
1405 }
1406 
1407 /*
1408  * Create a thread and schedule it for upcall on the KSE given.
1409  * Use our thread's standin so that we don't have to allocate one.
1410  */
1411 struct thread *
1412 thread_schedule_upcall(struct thread *td, struct kse_upcall *ku)
1413 {
1414 	struct thread *td2;
1415 
1416 	mtx_assert(&sched_lock, MA_OWNED);
1417 
1418 	/*
1419 	 * Schedule an upcall thread on specified kse_upcall,
1420 	 * the kse_upcall must be free.
1421 	 * td must have a spare thread.
1422 	 */
1423 	KASSERT(ku->ku_owner == NULL, ("%s: upcall has owner", __func__));
1424 	if ((td2 = td->td_standin) != NULL) {
1425 		td->td_standin = NULL;
1426 	} else {
1427 		panic("no reserve thread when scheduling an upcall");
1428 		return (NULL);
1429 	}
1430 	CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)",
1431 	     td2, td->td_proc->p_pid, td->td_proc->p_comm);
1432 	bcopy(&td->td_startcopy, &td2->td_startcopy,
1433 	    (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy));
1434 	thread_link(td2, ku->ku_ksegrp);
1435 	/* Let the new thread become owner of the upcall */
1436 	ku->ku_owner   = td2;
1437 	td2->td_upcall = ku;
1438 	td2->td_flags  = TDF_UPCALLING;
1439 	if (td->td_proc->p_sflag & PS_NEEDSIGCHK)
1440 		td2->td_flags |= TDF_ASTPENDING;
1441 	td2->td_kse    = NULL;
1442 	td2->td_state  = TDS_CAN_RUN;
1443 	td2->td_inhibitors = 0;
1444 	setrunqueue(td2);
1445 	return (td2);	/* bogus.. should be a void function */
1446 }
1447 
1448 void
1449 thread_signal_add(struct thread *td, int sig)
1450 {
1451 	struct kse_upcall *ku;
1452 	struct proc *p;
1453 	sigset_t ss;
1454 	int error;
1455 
1456 	PROC_LOCK_ASSERT(td->td_proc, MA_OWNED);
1457 	td = curthread;
1458 	ku = td->td_upcall;
1459 	p = td->td_proc;
1460 
1461 	PROC_UNLOCK(p);
1462 	error = copyin(&ku->ku_mailbox->km_sigscaught, &ss, sizeof(sigset_t));
1463 	if (error)
1464 		goto error;
1465 
1466 	SIGADDSET(ss, sig);
1467 
1468 	error = copyout(&ss, &ku->ku_mailbox->km_sigscaught, sizeof(sigset_t));
1469 	if (error)
1470 		goto error;
1471 
1472 	PROC_LOCK(p);
1473 	return;
1474 error:
1475 	PROC_LOCK(p);
1476 	sigexit(td, SIGILL);
1477 }
1478 
1479 
1480 /*
1481  * Schedule an upcall to notify a KSE process recieved signals.
1482  *
1483  */
1484 void
1485 thread_signal_upcall(struct thread *td)
1486 {
1487 	mtx_lock_spin(&sched_lock);
1488 	td->td_flags |= TDF_UPCALLING;
1489 	mtx_unlock_spin(&sched_lock);
1490 
1491 	return;
1492 }
1493 
1494 /*
1495  * Setup done on the thread when it enters the kernel.
1496  * XXXKSE Presently only for syscalls but eventually all kernel entries.
1497  */
1498 void
1499 thread_user_enter(struct proc *p, struct thread *td)
1500 {
1501 	struct ksegrp *kg;
1502 	struct kse_upcall *ku;
1503 
1504 	kg = td->td_ksegrp;
1505 	/*
1506 	 * First check that we shouldn't just abort.
1507 	 * But check if we are the single thread first!
1508 	 * XXX p_singlethread not locked, but should be safe.
1509 	 */
1510 	if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1511 		PROC_LOCK(p);
1512 		mtx_lock_spin(&sched_lock);
1513 		thread_exit();
1514 		/* NOTREACHED */
1515 	}
1516 
1517 	/*
1518 	 * If we are doing a syscall in a KSE environment,
1519 	 * note where our mailbox is. There is always the
1520 	 * possibility that we could do this lazily (in kse_reassign()),
1521 	 * but for now do it every time.
