xref: /freebsd/sys/kern/kern_thread.c (revision 28bd1fdc0be889d5b7d7f36f014709ccd0a112a4)
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/smp.h>
39 #include <sys/sysctl.h>
40 #include <sys/sched.h>
41 #include <sys/sleepqueue.h>
42 #include <sys/turnstile.h>
43 #include <sys/ktr.h>
44 #include <sys/umtx.h>
45 
46 #include <vm/vm.h>
47 #include <vm/vm_extern.h>
48 #include <vm/uma.h>
49 
50 /*
51  * KSEGRP related storage.
52  */
53 static uma_zone_t ksegrp_zone;
54 static uma_zone_t thread_zone;
55 
56 /* DEBUG ONLY */
57 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
58 static int thread_debug = 0;
59 SYSCTL_INT(_kern_threads, OID_AUTO, debug, CTLFLAG_RW,
60 	&thread_debug, 0, "thread debug");
61 
62 int max_threads_per_proc = 1500;
63 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
64 	&max_threads_per_proc, 0, "Limit on threads per proc");
65 
66 int max_groups_per_proc = 1500;
67 SYSCTL_INT(_kern_threads, OID_AUTO, max_groups_per_proc, CTLFLAG_RW,
68 	&max_groups_per_proc, 0, "Limit on thread groups per proc");
69 
70 int max_threads_hits;
71 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
72 	&max_threads_hits, 0, "");
73 
74 int virtual_cpu;
75 
76 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
77 TAILQ_HEAD(, ksegrp) zombie_ksegrps = TAILQ_HEAD_INITIALIZER(zombie_ksegrps);
78 struct mtx kse_zombie_lock;
79 MTX_SYSINIT(kse_zombie_lock, &kse_zombie_lock, "kse zombie lock", MTX_SPIN);
80 
81 static int
82 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
83 {
84 	int error, new_val;
85 	int def_val;
86 
87 	def_val = mp_ncpus;
88 	if (virtual_cpu == 0)
89 		new_val = def_val;
90 	else
91 		new_val = virtual_cpu;
92 	error = sysctl_handle_int(oidp, &new_val, 0, req);
93 	if (error != 0 || req->newptr == NULL)
94 		return (error);
95 	if (new_val < 0)
96 		return (EINVAL);
97 	virtual_cpu = new_val;
98 	return (0);
99 }
100 
101 /* DEBUG ONLY */
102 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
103 	0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
104 	"debug virtual cpus");
105 
106 struct mtx tid_lock;
107 static struct unrhdr *tid_unrhdr;
108 
109 /*
110  * Prepare a thread for use.
111  */
112 static int
113 thread_ctor(void *mem, int size, void *arg, int flags)
114 {
115 	struct thread	*td;
116 
117 	td = (struct thread *)mem;
118 	td->td_state = TDS_INACTIVE;
119 	td->td_oncpu = NOCPU;
120 
121 	td->td_tid = alloc_unr(tid_unrhdr);
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.  A context switch must occur inside a
127 	 * critical section, and in fact, includes hand-off of the sched_lock.
128 	 * After a context switch to a newly created thread, it will release
129 	 * sched_lock for the first time, and its td_critnest will hit 0 for
130 	 * the first time.  This happens on the far end of a context switch,
131 	 * and when it context switches away from itself, it will in fact go
132 	 * back into a critical section, and hand off the sched lock to the
133 	 * next thread.
134 	 */
135 	td->td_critnest = 1;
136 	return (0);
137 }
138 
139 /*
140  * Reclaim a thread after use.
141  */
142 static void
143 thread_dtor(void *mem, int size, void *arg)
144 {
145 	struct thread *td;
146 
147 	td = (struct thread *)mem;
148 
149 #ifdef INVARIANTS
150 	/* Verify that this thread is in a safe state to free. */
151 	switch (td->td_state) {
152 	case TDS_INHIBITED:
153 	case TDS_RUNNING:
154 	case TDS_CAN_RUN:
155 	case TDS_RUNQ:
156 		/*
157 		 * We must never unlink a thread that is in one of
158 		 * these states, because it is currently active.
