xref: /freebsd/sys/kern/kern_thread.c (revision d056fa046c6a91b90cd98165face0e42a33a5173)
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  * KSEGRP related storage.
55  */
56 static uma_zone_t ksegrp_zone;
57 static uma_zone_t thread_zone;
58 
59 /* DEBUG ONLY */
60 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
61 static int thread_debug = 0;
62 SYSCTL_INT(_kern_threads, OID_AUTO, debug, CTLFLAG_RW,
63 	&thread_debug, 0, "thread debug");
64 
65 int max_threads_per_proc = 1500;
66 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
67 	&max_threads_per_proc, 0, "Limit on threads per proc");
68 
69 int max_groups_per_proc = 1500;
70 SYSCTL_INT(_kern_threads, OID_AUTO, max_groups_per_proc, CTLFLAG_RW,
71 	&max_groups_per_proc, 0, "Limit on thread groups per proc");
72 
73 int max_threads_hits;
74 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
75 	&max_threads_hits, 0, "");
76 
77 int virtual_cpu;
78 
79 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
80 TAILQ_HEAD(, ksegrp) zombie_ksegrps = TAILQ_HEAD_INITIALIZER(zombie_ksegrps);
81 struct mtx kse_zombie_lock;
82 MTX_SYSINIT(kse_zombie_lock, &kse_zombie_lock, "kse zombie lock", MTX_SPIN);
83 
84 static int
85 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
86 {
87 	int error, new_val;
88 	int def_val;
89 
90 	def_val = mp_ncpus;
91 	if (virtual_cpu == 0)
92 		new_val = def_val;
93 	else
94 		new_val = virtual_cpu;
95 	error = sysctl_handle_int(oidp, &new_val, 0, req);
96 	if (error != 0 || req->newptr == NULL)
97 		return (error);
98 	if (new_val < 0)
99 		return (EINVAL);
100 	virtual_cpu = new_val;
101 	return (0);
102 }
103 
104 /* DEBUG ONLY */
105 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
106 	0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
107 	"debug virtual cpus");
108 
109 struct mtx tid_lock;
110 static struct unrhdr *tid_unrhdr;
111 
112 /*
113  * Prepare a thread for use.
114  */
115 static int
116 thread_ctor(void *mem, int size, void *arg, int flags)
117 {
118 	struct thread	*td;
119 
120 	td = (struct thread *)mem;
121 	td->td_state = TDS_INACTIVE;
122 	td->td_oncpu = NOCPU;
123 
124 	td->td_tid = alloc_unr(tid_unrhdr);
125 
126 	/*
127 	 * Note that td_critnest begins life as 1 because the thread is not
128 	 * running and is thereby implicitly waiting to be on the receiving
129 	 * end of a context switch.  A context switch must occur inside a
130 	 * critical section, and in fact, includes hand-off of the sched_lock.
131 	 * After a context switch to a newly created thread, it will release
132 	 * sched_lock for the first time, and its td_critnest will hit 0 for
133 	 * the first time.  This happens on the far end of a context switch,
134 	 * and when it context switches away from itself, it will in fact go
135 	 * back into a critical section, and hand off the sched lock to the
136 	 * next thread.
137 	 */
138 	td->td_critnest = 1;
139 
140 #ifdef AUDIT
141 	audit_thread_alloc(td);
142 #endif
143 	return (0);
144 }
145 
146 /*
147  * Reclaim a thread after use.
148  */
149 static void
150 thread_dtor(void *mem, int size, void *arg)
151 {
152 	struct thread *td;
153 
154 	td = (struct thread *)mem;
155 
156 #ifdef INVARIANTS
157 	/* Verify that this thread is in a safe state to free. */
158 	switch (td->td_state) {
159 	case TDS_INHIBITED:
160 	case TDS_RUNNING:
161 	case TDS_CAN_RUN:
162 	case TDS_RUNQ:
163 		/*
164 		 * We must never unlink a thread that is in one of
165 		 * these states, because it is currently active.
166 		 */
167 		panic("bad state for thread unlinking");
168 		/* NOTREACHED */
169 	case TDS_INACTIVE:
170 		break;
171 	default:
172 		panic("bad thread state");
173 		/* NOTREACHED */
174 	}
175 #endif
176 #ifdef AUDIT
177 	audit_thread_free(td);
178 #endif
179 	free_unr(tid_unrhdr, td->td_tid);
180 	sched_newthread(td);
181 }
182 
183 /*
184  * Initialize type-stable parts of a thread (when newly created).