1522 	 */
1523 	kg = td->td_ksegrp;
1524 	if (kg->kg_numupcalls) {
1525 		ku = td->td_upcall;
1526 		KASSERT(ku, ("%s: no upcall owned", __func__));
1527 		KASSERT((ku->ku_owner == td), ("%s: wrong owner", __func__));
1528 		td->td_mailbox =
1529 		    (void *)fuword((void *)&ku->ku_mailbox->km_curthread);
1530 		if ((td->td_mailbox == NULL) ||
1531 		    (td->td_mailbox == (void *)-1)) {
1532 		    	/* Don't schedule upcall when blocked */
1533 			td->td_mailbox = NULL;
1534 			mtx_lock_spin(&sched_lock);
1535 			td->td_flags &= ~TDF_CAN_UNBIND;
1536 			mtx_unlock_spin(&sched_lock);
1537 		} else {
1538 			if (td->td_standin == NULL)
1539 				thread_alloc_spare(td, NULL);
1540 			mtx_lock_spin(&sched_lock);
1541 			td->td_flags |= TDF_CAN_UNBIND;
1542 			mtx_unlock_spin(&sched_lock);
1543 		}
1544 	}
1545 }
1546 
1547 /*
1548  * The extra work we go through if we are a threaded process when we
1549  * return to userland.
1550  *
1551  * If we are a KSE process and returning to user mode, check for
1552  * extra work to do before we return (e.g. for more syscalls
1553  * to complete first).  If we were in a critical section, we should
1554  * just return to let it finish. Same if we were in the UTS (in
1555  * which case the mailbox's context's busy indicator will be set).
1556  * The only traps we suport will have set the mailbox.
1557  * We will clear it here.
1558  */
1559 int
1560 thread_userret(struct thread *td, struct trapframe *frame)
1561 {
1562 	int error = 0, upcalls;
1563 	struct kse_upcall *ku;
1564 	struct ksegrp *kg, *kg2;
1565 	struct proc *p;
1566 	struct timespec ts;
1567 
1568 	p = td->td_proc;
1569 	kg = td->td_ksegrp;
1570 
1571 	/* Nothing to do with non-threaded group/process */
1572 	if (td->td_ksegrp->kg_numupcalls == 0)
1573 		return (0);
1574 
1575 	/*
1576 	 * Stat clock interrupt hit in userland, it
1577 	 * is returning from interrupt, charge thread's
1578 	 * userland time for UTS.
1579 	 */
1580 	if (td->td_flags & TDF_USTATCLOCK) {
1581 		thread_update_usr_ticks(td, 1);
1582 		mtx_lock_spin(&sched_lock);
1583 		td->td_flags &= ~TDF_USTATCLOCK;
1584 		mtx_unlock_spin(&sched_lock);
1585 		if (kg->kg_completed ||
1586 		    (td->td_upcall->ku_flags & KUF_DOUPCALL))
1587 			thread_user_enter(p, td);
1588 	}
1589 
1590 	/*
1591 	 * Optimisation:
1592 	 * This thread has not started any upcall.
1593 	 * If there is no work to report other than ourself,
1594 	 * then it can return direct to userland.
1595 	 */
1596 	if (TD_CAN_UNBIND(td)) {
1597 		mtx_lock_spin(&sched_lock);
1598 		td->td_flags &= ~TDF_CAN_UNBIND;
1599 		mtx_unlock_spin(&sched_lock);
1600 		if ((kg->kg_completed == NULL) &&
1601 		    (td->td_upcall->ku_flags & KUF_DOUPCALL) == 0) {
1602 			thread_update_usr_ticks(td, 0);
1603 			if (!(kg->kg_completed ||
1604 			    (td->td_upcall->ku_flags & KUF_DOUPCALL))) {
1605 				td->td_mailbox = NULL;
1606 				return (0);
1607 			}
1608 		}
1609 		error = thread_export_context(td);
1610 		if (error) {
1611 			/*
1612 			 * Failing to do the KSE operation just defaults
1613 			 * back to synchonous operation, so just return from
1614 			 * the syscall.