159 		 */
160 		panic("bad state for thread unlinking");
161 		/* NOTREACHED */
162 	case TDS_INACTIVE:
163 		break;
164 	default:
165 		panic("bad thread state");
166 		/* NOTREACHED */
167 	}
168 #endif
169 
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_umtxq = umtxq_alloc();
189 	td->td_sched = (struct td_sched *)&td[1];
190 	sched_newthread(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 	umtxq_free(td->td_umtxq);
206 	vm_thread_dispose(td);
207 }
208 
209 /*
210  * Initialize type-stable parts of a ksegrp (when newly created).
211  */
212 static int
213 ksegrp_ctor(void *mem, int size, void *arg, int flags)
214 {
215 	struct ksegrp	*kg;
216 
217 	kg = (struct ksegrp *)mem;
218 	bzero(mem, size);
219 	kg->kg_sched = (struct kg_sched *)&kg[1];
220 	return (0);
221 }
222 
223 void
224 ksegrp_link(struct ksegrp *kg, struct proc *p)
225 {
226 
227 	TAILQ_INIT(&kg->kg_threads);
228 	TAILQ_INIT(&kg->kg_runq);	/* links with td_runq */
229 	TAILQ_INIT(&kg->kg_upcalls);	/* all upcall structure in ksegrp */
230 	kg->kg_proc = p;
231 	/*
232 	 * the following counters are in the -zero- section
233 	 * and may not need clearing
234 	 */
235 	kg->kg_numthreads = 0;
236 	kg->kg_numupcalls = 0;
237 	/* link it in now that it's consistent */
238 	p->p_numksegrps++;
239 	TAILQ_INSERT_HEAD(&p->p_ksegrps, kg, kg_ksegrp);
240 }
241 
242 /*
243  * Called from:
244  *   thread-exit()
245  */
246 void
247 ksegrp_unlink(struct ksegrp *kg)
248 {
249 	struct proc *p;
250 
251 	mtx_assert(&sched_lock, MA_OWNED);
252 	KASSERT((kg->kg_numthreads == 0), ("ksegrp_unlink: residual threads"));
253 	KASSERT((kg->kg_numupcalls == 0), ("ksegrp_unlink: residual upcalls"));
254 
255 	p = kg->kg_proc;
256 	TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
257 	p->p_numksegrps--;
258 	/*
259 	 * Aggregate stats from the KSE
260 	 */
261 	if (p->p_procscopegrp == kg)
262 		p->p_procscopegrp = NULL;
263 }
264 
265 /*
266  * For a newly created process,
267  * link up all the structures and its initial threads etc.
268  * called from:
269  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
270  * proc_dtor() (should go away)
271  * proc_init()
272  */
273 void
274 proc_linkup(struct proc *p, struct ksegrp *kg, struct thread *td)
275 {
276 
277 	TAILQ_INIT(&p->p_ksegrps);	     /* all ksegrps in proc */
278 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
279 	TAILQ_INIT(&p->p_suspended);	     /* Threads suspended */
280 	p->p_numksegrps = 0;
281 	p->p_numthreads = 0;
282 
283 	ksegrp_link(kg, p);
284 	thread_link(td, kg);
285 }
286 
287 /*
288  * Initialize global thread allocation resources.
289  */
290 void
291 threadinit(void)
292 {
293 
294 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
295 	tid_unrhdr = new_unrhdr(PID_MAX + 1, INT_MAX, &tid_lock);
296 
297 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
298 	    thread_ctor, thread_dtor, thread_init, thread_fini,
299 	    UMA_ALIGN_CACHE, 0);
300 	ksegrp_zone = uma_zcreate("KSEGRP", sched_sizeof_ksegrp(),
301 	    ksegrp_ctor, NULL, NULL, NULL,
302 	    UMA_ALIGN_CACHE, 0);
303 	kseinit();	/* set up kse specific stuff  e.g. upcall zone*/
304 }
305 
306 /*
307  * Stash an embarasingly extra thread into the zombie thread queue.