185  */
186 static int
187 thread_init(void *mem, int size, int flags)
188 {
189 	struct thread *td;
190 
191 	td = (struct thread *)mem;
192 
193 	vm_thread_new(td, 0);
194 	cpu_thread_setup(td);
195 	td->td_sleepqueue = sleepq_alloc();
196 	td->td_turnstile = turnstile_alloc();
197 	td->td_umtxq = umtxq_alloc();
198 	td->td_sched = (struct td_sched *)&td[1];
199 	sched_newthread(td);
200 	return (0);
201 }
202 
203 /*
204  * Tear down type-stable parts of a thread (just before being discarded).
205  */
206 static void
207 thread_fini(void *mem, int size)
208 {
209 	struct thread *td;
210 
211 	td = (struct thread *)mem;
212 	turnstile_free(td->td_turnstile);
213 	sleepq_free(td->td_sleepqueue);
214 	umtxq_free(td->td_umtxq);
215 	vm_thread_dispose(td);
216 }
217 
218 /*
219  * Initialize type-stable parts of a ksegrp (when newly created).
220  */
221 static int
222 ksegrp_ctor(void *mem, int size, void *arg, int flags)
223 {
224 	struct ksegrp	*kg;
225 
226 	kg = (struct ksegrp *)mem;
227 	bzero(mem, size);
228 	kg->kg_sched = (struct kg_sched *)&kg[1];
229 	return (0);
230 }
231 
232 void
233 ksegrp_link(struct ksegrp *kg, struct proc *p)
234 {
235 
236 	TAILQ_INIT(&kg->kg_threads);
237 	TAILQ_INIT(&kg->kg_runq);	/* links with td_runq */
238 	TAILQ_INIT(&kg->kg_upcalls);	/* all upcall structure in ksegrp */
239 	kg->kg_proc = p;
240 	/*
241 	 * the following counters are in the -zero- section
242 	 * and may not need clearing
243 	 */
244 	kg->kg_numthreads = 0;
245 	kg->kg_numupcalls = 0;
246 	/* link it in now that it's consistent */
247 	p->p_numksegrps++;
248 	TAILQ_INSERT_HEAD(&p->p_ksegrps, kg, kg_ksegrp);
249 }
250 
251 /*
252  * Called from:
253  *   thread-exit()
254  */
255 void
256 ksegrp_unlink(struct ksegrp *kg)
257 {
258 	struct proc *p;
259 
260 	mtx_assert(&sched_lock, MA_OWNED);
261 	KASSERT((kg->kg_numthreads == 0), ("ksegrp_unlink: residual threads"));
262 	KASSERT((kg->kg_numupcalls == 0), ("ksegrp_unlink: residual upcalls"));
263 
264 	p = kg->kg_proc;
265 	TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
266 	p->p_numksegrps--;
267 	/*
268 	 * Aggregate stats from the KSE
269 	 */
270 	if (p->p_procscopegrp == kg)
271 		p->p_procscopegrp = NULL;
272 }
273 
274 /*
275  * For a newly created process,
276  * link up all the structures and its initial threads etc.
277  * called from:
278  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
279  * proc_dtor() (should go away)
280  * proc_init()
281  */
282 void
283 proc_linkup(struct proc *p, struct ksegrp *kg, struct thread *td)
284 {
285 
286 	TAILQ_INIT(&p->p_ksegrps);	     /* all ksegrps in proc */
287 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
288 	TAILQ_INIT(&p->p_suspended);	     /* Threads suspended */
289 	sigqueue_init(&p->p_sigqueue, p);
290 	p->p_ksi = ksiginfo_alloc(1);
291 	if (p->p_ksi != NULL) {
292 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
293 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
294 	}
295 	LIST_INIT(&p->p_mqnotifier);
296 	p->p_numksegrps = 0;
297 	p->p_numthreads = 0;
298 
299 	ksegrp_link(kg, p);
300 	thread_link(td, kg);
301 }
302 
303 /*
304  * Initialize global thread allocation resources.
