xref: /freebsd/sys/kern/kern_thread.c (revision 72dec0792a09bb5f03d341642657bd6115d99c9e)
1 /*-
2  * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
3  *  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice(s), this list of conditions and the following disclaimer as
10  *    the first lines of this file unmodified other than the possible
11  *    addition of one or more copyright notices.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice(s), this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
26  * DAMAGE.
27  */
28 
29 #include "opt_witness.h"
30 #include "opt_hwpmc_hooks.h"
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/rangelock.h>
42 #include <sys/resourcevar.h>
43 #include <sys/sdt.h>
44 #include <sys/smp.h>
45 #include <sys/sched.h>
46 #include <sys/sleepqueue.h>
47 #include <sys/selinfo.h>
48 #include <sys/syscallsubr.h>
49 #include <sys/sysent.h>
50 #include <sys/turnstile.h>
51 #include <sys/ktr.h>
52 #include <sys/rwlock.h>
53 #include <sys/umtx.h>
54 #include <sys/cpuset.h>
55 #ifdef	HWPMC_HOOKS
56 #include <sys/pmckern.h>
57 #endif
58 
59 #include <security/audit/audit.h>
60 
61 #include <vm/vm.h>
62 #include <vm/vm_extern.h>
63 #include <vm/uma.h>
64 #include <vm/vm_domain.h>
65 #include <sys/eventhandler.h>
66 
67 SDT_PROVIDER_DECLARE(proc);
68 SDT_PROBE_DEFINE(proc, , , lwp__exit);
69 
70 /*
71  * thread related storage.
72  */
73 static uma_zone_t thread_zone;
74 
75 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
76 static struct mtx zombie_lock;
77 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
78 
79 static void thread_zombie(struct thread *);
80 static int thread_unsuspend_one(struct thread *td, struct proc *p,
81     bool boundary);
82 
83 #define TID_BUFFER_SIZE	1024
84 
85 struct mtx tid_lock;
86 static struct unrhdr *tid_unrhdr;
87 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
88 static int tid_head, tid_tail;
89 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
90 
91 struct	tidhashhead *tidhashtbl;
92 u_long	tidhash;
93 struct	rwlock tidhash_lock;
94 
95 static lwpid_t
96 tid_alloc(void)
97 {
98 	lwpid_t	tid;
99 
100 	tid = alloc_unr(tid_unrhdr);
101 	if (tid != -1)
102 		return (tid);
103 	mtx_lock(&tid_lock);
104 	if (tid_head == tid_tail) {
105 		mtx_unlock(&tid_lock);
106 		return (-1);
107 	}
108 	tid = tid_buffer[tid_head];
109 	tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
110 	mtx_unlock(&tid_lock);
111 	return (tid);
112 }
113 
114 static void
115 tid_free(lwpid_t tid)
116 {
117 	lwpid_t tmp_tid = -1;
118 
119 	mtx_lock(&tid_lock);
120 	if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
121 		tmp_tid = tid_buffer[tid_head];
122 		tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
123 	}
124 	tid_buffer[tid_tail] = tid;
125 	tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE;
126 	mtx_unlock(&tid_lock);
127 	if (tmp_tid != -1)
128 		free_unr(tid_unrhdr, tmp_tid);
129 }
130 
131 /*
132  * Prepare a thread for use.
133  */
134 static int
135 thread_ctor(void *mem, int size, void *arg, int flags)
136 {
137 	struct thread	*td;
138 
139 	td = (struct thread *)mem;
140 	td->td_state = TDS_INACTIVE;
141 	td->td_oncpu = NOCPU;
142 
143 	td->td_tid = tid_alloc();
144 
145 	/*
146 	 * Note that td_critnest begins life as 1 because the thread is not
147 	 * running and is thereby implicitly waiting to be on the receiving
148 	 * end of a context switch.
149 	 */
150 	td->td_critnest = 1;
151 	td->td_lend_user_pri = PRI_MAX;
152 	EVENTHANDLER_INVOKE(thread_ctor, td);
153 #ifdef AUDIT
154 	audit_thread_alloc(td);
155 #endif
156 	umtx_thread_alloc(td);
157 	return (0);
158 }
159 
160 /*
161  * Reclaim a thread after use.