308  */
309 void
310 thread_stash(struct thread *td)
311 {
312 	mtx_lock_spin(&kse_zombie_lock);
313 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq);
314 	mtx_unlock_spin(&kse_zombie_lock);
315 }
316 
317 /*
318  * Stash an embarasingly extra ksegrp into the zombie ksegrp queue.
319  */
320 void
321 ksegrp_stash(struct ksegrp *kg)
322 {
323 	mtx_lock_spin(&kse_zombie_lock);
324 	TAILQ_INSERT_HEAD(&zombie_ksegrps, kg, kg_ksegrp);
325 	mtx_unlock_spin(&kse_zombie_lock);
326 }
327 
328 /*
329  * Reap zombie kse resource.
330  */
331 void
332 thread_reap(void)
333 {
334 	struct thread *td_first, *td_next;
335 	struct ksegrp *kg_first, * kg_next;
336 
337 	/*
338 	 * Don't even bother to lock if none at this instant,
339 	 * we really don't care about the next instant..
340 	 */
341 	if ((!TAILQ_EMPTY(&zombie_threads))
342 	    || (!TAILQ_EMPTY(&zombie_ksegrps))) {
343 		mtx_lock_spin(&kse_zombie_lock);
344 		td_first = TAILQ_FIRST(&zombie_threads);
345 		kg_first = TAILQ_FIRST(&zombie_ksegrps);
346 		if (td_first)
347 			TAILQ_INIT(&zombie_threads);
348 		if (kg_first)
349 			TAILQ_INIT(&zombie_ksegrps);
350 		mtx_unlock_spin(&kse_zombie_lock);
351 		while (td_first) {
352 			td_next = TAILQ_NEXT(td_first, td_runq);
353 			if (td_first->td_ucred)
354 				crfree(td_first->td_ucred);
355 			thread_free(td_first);
356 			td_first = td_next;
357 		}
358 		while (kg_first) {
359 			kg_next = TAILQ_NEXT(kg_first, kg_ksegrp);
360 			ksegrp_free(kg_first);
361 			kg_first = kg_next;
362 		}
363 		/*
364 		 * there will always be a thread on the list if one of these
365 		 * is there.
366 		 */
367 		kse_GC();
368 	}
369 }
370 
371 /*
372  * Allocate a ksegrp.
373  */
374 struct ksegrp *
375 ksegrp_alloc(void)
376 {
377 	return (uma_zalloc(ksegrp_zone, M_WAITOK));
378 }
379 
380 /*
381  * Allocate a thread.
382  */
383 struct thread *
384 thread_alloc(void)
385 {
386 	thread_reap(); /* check if any zombies to get */
387 	return (uma_zalloc(thread_zone, M_WAITOK));
388 }
389 
390 /*
391  * Deallocate a ksegrp.
392  */
393 void
394 ksegrp_free(struct ksegrp *td)
395 {
396 	uma_zfree(ksegrp_zone, td);
397 }
398 
399 /*
400  * Deallocate a thread.
401  */
402 void
403 thread_free(struct thread *td)
404 {
405 
406 	cpu_thread_clean(td);
407 	uma_zfree(thread_zone, td);
408 }
409 
410 /*
411  * Discard the current thread and exit from its context.
412  * Always called with scheduler locked.
413  *
414  * Because we can't free a thread while we're operating under its context,
415  * push the current thread into our CPU's deadthread holder. This means
416  * we needn't worry about someone else grabbing our context before we
417  * do a cpu_throw().  This may not be needed now as we are under schedlock.
418  * Maybe we can just do a thread_stash() as thr_exit1 does.
419  */
420 /*  XXX
421  * libthr expects its thread exit to return for the last
422  * thread, meaning that the program is back to non-threaded
423  * mode I guess. Because we do this (cpu_throw) unconditionally
424  * here, they have their own version of it. (thr_exit1())
425  * that doesn't do it all if this was the last thread.
426  * It is also called from thread_suspend_check().
427  * Of course in the end, they end up coming here through exit1
428  * anyhow..  After fixing 'thr' to play by the rules we should be able
429  * to merge these two functions together.