1615 			 */
1616 			return (0);
1617 		}
1618 		/*
1619 		 * There is something to report, and we own an upcall
1620 		 * strucuture, we can go to userland.
1621 		 * Turn ourself into an upcall thread.
1622 		 */
1623 		mtx_lock_spin(&sched_lock);
1624 		td->td_flags |= TDF_UPCALLING;
1625 		mtx_unlock_spin(&sched_lock);
1626 	} else if (td->td_mailbox) {
1627 		error = thread_export_context(td);
1628 		if (error) {
1629 			PROC_LOCK(td->td_proc);
1630 			mtx_lock_spin(&sched_lock);
1631 			/* possibly upcall with error? */
1632 		} else {
1633 			PROC_LOCK(td->td_proc);
1634 			mtx_lock_spin(&sched_lock);
1635 			/*
1636 			 * There are upcall threads waiting for
1637 			 * work to do, wake one of them up.
1638 			 * XXXKSE Maybe wake all of them up.
1639 			 */
1640 			if (kg->kg_upsleeps)
1641 				wakeup_one(&kg->kg_completed);
1642 		}
1643 		thread_exit();
1644 		/* NOTREACHED */
1645 	}
1646 
1647 	KASSERT(TD_CAN_UNBIND(td) == 0, ("can unbind"));
1648 
1649 	if (p->p_numthreads > max_threads_per_proc) {
1650 		max_threads_hits++;
1651 		PROC_LOCK(p);
1652 		while (p->p_numthreads > max_threads_per_proc) {
1653 			if (P_SHOULDSTOP(p))
1654 				break;
1655 			upcalls = 0;
1656 			mtx_lock_spin(&sched_lock);
1657 			FOREACH_KSEGRP_IN_PROC(p, kg2) {
1658 				if (kg2->kg_numupcalls == 0)
1659 					upcalls++;
1660 				else
1661 					upcalls += kg2->kg_numupcalls;
1662 			}
1663 			mtx_unlock_spin(&sched_lock);
1664 			if (upcalls >= max_threads_per_proc)
1665 				break;
1666 			p->p_maxthrwaits++;
1667 			msleep(&p->p_numthreads, &p->p_mtx, PPAUSE|PCATCH,
1668 			    "maxthreads", NULL);
1669 			p->p_maxthrwaits--;
1670 		}
1671 		PROC_UNLOCK(p);
1672 	}
1673 
1674 	if (td->td_flags & TDF_UPCALLING) {
1675 		ku = td->td_upcall;
1676 		/*
1677 		 * There is no more work to do and we are going to ride
1678 		 * this thread up to userland as an upcall.
1679 		 * Do the last parts of the setup needed for the upcall.
1680 		 */
1681 		CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)",
1682 		    td, td->td_proc->p_pid, td->td_proc->p_comm);
1683 
1684 		/*
1685 		 * Set user context to the UTS.
1686 		 * Will use Giant in cpu_thread_clean() because it uses
1687 		 * kmem_free(kernel_map, ...)
1688 		 */
1689 		cpu_set_upcall_kse(td, ku);
1690 		mtx_lock_spin(&sched_lock);
1691 		td->td_flags &= ~TDF_UPCALLING;
1692 		if (ku->ku_flags & KUF_DOUPCALL)
1693 			ku->ku_flags &= ~KUF_DOUPCALL;
1694 		mtx_unlock_spin(&sched_lock);
1695 
1696 		/*
1697 		 * Unhook the list of completed threads.
1698 		 * anything that completes after this gets to
1699 		 * come in next time.
1700 		 * Put the list of completed thread mailboxes on
1701 		 * this KSE's mailbox.
1702 		 */
1703 		error = thread_link_mboxes(kg, ku);
1704 		if (error)
1705 			goto out;
1706 
1707 		/*
1708 		 * Set state and clear the  thread mailbox pointer.
1709 		 * From now on we are just a bound outgoing process.
1710 		 * **Problem** userret is often called several times.
1711 		 * it would be nice if this all happenned only on the first
1712 		 * time through. (the scan for extra work etc.)