162  */
163 static void
164 thread_dtor(void *mem, int size, void *arg)
165 {
166 	struct thread *td;
167 
168 	td = (struct thread *)mem;
169 
170 #ifdef INVARIANTS
171 	/* Verify that this thread is in a safe state to free. */
172 	switch (td->td_state) {
173 	case TDS_INHIBITED:
174 	case TDS_RUNNING:
175 	case TDS_CAN_RUN:
176 	case TDS_RUNQ:
177 		/*
178 		 * We must never unlink a thread that is in one of
179 		 * these states, because it is currently active.
180 		 */
181 		panic("bad state for thread unlinking");
182 		/* NOTREACHED */
183 	case TDS_INACTIVE:
184 		break;
185 	default:
186 		panic("bad thread state");
187 		/* NOTREACHED */
188 	}
189 #endif
190 #ifdef AUDIT
191 	audit_thread_free(td);
192 #endif
193 	/* Free all OSD associated to this thread. */
194 	osd_thread_exit(td);
195 	td_softdep_cleanup(td);
196 	MPASS(td->td_su == NULL);
197 
198 	EVENTHANDLER_INVOKE(thread_dtor, td);
199 	tid_free(td->td_tid);
200 }
201 
202 /*
203  * Initialize type-stable parts of a thread (when newly created).
204  */
205 static int
206 thread_init(void *mem, int size, int flags)
207 {
208 	struct thread *td;
209 
210 	td = (struct thread *)mem;
211 
212 	td->td_sleepqueue = sleepq_alloc();
213 	td->td_turnstile = turnstile_alloc();
214 	td->td_rlqe = NULL;
215 	EVENTHANDLER_INVOKE(thread_init, td);
216 	umtx_thread_init(td);
217 	td->td_kstack = 0;
218 	td->td_sel = NULL;
219 	return (0);
220 }
221 
222 /*
223  * Tear down type-stable parts of a thread (just before being discarded).
224  */
225 static void
226 thread_fini(void *mem, int size)
227 {
228 	struct thread *td;
229 
230 	td = (struct thread *)mem;
231 	EVENTHANDLER_INVOKE(thread_fini, td);
232 	rlqentry_free(td->td_rlqe);
233 	turnstile_free(td->td_turnstile);
234 	sleepq_free(td->td_sleepqueue);
235 	umtx_thread_fini(td);
236 	seltdfini(td);
237 }
238 
239 /*
240  * For a newly created process,
241  * link up all the structures and its initial threads etc.
242  * called from:
243  * {arch}/{arch}/machdep.c   {arch}_init(), init386() etc.
244  * proc_dtor() (should go away)
245  * proc_init()
246  */
247 void
248 proc_linkup0(struct proc *p, struct thread *td)
249 {
250 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
251 	proc_linkup(p, td);
252 }
253 
254 void
255 proc_linkup(struct proc *p, struct thread *td)
256 {
257 
258 	sigqueue_init(&p->p_sigqueue, p);
259 	p->p_ksi = ksiginfo_alloc(1);
260 	if (p->p_ksi != NULL) {
261 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
262 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
263 	}
264 	LIST_INIT(&p->p_mqnotifier);
265 	p->p_numthreads = 0;
266 	thread_link(td, p);
267 }
268 
269 /*
270  * Initialize global thread allocation resources.
271  */
272 void
273 threadinit(void)
274 {
275 
276 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
277 
278 	/*
279 	 * pid_max cannot be greater than PID_MAX.
280 	 * leave one number for thread0.
281 	 */
282 	tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
283 
284 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
285 	    thread_ctor, thread_dtor, thread_init, thread_fini,
286 	    32 - 1, UMA_ZONE_NOFREE);
287 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
288 	rw_init(&tidhash_lock, "tidhash");
289 }
290 
291 /*
292  * Place an unused thread on the zombie list.
293  * Use the slpq as that must be unused by now.