430  *
431  * called from:
432  * exit1()
433  * kse_exit()
434  * thr_exit()
435  * thread_user_enter()
436  * thread_userret()
437  * thread_suspend_check()
438  */
439 void
440 thread_exit(void)
441 {
442 	struct thread *td;
443 	struct proc *p;
444 	struct ksegrp	*kg;
445 
446 	td = curthread;
447 	kg = td->td_ksegrp;
448 	p = td->td_proc;
449 
450 	mtx_assert(&sched_lock, MA_OWNED);
451 	mtx_assert(&Giant, MA_NOTOWNED);
452 	PROC_LOCK_ASSERT(p, MA_OWNED);
453 	KASSERT(p != NULL, ("thread exiting without a process"));
454 	KASSERT(kg != NULL, ("thread exiting without a kse group"));
455 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
456 	    (long)p->p_pid, p->p_comm);
457 
458 	if (td->td_standin != NULL) {
459 		/*
460 		 * Note that we don't need to free the cred here as it
461 		 * is done in thread_reap().
462 		 */
463 		thread_stash(td->td_standin);
464 		td->td_standin = NULL;
465 	}
466 
467 	/*
468 	 * drop FPU & debug register state storage, or any other
469 	 * architecture specific resources that
470 	 * would not be on a new untouched process.
471 	 */
472 	cpu_thread_exit(td);	/* XXXSMP */
473 
474 	/*
475 	 * The thread is exiting. scheduler can release its stuff
476 	 * and collect stats etc.
477 	 */
478 	sched_thread_exit(td);
479 
480 	/*
481 	 * The last thread is left attached to the process
482 	 * So that the whole bundle gets recycled. Skip
483 	 * all this stuff if we never had threads.
484 	 * EXIT clears all sign of other threads when
485 	 * it goes to single threading, so the last thread always
486 	 * takes the short path.
487 	 */
488 	if (p->p_flag & P_HADTHREADS) {
489 		if (p->p_numthreads > 1) {
490 			thread_unlink(td);
491 
492 			/* XXX first arg not used in 4BSD or ULE */
493 			sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
494 
495 			/*
496 			 * The test below is NOT true if we are the
497 			 * sole exiting thread. P_STOPPED_SNGL is unset
498 			 * in exit1() after it is the only survivor.
499 			 */
500 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
501 				if (p->p_numthreads == p->p_suspcount) {
502 					thread_unsuspend_one(p->p_singlethread);
503 				}
504 			}
505 
506 			/*
507 			 * Because each upcall structure has an owner thread,
508 			 * owner thread exits only when process is in exiting
509 			 * state, so upcall to userland is no longer needed,
510 			 * deleting upcall structure is safe here.
511 			 * So when all threads in a group is exited, all upcalls
512 			 * in the group should be automatically freed.
513 			 *  XXXKSE This is a KSE thing and should be exported
514 			 * there somehow.
515 			 */
516 			upcall_remove(td);
517 
518 			/*
519 			 * If the thread we unlinked above was the last one,
520 			 * then this ksegrp should go away too.
521 			 */
522 			if (kg->kg_numthreads == 0) {
523 				/*
524 				 * let the scheduler know about this in case
525 				 * it needs to recover stats or resources.
526 				 * Theoretically we could let
527 				 * sched_exit_ksegrp()  do the equivalent of
528 				 * setting the concurrency to 0
529 				 * but don't do it yet to avoid changing
530 				 * the existing scheduler code until we
531 				 * are ready.
532 				 * We supply a random other ksegrp
533 				 * as the recipient of any built up
534 				 * cpu usage etc. (If the scheduler wants it).
535 				 * XXXKSE
536 				 * This is probably not fair so think of
537  				 * a better answer.
538 				 */
539 				sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), td);
540 				sched_set_concurrency(kg, 0); /* XXX TEMP */
541 				ksegrp_unlink(kg);
542 				ksegrp_stash(kg);
543 			}
544 			PROC_UNLOCK(p);
545 			td->td_ksegrp	= NULL;
546 			PCPU_SET(deadthread, td);
547 		} else {
548 			/*
549 			 * The last thread is exiting.. but not through exit()
550 			 * what should we do?