305  */
306 void
307 threadinit(void)
308 {
309 
310 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
311 	tid_unrhdr = new_unrhdr(PID_MAX + 1, INT_MAX, &tid_lock);
312 
313 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
314 	    thread_ctor, thread_dtor, thread_init, thread_fini,
315 	    UMA_ALIGN_CACHE, 0);
316 	ksegrp_zone = uma_zcreate("KSEGRP", sched_sizeof_ksegrp(),
317 	    ksegrp_ctor, NULL, NULL, NULL,
318 	    UMA_ALIGN_CACHE, 0);
319 	kseinit();	/* set up kse specific stuff  e.g. upcall zone*/
320 }
321 
322 /*
323  * Stash an embarasingly extra thread into the zombie thread queue.
324  */
325 void
326 thread_stash(struct thread *td)
327 {
328 	mtx_lock_spin(&kse_zombie_lock);
329 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq);
330 	mtx_unlock_spin(&kse_zombie_lock);
331 }
332 
333 /*
334  * Stash an embarasingly extra ksegrp into the zombie ksegrp queue.
335  */
336 void
337 ksegrp_stash(struct ksegrp *kg)
338 {
339 	mtx_lock_spin(&kse_zombie_lock);
340 	TAILQ_INSERT_HEAD(&zombie_ksegrps, kg, kg_ksegrp);
341 	mtx_unlock_spin(&kse_zombie_lock);
342 }
343 
344 /*
345  * Reap zombie kse resource.
346  */
347 void
348 thread_reap(void)
349 {
350 	struct thread *td_first, *td_next;
351 	struct ksegrp *kg_first, * kg_next;
352 
353 	/*
354 	 * Don't even bother to lock if none at this instant,
355 	 * we really don't care about the next instant..
356 	 */
357 	if ((!TAILQ_EMPTY(&zombie_threads))
358 	    || (!TAILQ_EMPTY(&zombie_ksegrps))) {
359 		mtx_lock_spin(&kse_zombie_lock);
360 		td_first = TAILQ_FIRST(&zombie_threads);
361 		kg_first = TAILQ_FIRST(&zombie_ksegrps);
362 		if (td_first)
363 			TAILQ_INIT(&zombie_threads);
364 		if (kg_first)
365 			TAILQ_INIT(&zombie_ksegrps);
366 		mtx_unlock_spin(&kse_zombie_lock);
367 		while (td_first) {
368 			td_next = TAILQ_NEXT(td_first, td_runq);
369 			if (td_first->td_ucred)
370 				crfree(td_first->td_ucred);
371 			thread_free(td_first);
372 			td_first = td_next;
373 		}
374 		while (kg_first) {
375 			kg_next = TAILQ_NEXT(kg_first, kg_ksegrp);
376 			ksegrp_free(kg_first);
377 			kg_first = kg_next;
378 		}
379 		/*
380 		 * there will always be a thread on the list if one of these
381 		 * is there.
382 		 */
383 		kse_GC();
384 	}
385 }
386 
387 /*
388  * Allocate a ksegrp.
389  */
390 struct ksegrp *
391 ksegrp_alloc(void)
392 {
393 	return (uma_zalloc(ksegrp_zone, M_WAITOK));
394 }
395 
396 /*
397  * Allocate a thread.
398  */
399 struct thread *
400 thread_alloc(void)
401 {
402 	thread_reap(); /* check if any zombies to get */
403 	return (uma_zalloc(thread_zone, M_WAITOK));
404 }
405 
406 /*
407  * Deallocate a ksegrp.
408  */
409 void
410 ksegrp_free(struct ksegrp *td)
411 {
412 	uma_zfree(ksegrp_zone, td);
413 }
414 
415 /*
416  * Deallocate a thread.
417  */
418 void
419 thread_free(struct thread *td)
420 {
421 
422 	cpu_thread_clean(td);
423 	uma_zfree(thread_zone, td);
424 }
425 
426 /*
427  * Discard the current thread and exit from its context.
428  * Always called with scheduler locked.
429  *
430  * Because we can't free a thread while we're operating under its context,
431  * push the current thread into our CPU's deadthread holder. This means
432  * we needn't worry about someone else grabbing our context before we
433  * do a cpu_throw().  This may not be needed now as we are under schedlock.
434  * Maybe we can just do a thread_stash() as thr_exit1 does.
435  */
436 /*  XXX
437  * libthr expects its thread exit to return for the last
438  * thread, meaning that the program is back to non-threaded
439  * mode I guess. Because we do this (cpu_throw) unconditionally
440  * here, they have their own version of it. (thr_exit1())
441  * that doesn't do it all if this was the last thread.
442  * It is also called from thread_suspend_check().