1713 		 */
1714 		error = suword((caddr_t)&ku->ku_mailbox->km_curthread, 0);
1715 		if (error)
1716 			goto out;
1717 
1718 		/* Export current system time */
1719 		nanotime(&ts);
1720 		error = copyout(&ts, (caddr_t)&ku->ku_mailbox->km_timeofday,
1721 			sizeof(ts));
1722 	}
1723 
1724 out:
1725 	if (error) {
1726 		/*
1727 		 * Things are going to be so screwed we should just kill
1728 		 * the process.
1729 		 * how do we do that?
1730 		 */
1731 		PROC_LOCK(td->td_proc);
1732 		psignal(td->td_proc, SIGSEGV);
1733 		PROC_UNLOCK(td->td_proc);
1734 	} else {
1735 		/*
1736 		 * Optimisation:
1737 		 * Ensure that we have a spare thread available,
1738 		 * for when we re-enter the kernel.
1739 		 */
1740 		if (td->td_standin == NULL)
1741 			thread_alloc_spare(td, NULL);
1742 	}
1743 
1744 	/*
1745 	 * Clear thread mailbox first, then clear system tick count.
1746 	 * The order is important because thread_statclock() use
1747 	 * mailbox pointer to see if it is an userland thread or
1748 	 * an UTS kernel thread.
1749 	 */
1750 	td->td_mailbox = NULL;
1751 	td->td_usticks = 0;
1752 	return (error);	/* go sync */
1753 }
1754 
1755 /*
1756  * Enforce single-threading.
1757  *
1758  * Returns 1 if the caller must abort (another thread is waiting to
1759  * exit the process or similar). Process is locked!
1760  * Returns 0 when you are successfully the only thread running.
1761  * A process has successfully single threaded in the suspend mode when
1762  * There are no threads in user mode. Threads in the kernel must be
1763  * allowed to continue until they get to the user boundary. They may even
1764  * copy out their return values and data before suspending. They may however be
1765  * accellerated in reaching the user boundary as we will wake up
1766  * any sleeping threads that are interruptable. (PCATCH).
1767  */
1768 int
1769 thread_single(int force_exit)
1770 {
1771 	struct thread *td;
1772 	struct thread *td2;
1773 	struct proc *p;
1774 
1775 	td = curthread;
1776 	p = td->td_proc;
1777 	mtx_assert(&Giant, MA_OWNED);
1778 	PROC_LOCK_ASSERT(p, MA_OWNED);
1779 	KASSERT((td != NULL), ("curthread is NULL"));
1780 
1781 	if ((p->p_flag & P_THREADED) == 0)
1782 		return (0);
1783 
1784 	/* Is someone already single threading? */
1785 	if (p->p_singlethread)
1786 		return (1);
1787 
1788 	if (force_exit == SINGLE_EXIT) {
1789 		p->p_flag |= P_SINGLE_EXIT;
1790 	} else
1791 		p->p_flag &= ~P_SINGLE_EXIT;
1792 	p->p_flag |= P_STOPPED_SINGLE;
1793 	p->p_singlethread = td;
1794 	/* XXXKSE Which lock protects the below values? */
1795 	while ((p->p_numthreads - p->p_suspcount) != 1) {
1796 		mtx_lock_spin(&sched_lock);
1797 		FOREACH_THREAD_IN_PROC(p, td2) {
1798 			if (td2 == td)
1799 				continue;
1800 			td->td_flags |= TDF_ASTPENDING;
1801 			if (TD_IS_INHIBITED(td2)) {
1802 				if (force_exit == SINGLE_EXIT) {
1803 					if (TD_IS_SUSPENDED(td2)) {
1804 						thread_unsuspend_one(td2);
1805 					}
1806 					if (TD_ON_SLEEPQ(td2) &&
1807 					    (td2->td_flags & TDF_SINTR)) {
1808 						if (td2->td_flags & TDF_CVWAITQ)
1809 							cv_abort(td2);
1810 						else
1811 							abortsleep(td2);
1812 					}
1813 				} else {
1814 					if (TD_IS_SUSPENDED(td2))
1815 						continue;
1816 					/*
1817 					 * maybe other inhibitted states too?
1818 					 * XXXKSE Is it totally safe to
1819 					 * suspend a non-interruptable thread?