551 			 * Theoretically this can't happen
552  			 * exit1() - clears threading flags before coming here
553  			 * kse_exit() - treats last thread specially
554  			 * thr_exit() - treats last thread specially
555  			 * thread_user_enter() - only if more exist
556  			 * thread_userret() - only if more exist
557  			 * thread_suspend_check() - only if more exist
558 			 */
559 			panic ("thread_exit: Last thread exiting on its own");
560 		}
561 	} else {
562 		/*
563 		 * non threaded process comes here.
564 		 * This includes an EX threaded process that is coming
565 		 * here via exit1(). (exit1 dethreads the proc first).
566 		 */
567 		PROC_UNLOCK(p);
568 	}
569 	td->td_state = TDS_INACTIVE;
570 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
571 	cpu_throw(td, choosethread());
572 	panic("I'm a teapot!");
573 	/* NOTREACHED */
574 }
575 
576 /*
577  * Do any thread specific cleanups that may be needed in wait()
578  * called with Giant, proc and schedlock not held.
579  */
580 void
581 thread_wait(struct proc *p)
582 {
583 	struct thread *td;
584 
585 	mtx_assert(&Giant, MA_NOTOWNED);
586 	KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
587 	KASSERT((p->p_numksegrps == 1), ("Multiple ksegrps in wait1()"));
588 	FOREACH_THREAD_IN_PROC(p, td) {
589 		if (td->td_standin != NULL) {
590 			if (td->td_standin->td_ucred != NULL) {
591 				crfree(td->td_standin->td_ucred);
592 				td->td_standin->td_ucred = NULL;
593 			}
594 			thread_free(td->td_standin);
595 			td->td_standin = NULL;
596 		}
597 		cpu_thread_clean(td);
598 		crfree(td->td_ucred);
599 	}
600 	thread_reap();	/* check for zombie threads etc. */
601 }
602 
603 /*
604  * Link a thread to a process.
605  * set up anything that needs to be initialized for it to
606  * be used by the process.
607  *
608  * Note that we do not link to the proc's ucred here.
609  * The thread is linked as if running but no KSE assigned.
610  * Called from:
611  *  proc_linkup()
612  *  thread_schedule_upcall()
613  *  thr_create()
614  */
615 void
616 thread_link(struct thread *td, struct ksegrp *kg)
617 {
618 	struct proc *p;
619 
620 	p = kg->kg_proc;
621 	td->td_state    = TDS_INACTIVE;
622 	td->td_proc     = p;
623 	td->td_ksegrp   = kg;
624 	td->td_flags    = 0;
625 	td->td_kflags	= 0;
626 
627 	LIST_INIT(&td->td_contested);
628 	callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
629 	TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
630 	TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist);
631 	p->p_numthreads++;
632 	kg->kg_numthreads++;
633 }
634 
635 /*
636  * Convert a process with one thread to an unthreaded process.
637  * Called from:
638  *  thread_single(exit)  (called from execve and exit)
639  *  kse_exit()		XXX may need cleaning up wrt KSE stuff
640  */
641 void
642 thread_unthread(struct thread *td)
643 {
644 	struct proc *p = td->td_proc;
645 
646 	KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
647 	upcall_remove(td);
648 	p->p_flag &= ~(P_SA|P_HADTHREADS);
649 	td->td_mailbox = NULL;
650 	td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
651 	if (td->td_standin != NULL) {
652 		thread_stash(td->td_standin);
653 		td->td_standin = NULL;
654 	}
655 	sched_set_concurrency(td->td_ksegrp, 1);
656 }
657 
658 /*
659  * Called from:
660  *  thread_exit()
661  */
662 void
663 thread_unlink(struct thread *td)
664 {
665 	struct proc *p = td->td_proc;
666 	struct ksegrp *kg = td->td_ksegrp;
667 
668 	mtx_assert(&sched_lock, MA_OWNED);
669 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
670 	p->p_numthreads--;
671 	TAILQ_REMOVE(&kg->kg_threads, td, td_kglist);
672 	kg->kg_numthreads--;
673 	/* could clear a few other things here */
674 	/* Must  NOT clear links to proc and ksegrp! */
675 }
676 
677 /*
678  * Enforce single-threading.