443  * Of course in the end, they end up coming here through exit1
444  * anyhow..  After fixing 'thr' to play by the rules we should be able
445  * to merge these two functions together.
446  *
447  * called from:
448  * exit1()
449  * kse_exit()
450  * thr_exit()
451  * thread_user_enter()
452  * thread_userret()
453  * thread_suspend_check()
454  */
455 void
456 thread_exit(void)
457 {
458 	uint64_t new_switchtime;
459 	struct thread *td;
460 	struct proc *p;
461 	struct ksegrp	*kg;
462 
463 	td = curthread;
464 	kg = td->td_ksegrp;
465 	p = td->td_proc;
466 
467 	mtx_assert(&sched_lock, MA_OWNED);
468 	mtx_assert(&Giant, MA_NOTOWNED);
469 	PROC_LOCK_ASSERT(p, MA_OWNED);
470 	KASSERT(p != NULL, ("thread exiting without a process"));
471 	KASSERT(kg != NULL, ("thread exiting without a kse group"));
472 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
473 	    (long)p->p_pid, p->p_comm);
474 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
475 
476 #ifdef AUDIT
477 	AUDIT_SYSCALL_EXIT(0, td);
478 #endif
479 
480 	if (td->td_standin != NULL) {
481 		/*
482 		 * Note that we don't need to free the cred here as it
483 		 * is done in thread_reap().
484 		 */
485 		thread_stash(td->td_standin);
486 		td->td_standin = NULL;
487 	}
488 
489 	/*
490 	 * drop FPU & debug register state storage, or any other
491 	 * architecture specific resources that
492 	 * would not be on a new untouched process.
493 	 */
494 	cpu_thread_exit(td);	/* XXXSMP */
495 
496 	/*
497 	 * The thread is exiting. scheduler can release its stuff
498 	 * and collect stats etc.
499 	 * XXX this is not very right, since PROC_UNLOCK may still
500 	 * need scheduler stuff.
501 	 */
502 	sched_thread_exit(td);
503 
504 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
505 	new_switchtime = cpu_ticks();
506 	p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
507 	p->p_rux.rux_uticks += td->td_uticks;
508 	p->p_rux.rux_sticks += td->td_sticks;
509 	p->p_rux.rux_iticks += td->td_iticks;
510 	PCPU_SET(switchtime, new_switchtime);
511 	PCPU_SET(switchticks, ticks);
512 	cnt.v_swtch++;
513 
514 	/* Add our usage into the usage of all our children. */
515 	if (p->p_numthreads == 1)
516 		ruadd(p->p_ru, &p->p_rux, &p->p_stats->p_cru, &p->p_crux);
517 
518 	/*
519 	 * The last thread is left attached to the process
520 	 * So that the whole bundle gets recycled. Skip
521 	 * all this stuff if we never had threads.
522 	 * EXIT clears all sign of other threads when
523 	 * it goes to single threading, so the last thread always
524 	 * takes the short path.
525 	 */
526 	if (p->p_flag & P_HADTHREADS) {
527 		if (p->p_numthreads > 1) {
528 			thread_unlink(td);
529 
530 			/* XXX first arg not used in 4BSD or ULE */
531 			sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
532 
533 			/*
534 			 * The test below is NOT true if we are the
535 			 * sole exiting thread. P_STOPPED_SNGL is unset
536 			 * in exit1() after it is the only survivor.
537 			 */
538 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
539 				if (p->p_numthreads == p->p_suspcount) {
540 					thread_unsuspend_one(p->p_singlethread);
541 				}
542 			}
543 
544 			/*
545 			 * Because each upcall structure has an owner thread,
546 			 * owner thread exits only when process is in exiting
547 			 * state, so upcall to userland is no longer needed,
548 			 * deleting upcall structure is safe here.
549 			 * So when all threads in a group is exited, all upcalls
550 			 * in the group should be automatically freed.
551 			 *  XXXKSE This is a KSE thing and should be exported
552 			 * there somehow.
553 			 */
554 			upcall_remove(td);
555 
556 			/*
557 			 * If the thread we unlinked above was the last one,
558 			 * then this ksegrp should go away too.
559 			 */
560 			if (kg->kg_numthreads == 0) {
561 				/*
562 				 * let the scheduler know about this in case
563 				 * it needs to recover stats or resources.