1820 					 */
1821 					if (td2->td_inhibitors &
1822 					    (TDI_SLEEPING | TDI_SWAPPED))
1823 						thread_suspend_one(td2);
1824 				}
1825 			}
1826 		}
1827 		/*
1828 		 * Maybe we suspended some threads.. was it enough?
1829 		 */
1830 		if ((p->p_numthreads - p->p_suspcount) == 1) {
1831 			mtx_unlock_spin(&sched_lock);
1832 			break;
1833 		}
1834 
1835 		/*
1836 		 * Wake us up when everyone else has suspended.
1837 		 * In the mean time we suspend as well.
1838 		 */
1839 		thread_suspend_one(td);
1840 		mtx_unlock(&Giant);
1841 		PROC_UNLOCK(p);
1842 		p->p_stats->p_ru.ru_nvcsw++;
1843 		mi_switch();
1844 		mtx_unlock_spin(&sched_lock);
1845 		mtx_lock(&Giant);
1846 		PROC_LOCK(p);
1847 	}
1848 	if (force_exit == SINGLE_EXIT) {
1849 		if (td->td_upcall) {
1850 			mtx_lock_spin(&sched_lock);
1851 			upcall_remove(td);
1852 			mtx_unlock_spin(&sched_lock);
1853 		}
1854 		kse_purge(p, td);
1855 	}
1856 	return (0);
1857 }
1858 
1859 /*
1860  * Called in from locations that can safely check to see
1861  * whether we have to suspend or at least throttle for a
1862  * single-thread event (e.g. fork).
1863  *
1864  * Such locations include userret().
1865  * If the "return_instead" argument is non zero, the thread must be able to
1866  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1867  *
1868  * The 'return_instead' argument tells the function if it may do a
1869  * thread_exit() or suspend, or whether the caller must abort and back
1870  * out instead.
1871  *
1872  * If the thread that set the single_threading request has set the
1873  * P_SINGLE_EXIT bit in the process flags then this call will never return
1874  * if 'return_instead' is false, but will exit.
1875  *
1876  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1877  *---------------+--------------------+---------------------
1878  *       0       | returns 0          |   returns 0 or 1
1879  *               | when ST ends       |   immediatly
1880  *---------------+--------------------+---------------------
1881  *       1       | thread exits       |   returns 1
1882  *               |                    |  immediatly
1883  * 0 = thread_exit() or suspension ok,
1884  * other = return error instead of stopping the thread.
1885  *
1886  * While a full suspension is under effect, even a single threading
1887  * thread would be suspended if it made this call (but it shouldn't).
1888  * This call should only be made from places where
1889  * thread_exit() would be safe as that may be the outcome unless
1890  * return_instead is set.
1891  */
1892 int
1893 thread_suspend_check(int return_instead)
1894 {
1895 	struct thread *td;
1896 	struct proc *p;
1897 	struct ksegrp *kg;
1898 
1899 	td = curthread;
1900 	p = td->td_proc;
1901 	kg = td->td_ksegrp;
1902 	PROC_LOCK_ASSERT(p, MA_OWNED);
1903 	while (P_SHOULDSTOP(p)) {
1904 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1905 			KASSERT(p->p_singlethread != NULL,
1906 			    ("singlethread not set"));
1907 			/*
1908 			 * The only suspension in action is a
1909 			 * single-threading. Single threader need not stop.
1910 			 * XXX Should be safe to access unlocked
1911 			 * as it can only be set to be true by us.
1912 			 */
1913 			if (p->p_singlethread == td)
1914 				return (0);	/* Exempt from stopping. */
1915 		}
1916 		if (return_instead)
1917 			return (1);
1918 
1919 		/*
1920 		 * If the process is waiting for us to exit,
1921 		 * this thread should just suicide.
1922 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1923 		 */
1924 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1925 			mtx_lock_spin(&sched_lock);
1926 			while (mtx_owned(&Giant))
1927 				mtx_unlock(&Giant);
1928 			thread_exit();
1929 		}
1930 
1931 		/*
1932 		 * When a thread suspends, it just
1933 		 * moves to the processes's suspend queue
1934 		 * and stays there.