679  *
680  * Returns 1 if the caller must abort (another thread is waiting to
681  * exit the process or similar). Process is locked!
682  * Returns 0 when you are successfully the only thread running.
683  * A process has successfully single threaded in the suspend mode when
684  * There are no threads in user mode. Threads in the kernel must be
685  * allowed to continue until they get to the user boundary. They may even
686  * copy out their return values and data before suspending. They may however be
687  * accellerated in reaching the user boundary as we will wake up
688  * any sleeping threads that are interruptable. (PCATCH).
689  */
690 int
691 thread_single(int mode)
692 {
693 	struct thread *td;
694 	struct thread *td2;
695 	struct proc *p;
696 	int remaining;
697 
698 	td = curthread;
699 	p = td->td_proc;
700 	mtx_assert(&Giant, MA_NOTOWNED);
701 	PROC_LOCK_ASSERT(p, MA_OWNED);
702 	KASSERT((td != NULL), ("curthread is NULL"));
703 
704 	if ((p->p_flag & P_HADTHREADS) == 0)
705 		return (0);
706 
707 	/* Is someone already single threading? */
708 	if (p->p_singlethread != NULL && p->p_singlethread != td)
709 		return (1);
710 
711 	if (mode == SINGLE_EXIT) {
712 		p->p_flag |= P_SINGLE_EXIT;
713 		p->p_flag &= ~P_SINGLE_BOUNDARY;
714 	} else {
715 		p->p_flag &= ~P_SINGLE_EXIT;
716 		if (mode == SINGLE_BOUNDARY)
717 			p->p_flag |= P_SINGLE_BOUNDARY;
718 		else
719 			p->p_flag &= ~P_SINGLE_BOUNDARY;
720 	}
721 	p->p_flag |= P_STOPPED_SINGLE;
722 	mtx_lock_spin(&sched_lock);
723 	p->p_singlethread = td;
724 	if (mode == SINGLE_EXIT)
725 		remaining = p->p_numthreads;
726 	else if (mode == SINGLE_BOUNDARY)
727 		remaining = p->p_numthreads - p->p_boundary_count;
728 	else
729 		remaining = p->p_numthreads - p->p_suspcount;
730 	while (remaining != 1) {
731 		FOREACH_THREAD_IN_PROC(p, td2) {
732 			if (td2 == td)
733 				continue;
734 			td2->td_flags |= TDF_ASTPENDING;
735 			if (TD_IS_INHIBITED(td2)) {
736 				switch (mode) {
737 				case SINGLE_EXIT:
738 					if (td->td_flags & TDF_DBSUSPEND)
739 						td->td_flags &= ~TDF_DBSUSPEND;
740 					if (TD_IS_SUSPENDED(td2))
741 						thread_unsuspend_one(td2);
742 					if (TD_ON_SLEEPQ(td2) &&
743 					    (td2->td_flags & TDF_SINTR))
744 						sleepq_abort(td2);
745 					break;
746 				case SINGLE_BOUNDARY:
747 					if (TD_IS_SUSPENDED(td2) &&
748 					    !(td2->td_flags & TDF_BOUNDARY))
749 						thread_unsuspend_one(td2);
750 					if (TD_ON_SLEEPQ(td2) &&
751 					    (td2->td_flags & TDF_SINTR))
752 						sleepq_abort(td2);
753 					break;
754 				default:
755 					if (TD_IS_SUSPENDED(td2))
756 						continue;
757 					/*
758 					 * maybe other inhibitted states too?
759 					 */
760 					if ((td2->td_flags & TDF_SINTR) &&
761 					    (td2->td_inhibitors &
762 					    (TDI_SLEEPING | TDI_SWAPPED)))
763 						thread_suspend_one(td2);
764 					break;
765 				}
766 			}
767 		}
768 		if (mode == SINGLE_EXIT)
769 			remaining = p->p_numthreads;
770 		else if (mode == SINGLE_BOUNDARY)
771 			remaining = p->p_numthreads - p->p_boundary_count;
772 		else
773 			remaining = p->p_numthreads - p->p_suspcount;
774 
775 		/*
776 		 * Maybe we suspended some threads.. was it enough?