564 				 * Theoretically we could let
565 				 * sched_exit_ksegrp()  do the equivalent of
566 				 * setting the concurrency to 0
567 				 * but don't do it yet to avoid changing
568 				 * the existing scheduler code until we
569 				 * are ready.
570 				 * We supply a random other ksegrp
571 				 * as the recipient of any built up
572 				 * cpu usage etc. (If the scheduler wants it).
573 				 * XXXKSE
574 				 * This is probably not fair so think of
575  				 * a better answer.
576 				 */
577 				sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), td);
578 				sched_set_concurrency(kg, 0); /* XXX TEMP */
579 				ksegrp_unlink(kg);
580 				ksegrp_stash(kg);
581 			}
582 			PROC_UNLOCK(p);
583 			td->td_ksegrp	= NULL;
584 			PCPU_SET(deadthread, td);
585 		} else {
586 			/*
587 			 * The last thread is exiting.. but not through exit()
588 			 * what should we do?
589 			 * Theoretically this can't happen
590  			 * exit1() - clears threading flags before coming here
591  			 * kse_exit() - treats last thread specially
592  			 * thr_exit() - treats last thread specially
593  			 * thread_user_enter() - only if more exist
594  			 * thread_userret() - only if more exist
595  			 * thread_suspend_check() - only if more exist
596 			 */
597 			panic ("thread_exit: Last thread exiting on its own");
598 		}
599 	} else {
600 		/*
601 		 * non threaded process comes here.
602 		 * This includes an EX threaded process that is coming
603 		 * here via exit1(). (exit1 dethreads the proc first).
604 		 */
605 		PROC_UNLOCK(p);
606 	}
607 	td->td_state = TDS_INACTIVE;
608 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
609 	cpu_throw(td, choosethread());
610 	panic("I'm a teapot!");
611 	/* NOTREACHED */
612 }
613 
614 /*
615  * Do any thread specific cleanups that may be needed in wait()
616  * called with Giant, proc and schedlock not held.
617  */
618 void
619 thread_wait(struct proc *p)
620 {
621 	struct thread *td;
622 
623 	mtx_assert(&Giant, MA_NOTOWNED);
624 	KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
625 	KASSERT((p->p_numksegrps == 1), ("Multiple ksegrps in wait1()"));
626 	FOREACH_THREAD_IN_PROC(p, td) {
627 		if (td->td_standin != NULL) {
628 			if (td->td_standin->td_ucred != NULL) {
629 				crfree(td->td_standin->td_ucred);
630 				td->td_standin->td_ucred = NULL;
631 			}
632 			thread_free(td->td_standin);
633 			td->td_standin = NULL;
634 		}
635 		cpu_thread_clean(td);
636 		crfree(td->td_ucred);
637 	}
638 	thread_reap();	/* check for zombie threads etc. */
639 }
640 
641 /*
642  * Link a thread to a process.
643  * set up anything that needs to be initialized for it to
644  * be used by the process.
645  *
646  * Note that we do not link to the proc's ucred here.
647  * The thread is linked as if running but no KSE assigned.
648  * Called from:
649  *  proc_linkup()
650  *  thread_schedule_upcall()
651  *  thr_create()
652  */
653 void
654 thread_link(struct thread *td, struct ksegrp *kg)
655 {
656 	struct proc *p;
657 
658 	p = kg->kg_proc;
659 	td->td_state    = TDS_INACTIVE;
660 	td->td_proc     = p;
661 	td->td_ksegrp   = kg;
662 	td->td_flags    = 0;
663 	td->td_kflags	= 0;
664 
665 	LIST_INIT(&td->td_contested);
666 	sigqueue_init(&td->td_sigqueue, p);
667 	callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
668 	TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
669 	TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist);
670 	p->p_numthreads++;
671 	kg->kg_numthreads++;
672 }
673 
674 /*
675  * Convert a process with one thread to an unthreaded process.