1935 		 */
1936 		mtx_lock_spin(&sched_lock);
1937 		if ((p->p_flag & P_STOPPED_SIG) &&
1938 		    (p->p_suspcount+1 == p->p_numthreads)) {
1939 			mtx_unlock_spin(&sched_lock);
1940 			PROC_LOCK(p->p_pptr);
1941 			if ((p->p_pptr->p_procsig->ps_flag &
1942 				PS_NOCLDSTOP) == 0) {
1943 				psignal(p->p_pptr, SIGCHLD);
1944 			}
1945 			PROC_UNLOCK(p->p_pptr);
1946 			mtx_lock_spin(&sched_lock);
1947 		}
1948 		mtx_assert(&Giant, MA_NOTOWNED);
1949 		thread_suspend_one(td);
1950 		PROC_UNLOCK(p);
1951 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1952 			if (p->p_numthreads == p->p_suspcount) {
1953 				thread_unsuspend_one(p->p_singlethread);
1954 			}
1955 		}
1956 		p->p_stats->p_ru.ru_nivcsw++;
1957 		mi_switch();
1958 		mtx_unlock_spin(&sched_lock);
1959 		PROC_LOCK(p);
1960 	}
1961 	return (0);
1962 }
1963 
1964 void
1965 thread_suspend_one(struct thread *td)
1966 {
1967 	struct proc *p = td->td_proc;
1968 
1969 	mtx_assert(&sched_lock, MA_OWNED);
1970 	p->p_suspcount++;
1971 	TD_SET_SUSPENDED(td);
1972 	TAILQ_INSERT_TAIL(&p->p_suspended, td, td_runq);
1973 	/*
1974 	 * Hack: If we are suspending but are on the sleep queue
1975 	 * then we are in msleep or the cv equivalent. We
1976 	 * want to look like we have two Inhibitors.
1977 	 * May already be set.. doesn't matter.
1978 	 */
1979 	if (TD_ON_SLEEPQ(td))
1980 		TD_SET_SLEEPING(td);
1981 }
1982 
1983 void
1984 thread_unsuspend_one(struct thread *td)
1985 {
1986 	struct proc *p = td->td_proc;
1987 
1988 	mtx_assert(&sched_lock, MA_OWNED);
1989 	TAILQ_REMOVE(&p->p_suspended, td, td_runq);
1990 	TD_CLR_SUSPENDED(td);
1991 	p->p_suspcount--;
1992 	setrunnable(td);
1993 }
1994 
1995 /*
1996  * Allow all threads blocked by single threading to continue running.
1997  */
1998 void
1999 thread_unsuspend(struct proc *p)
2000 {
2001 	struct thread *td;
2002 
2003 	mtx_assert(&sched_lock, MA_OWNED);
2004 	PROC_LOCK_ASSERT(p, MA_OWNED);
2005 	if (!P_SHOULDSTOP(p)) {
2006 		while (( td = TAILQ_FIRST(&p->p_suspended))) {
2007 			thread_unsuspend_one(td);
2008 		}
2009 	} else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
2010 	    (p->p_numthreads == p->p_suspcount)) {
2011 		/*
2012 		 * Stopping everything also did the job for the single
2013 		 * threading request. Now we've downgraded to single-threaded,
2014 		 * let it continue.
2015 		 */
2016 		thread_unsuspend_one(p->p_singlethread);
2017 	}
2018 }
2019 
2020 void
2021 thread_single_end(void)
2022 {
2023 	struct thread *td;
2024 	struct proc *p;
2025 
2026 	td = curthread;
2027 	p = td->td_proc;
2028 	PROC_LOCK_ASSERT(p, MA_OWNED);
2029 	p->p_flag &= ~P_STOPPED_SINGLE;
2030 	p->p_singlethread = NULL;
2031 	/*
2032 	 * If there are other threads they mey now run,
2033 	 * unless of course there is a blanket 'stop order'
2034 	 * on the process. The single threader must be allowed
2035 	 * to continue however as this is a bad place to stop.
2036 	 */
2037 	if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
2038 		mtx_lock_spin(&sched_lock);
2039 		while (( td = TAILQ_FIRST(&p->p_suspended))) {
2040 			thread_unsuspend_one(td);
2041 		}
2042 		mtx_unlock_spin(&sched_lock);
2043 	}
2044 }
2045 
2046 
2047