777 		 */
778 		if (remaining == 1)
779 			break;
780 
781 		/*
782 		 * Wake us up when everyone else has suspended.
783 		 * In the mean time we suspend as well.
784 		 */
785 		thread_suspend_one(td);
786 		PROC_UNLOCK(p);
787 		mi_switch(SW_VOL, NULL);
788 		mtx_unlock_spin(&sched_lock);
789 		PROC_LOCK(p);
790 		mtx_lock_spin(&sched_lock);
791 		if (mode == SINGLE_EXIT)
792 			remaining = p->p_numthreads;
793 		else if (mode == SINGLE_BOUNDARY)
794 			remaining = p->p_numthreads - p->p_boundary_count;
795 		else
796 			remaining = p->p_numthreads - p->p_suspcount;
797 	}
798 	if (mode == SINGLE_EXIT) {
799 		/*
800 		 * We have gotten rid of all the other threads and we
801 		 * are about to either exit or exec. In either case,
802 		 * we try our utmost  to revert to being a non-threaded
803 		 * process.
804 		 */
805 		p->p_singlethread = NULL;
806 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
807 		thread_unthread(td);
808 	}
809 	mtx_unlock_spin(&sched_lock);
810 	return (0);
811 }
812 
813 /*
814  * Called in from locations that can safely check to see
815  * whether we have to suspend or at least throttle for a
816  * single-thread event (e.g. fork).
817  *
818  * Such locations include userret().
819  * If the "return_instead" argument is non zero, the thread must be able to
820  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
821  *
822  * The 'return_instead' argument tells the function if it may do a
823  * thread_exit() or suspend, or whether the caller must abort and back
824  * out instead.
825  *
826  * If the thread that set the single_threading request has set the
827  * P_SINGLE_EXIT bit in the process flags then this call will never return
828  * if 'return_instead' is false, but will exit.
829  *
830  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
831  *---------------+--------------------+---------------------
832  *       0       | returns 0          |   returns 0 or 1
833  *               | when ST ends       |   immediatly
834  *---------------+--------------------+---------------------
835  *       1       | thread exits       |   returns 1
836  *               |                    |  immediatly
837  * 0 = thread_exit() or suspension ok,
838  * other = return error instead of stopping the thread.
839  *
840  * While a full suspension is under effect, even a single threading
841  * thread would be suspended if it made this call (but it shouldn't).
842  * This call should only be made from places where
843  * thread_exit() would be safe as that may be the outcome unless
844  * return_instead is set.
845  */
846 int
847 thread_suspend_check(int return_instead)
848 {
849 	struct thread *td;
850 	struct proc *p;
851 
852 	td = curthread;
853 	p = td->td_proc;
854 	mtx_assert(&Giant, MA_NOTOWNED);
855 	PROC_LOCK_ASSERT(p, MA_OWNED);
856 	while (P_SHOULDSTOP(p) ||
857 	      ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
858 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
859 			KASSERT(p->p_singlethread != NULL,
860 			    ("singlethread not set"));
861 			/*
862 			 * The only suspension in action is a
863 			 * single-threading. Single threader need not stop.
864 			 * XXX Should be safe to access unlocked
865 			 * as it can only be set to be true by us.
866 			 */
867 			if (p->p_singlethread == td)
868 				return (0);	/* Exempt from stopping. */
869 		}
870 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
871 			return (1);
872 
873 		/* Should we goto user boundary if we didn't come from there? */
874 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
875 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
876 			return (1);
877 
878 		mtx_lock_spin(&sched_lock);
879 		thread_stopped(p);
880 		/*
881 		 * If the process is waiting for us to exit,
882 		 * this thread should just suicide.
883 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
884 		 */
885 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
886 			thread_exit();
887 
888 		/*
889 		 * When a thread suspends, it just
890 		 * moves to the processes's suspend queue
891 		 * and stays there.