676  * Called from:
677  *  thread_single(exit)  (called from execve and exit)
678  *  kse_exit()		XXX may need cleaning up wrt KSE stuff
679  */
680 void
681 thread_unthread(struct thread *td)
682 {
683 	struct proc *p = td->td_proc;
684 
685 	KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
686 	upcall_remove(td);
687 	p->p_flag &= ~(P_SA|P_HADTHREADS);
688 	td->td_mailbox = NULL;
689 	td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
690 	if (td->td_standin != NULL) {
691 		thread_stash(td->td_standin);
692 		td->td_standin = NULL;
693 	}
694 	sched_set_concurrency(td->td_ksegrp, 1);
695 }
696 
697 /*
698  * Called from:
699  *  thread_exit()
700  */
701 void
702 thread_unlink(struct thread *td)
703 {
704 	struct proc *p = td->td_proc;
705 	struct ksegrp *kg = td->td_ksegrp;
706 
707 	mtx_assert(&sched_lock, MA_OWNED);
708 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
709 	p->p_numthreads--;
710 	TAILQ_REMOVE(&kg->kg_threads, td, td_kglist);
711 	kg->kg_numthreads--;
712 	/* could clear a few other things here */
713 	/* Must  NOT clear links to proc and ksegrp! */
714 }
715 
716 /*
717  * Enforce single-threading.
718  *
719  * Returns 1 if the caller must abort (another thread is waiting to
720  * exit the process or similar). Process is locked!
721  * Returns 0 when you are successfully the only thread running.
722  * A process has successfully single threaded in the suspend mode when
723  * There are no threads in user mode. Threads in the kernel must be
724  * allowed to continue until they get to the user boundary. They may even
725  * copy out their return values and data before suspending. They may however be
726  * accelerated in reaching the user boundary as we will wake up
727  * any sleeping threads that are interruptable. (PCATCH).
728  */
729 int
730 thread_single(int mode)
731 {
732 	struct thread *td;
733 	struct thread *td2;
734 	struct proc *p;
735 	int remaining;
736 
737 	td = curthread;
738 	p = td->td_proc;
739 	mtx_assert(&Giant, MA_NOTOWNED);
740 	PROC_LOCK_ASSERT(p, MA_OWNED);
741 	KASSERT((td != NULL), ("curthread is NULL"));
742 
743 	if ((p->p_flag & P_HADTHREADS) == 0)
744 		return (0);
745 
746 	/* Is someone already single threading? */
747 	if (p->p_singlethread != NULL && p->p_singlethread != td)
748 		return (1);
749 
750 	if (mode == SINGLE_EXIT) {
751 		p->p_flag |= P_SINGLE_EXIT;
752 		p->p_flag &= ~P_SINGLE_BOUNDARY;
753 	} else {
754 		p->p_flag &= ~P_SINGLE_EXIT;
755 		if (mode == SINGLE_BOUNDARY)
756 			p->p_flag |= P_SINGLE_BOUNDARY;
757 		else
758 			p->p_flag &= ~P_SINGLE_BOUNDARY;
759 	}
760 	p->p_flag |= P_STOPPED_SINGLE;
761 	mtx_lock_spin(&sched_lock);
762 	p->p_singlethread = td;
763 	if (mode == SINGLE_EXIT)
764 		remaining = p->p_numthreads;
765 	else if (mode == SINGLE_BOUNDARY)
766 		remaining = p->p_numthreads - p->p_boundary_count;
767 	else
768 		remaining = p->p_numthreads - p->p_suspcount;
769 	while (remaining != 1) {
770 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
771 			goto stopme;
772 		FOREACH_THREAD_IN_PROC(p, td2) {
773 			if (td2 == td)
774 				continue;
775 			td2->td_flags |= TDF_ASTPENDING;
776 			if (TD_IS_INHIBITED(td2)) {
777 				switch (mode) {
778 				case SINGLE_EXIT:
779 					if (td->td_flags & TDF_DBSUSPEND)
780 						td->td_flags &= ~TDF_DBSUSPEND;
781 					if (TD_IS_SUSPENDED(td2))
782 						thread_unsuspend_one(td2);
783 					if (TD_ON_SLEEPQ(td2) &&
784 					    (td2->td_flags & TDF_SINTR))
785 						sleepq_abort(td2, EINTR);
786 					break;
787 				case SINGLE_BOUNDARY:
788 					if (TD_IS_SUSPENDED(td2) &&
789 					    !(td2->td_flags & TDF_BOUNDARY))
790 						thread_unsuspend_one(td2);
791 					if (TD_ON_SLEEPQ(td2) &&
792 					    (td2->td_flags & TDF_SINTR))
793 						sleepq_abort(td2, ERESTART);
794 					break;
795 				default:
796 					if (TD_IS_SUSPENDED(td2))
797 						continue;
798 					/*
799 					 * maybe other inhibitted states too?
800 					 */
801 					if ((td2->td_flags & TDF_SINTR) &&
802 					    (td2->td_inhibitors &
803 					    (TDI_SLEEPING | TDI_SWAPPED)))
804 						thread_suspend_one(td2);
805 					break;
806 				}
807 			}
808 #ifdef SMP
809 			else if (TD_IS_RUNNING(td2) && td != td2) {
810 				forward_signal(td2);
811 			}
812 #endif
813 		}
814 		if (mode == SINGLE_EXIT)
815 			remaining = p->p_numthreads;
816 		else if (mode == SINGLE_BOUNDARY)
817 			remaining = p->p_numthreads - p->p_boundary_count;
818 		else
819 			remaining = p->p_numthreads - p->p_suspcount;
820 
821 		/*
822 		 * Maybe we suspended some threads.. was it enough?
823 		 */
824 		if (remaining == 1)
825 			break;
826 
827 stopme:
828 		/*
829 		 * Wake us up when everyone else has suspended.
830 		 * In the mean time we suspend as well.
831 		 */
832 		thread_stopped(p);
833 		thread_suspend_one(td);
834 		PROC_UNLOCK(p);
835 		mi_switch(SW_VOL, NULL);
836 		mtx_unlock_spin(&sched_lock);
837 		PROC_LOCK(p);
838 		mtx_lock_spin(&sched_lock);
839 		if (mode == SINGLE_EXIT)
840 			remaining = p->p_numthreads;
841 		else if (mode == SINGLE_BOUNDARY)
842 			remaining = p->p_numthreads - p->p_boundary_count;
843 		else
844 			remaining = p->p_numthreads - p->p_suspcount;
845 	}
846 	if (mode == SINGLE_EXIT) {
847 		/*
848 		 * We have gotten rid of all the other threads and we
849 		 * are about to either exit or exec. In either case,
850 		 * we try our utmost  to revert to being a non-threaded
851 		 * process.
852 		 */
853 		p->p_singlethread = NULL;
854 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
855 		thread_unthread(td);
856 	}
857 	mtx_unlock_spin(&sched_lock);
858 	return (0);
859 }
860 
861 /*
862  * Called in from locations that can safely check to see
863  * whether we have to suspend or at least throttle for a
864  * single-thread event (e.g. fork).
865  *
866  * Such locations include userret().
867  * If the "return_instead" argument is non zero, the thread must be able to
868  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
869  *
870  * The 'return_instead' argument tells the function if it may do a
871  * thread_exit() or suspend, or whether the caller must abort and back
872  * out instead.
873  *
874  * If the thread that set the single_threading request has set the
875  * P_SINGLE_EXIT bit in the process flags then this call will never return
876  * if 'return_instead' is false, but will exit.
877  *
878  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
879  *---------------+--------------------+---------------------
880  *       0       | returns 0          |   returns 0 or 1
881  *               | when ST ends       |   immediatly
882  *---------------+--------------------+---------------------
883  *       1       | thread exits       |   returns 1
884  *               |                    |  immediatly
885  * 0 = thread_exit() or suspension ok,
886  * other = return error instead of stopping the thread.
887  *
888  * While a full suspension is under effect, even a single threading
889  * thread would be suspended if it made this call (but it shouldn't).
890  * This call should only be made from places where
891  * thread_exit() would be safe as that may be the outcome unless
892  * return_instead is set.
893  */
894 int
895 thread_suspend_check(int return_instead)
896 {
897 	struct thread *td;
898 	struct proc *p;
899 
900 	td = curthread;
901 	p = td->td_proc;
902 	mtx_assert(&Giant, MA_NOTOWNED);
903 	PROC_LOCK_ASSERT(p, MA_OWNED);
904 	while (P_SHOULDSTOP(p) ||
905 	      ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
906 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
907 			KASSERT(p->p_singlethread != NULL,
908 			    ("singlethread not set"));
909 			/*
910 			 * The only suspension in action is a
911 			 * single-threading. Single threader need not stop.
912 			 * XXX Should be safe to access unlocked
913 			 * as it can only be set to be true by us.
914 			 */
915 			if (p->p_singlethread == td)
916 				return (0);	/* Exempt from stopping. */
917 		}
918 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
919 			return (EINTR);
920 
921 		/* Should we goto user boundary if we didn't come from there? */
922 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
923 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
924 			return (ERESTART);
925 
926 		/* If thread will exit, flush its pending signals */
927 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
928 			sigqueue_flush(&td->td_sigqueue);
929 
930 		mtx_lock_spin(&sched_lock);
931 		thread_stopped(p);
932 		/*
933 		 * If the process is waiting for us to exit,
934 		 * this thread should just suicide.