892 		 */
893 		thread_suspend_one(td);
894 		if (return_instead == 0) {
895 			p->p_boundary_count++;
896 			td->td_flags |= TDF_BOUNDARY;
897 		}
898 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
899 			if (p->p_numthreads == p->p_suspcount)
900 				thread_unsuspend_one(p->p_singlethread);
901 		}
902 		PROC_UNLOCK(p);
903 		mi_switch(SW_INVOL, NULL);
904 		if (return_instead == 0) {
905 			p->p_boundary_count--;
906 			td->td_flags &= ~TDF_BOUNDARY;
907 		}
908 		mtx_unlock_spin(&sched_lock);
909 		PROC_LOCK(p);
910 	}
911 	return (0);
912 }
913 
914 void
915 thread_suspend_one(struct thread *td)
916 {
917 	struct proc *p = td->td_proc;
918 
919 	mtx_assert(&sched_lock, MA_OWNED);
920 	PROC_LOCK_ASSERT(p, MA_OWNED);
921 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
922 	p->p_suspcount++;
923 	TD_SET_SUSPENDED(td);
924 	TAILQ_INSERT_TAIL(&p->p_suspended, td, td_runq);
925 }
926 
927 void
928 thread_unsuspend_one(struct thread *td)
929 {
930 	struct proc *p = td->td_proc;
931 
932 	mtx_assert(&sched_lock, MA_OWNED);
933 	PROC_LOCK_ASSERT(p, MA_OWNED);
934 	TAILQ_REMOVE(&p->p_suspended, td, td_runq);
935 	TD_CLR_SUSPENDED(td);
936 	p->p_suspcount--;
937 	setrunnable(td);
938 }
939 
940 /*
941  * Allow all threads blocked by single threading to continue running.
942  */
943 void
944 thread_unsuspend(struct proc *p)
945 {
946 	struct thread *td;
947 
948 	mtx_assert(&sched_lock, MA_OWNED);
949 	PROC_LOCK_ASSERT(p, MA_OWNED);
950 	if (!P_SHOULDSTOP(p)) {
951 		while ((td = TAILQ_FIRST(&p->p_suspended))) {
952 			thread_unsuspend_one(td);
953 		}
954 	} else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
955 	    (p->p_numthreads == p->p_suspcount)) {
956 		/*
957 		 * Stopping everything also did the job for the single
958 		 * threading request. Now we've downgraded to single-threaded,
959 		 * let it continue.
960 		 */
961 		thread_unsuspend_one(p->p_singlethread);
962 	}
963 }
964 
965 /*
966  * End the single threading mode..
967  */
968 void
969 thread_single_end(void)
970 {
971 	struct thread *td;
972 	struct proc *p;
973 
974 	td = curthread;
975 	p = td->td_proc;
976 	PROC_LOCK_ASSERT(p, MA_OWNED);
977 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
978 	mtx_lock_spin(&sched_lock);
979 	p->p_singlethread = NULL;
980 	p->p_procscopegrp = NULL;
981 	/*
982 	 * If there are other threads they mey now run,
983 	 * unless of course there is a blanket 'stop order'
984 	 * on the process. The single threader must be allowed
985 	 * to continue however as this is a bad place to stop.
986 	 */
987 	if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
988 		while ((td = TAILQ_FIRST(&p->p_suspended))) {
989 			thread_unsuspend_one(td);
990 		}
991 	}
992 	mtx_unlock_spin(&sched_lock);
993 }
994 
995 /*
996  * Called before going into an interruptible sleep to see if we have been
997  * interrupted or requested to exit.
998  */
999 int
1000 thread_sleep_check(struct thread *td)
1001 {
1002 	struct proc *p;
1003 
1004 	p = td->td_proc;
1005 	mtx_assert(&sched_lock, MA_OWNED);
1006 	if (p->p_flag & P_HADTHREADS) {
1007 		if (p->p_singlethread != td) {
1008 			if (p->p_flag & P_SINGLE_EXIT)
1009 				return (EINTR);
1010 			if (p->p_flag & P_SINGLE_BOUNDARY)
1011 				return (ERESTART);
1012 		}
1013 		if (td->td_flags & TDF_INTERRUPT)
1014 			return (td->td_intrval);
1015 	}
1016 	return (0);
1017 }
1018