935 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
936 		 */
937 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
938 			thread_exit();
939 
940 		/*
941 		 * When a thread suspends, it just
942 		 * moves to the processes's suspend queue
943 		 * and stays there.
944 		 */
945 		thread_suspend_one(td);
946 		if (return_instead == 0) {
947 			p->p_boundary_count++;
948 			td->td_flags |= TDF_BOUNDARY;
949 		}
950 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
951 			if (p->p_numthreads == p->p_suspcount)
952 				thread_unsuspend_one(p->p_singlethread);
953 		}
954 		PROC_UNLOCK(p);
955 		mi_switch(SW_INVOL, NULL);
956 		if (return_instead == 0) {
957 			p->p_boundary_count--;
958 			td->td_flags &= ~TDF_BOUNDARY;
959 		}
960 		mtx_unlock_spin(&sched_lock);
961 		PROC_LOCK(p);
962 	}
963 	return (0);
964 }
965 
966 void
967 thread_suspend_one(struct thread *td)
968 {
969 	struct proc *p = td->td_proc;
970 
971 	mtx_assert(&sched_lock, MA_OWNED);
972 	PROC_LOCK_ASSERT(p, MA_OWNED);
973 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
974 	p->p_suspcount++;
975 	TD_SET_SUSPENDED(td);
976 	TAILQ_INSERT_TAIL(&p->p_suspended, td, td_runq);
977 }
978 
979 void
980 thread_unsuspend_one(struct thread *td)
981 {
982 	struct proc *p = td->td_proc;
983 
984 	mtx_assert(&sched_lock, MA_OWNED);
985 	PROC_LOCK_ASSERT(p, MA_OWNED);
986 	TAILQ_REMOVE(&p->p_suspended, td, td_runq);
987 	TD_CLR_SUSPENDED(td);
988 	p->p_suspcount--;
989 	setrunnable(td);
990 }
991 
992 /*
993  * Allow all threads blocked by single threading to continue running.
994  */
995 void
996 thread_unsuspend(struct proc *p)
997 {
998 	struct thread *td;
999 
1000 	mtx_assert(&sched_lock, MA_OWNED);
1001 	PROC_LOCK_ASSERT(p, MA_OWNED);
1002 	if (!P_SHOULDSTOP(p)) {
1003 		while ((td = TAILQ_FIRST(&p->p_suspended))) {
1004 			thread_unsuspend_one(td);
1005 		}
1006 	} else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
1007 	    (p->p_numthreads == p->p_suspcount)) {
1008 		/*
1009 		 * Stopping everything also did the job for the single
1010 		 * threading request. Now we've downgraded to single-threaded,
1011 		 * let it continue.
1012 		 */
1013 		thread_unsuspend_one(p->p_singlethread);
1014 	}
1015 }
1016 
1017 /*
1018  * End the single threading mode..
1019  */
1020 void
1021 thread_single_end(void)
1022 {
1023 	struct thread *td;
1024 	struct proc *p;
1025 
1026 	td = curthread;
1027 	p = td->td_proc;
1028 	PROC_LOCK_ASSERT(p, MA_OWNED);
1029 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
1030 	mtx_lock_spin(&sched_lock);
1031 	p->p_singlethread = NULL;
1032 	p->p_procscopegrp = NULL;
1033 	/*
1034 	 * If there are other threads they mey now run,
1035 	 * unless of course there is a blanket 'stop order'
1036 	 * on the process. The single threader must be allowed
1037 	 * to continue however as this is a bad place to stop.
1038 	 */
1039 	if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
1040 		while ((td = TAILQ_FIRST(&p->p_suspended))) {
1041 			thread_unsuspend_one(td);
1042 		}
1043 	}
1044 	mtx_unlock_spin(&sched_lock);
1045 }
1046 
1047 struct thread *
1048 thread_find(struct proc *p, lwpid_t tid)
1049 {
1050 	struct thread *td;
1051 
1052 	PROC_LOCK_ASSERT(p, MA_OWNED);
1053 	mtx_lock_spin(&sched_lock);
1054 	FOREACH_THREAD_IN_PROC(p, td) {
1055 		if (td->td_tid == tid)
1056 			break;
1057 	}
1058 	mtx_unlock_spin(&sched_lock);
1059 	return (td);
1060 }
1061