294  */
295 void
296 thread_zombie(struct thread *td)
297 {
298 	mtx_lock_spin(&zombie_lock);
299 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
300 	mtx_unlock_spin(&zombie_lock);
301 }
302 
303 /*
304  * Release a thread that has exited after cpu_throw().
305  */
306 void
307 thread_stash(struct thread *td)
308 {
309 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
310 	thread_zombie(td);
311 }
312 
313 /*
314  * Reap zombie resources.
315  */
316 void
317 thread_reap(void)
318 {
319 	struct thread *td_first, *td_next;
320 
321 	/*
322 	 * Don't even bother to lock if none at this instant,
323 	 * we really don't care about the next instant.
324 	 */
325 	if (!TAILQ_EMPTY(&zombie_threads)) {
326 		mtx_lock_spin(&zombie_lock);
327 		td_first = TAILQ_FIRST(&zombie_threads);
328 		if (td_first)
329 			TAILQ_INIT(&zombie_threads);
330 		mtx_unlock_spin(&zombie_lock);
331 		while (td_first) {
332 			td_next = TAILQ_NEXT(td_first, td_slpq);
333 			thread_cow_free(td_first);
334 			thread_free(td_first);
335 			td_first = td_next;
336 		}
337 	}
338 }
339 
340 /*
341  * Allocate a thread.
342  */
343 struct thread *
344 thread_alloc(int pages)
345 {
346 	struct thread *td;
347 
348 	thread_reap(); /* check if any zombies to get */
349 
350 	td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
351 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
352 	if (!vm_thread_new(td, pages)) {
353 		uma_zfree(thread_zone, td);
354 		return (NULL);
355 	}
356 	cpu_thread_alloc(td);
357 	vm_domain_policy_init(&td->td_vm_dom_policy);
358 	return (td);
359 }
360 
361 int
362 thread_alloc_stack(struct thread *td, int pages)
363 {
364 
365 	KASSERT(td->td_kstack == 0,
366 	    ("thread_alloc_stack called on a thread with kstack"));
367 	if (!vm_thread_new(td, pages))
368 		return (0);
369 	cpu_thread_alloc(td);
370 	return (1);
371 }
372 
373 /*
374  * Deallocate a thread.
375  */
376 void
377 thread_free(struct thread *td)
378 {
379 
380 	lock_profile_thread_exit(td);
381 	if (td->td_cpuset)
382 		cpuset_rel(td->td_cpuset);
383 	td->td_cpuset = NULL;
384 	cpu_thread_free(td);
385 	if (td->td_kstack != 0)
386 		vm_thread_dispose(td);
387 	vm_domain_policy_cleanup(&td->td_vm_dom_policy);
388 	callout_drain(&td->td_slpcallout);
389 	uma_zfree(thread_zone, td);
390 }
391 
392 void
393 thread_cow_get_proc(struct thread *newtd, struct proc *p)
394 {
395 
396 	PROC_LOCK_ASSERT(p, MA_OWNED);
397 	newtd->td_ucred = crhold(p->p_ucred);
398 	newtd->td_limit = lim_hold(p->p_limit);
399 	newtd->td_cowgen = p->p_cowgen;
400 }
401 
402 void
403 thread_cow_get(struct thread *newtd, struct thread *td)
404 {
405 
406 	newtd->td_ucred = crhold(td->td_ucred);
407 	newtd->td_limit = lim_hold(td->td_limit);
408 	newtd->td_cowgen = td->td_cowgen;
409 }
410 
411 void
412 thread_cow_free(struct thread *td)
413 {
414 
415 	if (td->td_ucred != NULL)
416 		crfree(td->td_ucred);
417 	if (td->td_limit != NULL)
418 		lim_free(td->td_limit);
419 }
420 
421 void
422 thread_cow_update(struct thread *td)
423 {
424 	struct proc *p;
425 	struct ucred *oldcred;
426 	struct plimit *oldlimit;
427 
428 	p = td->td_proc;
429 	oldcred = NULL;
430 	oldlimit = NULL;
431 	PROC_LOCK(p);
432 	if (td->td_ucred != p->p_ucred) {
433 		oldcred = td->td_ucred;
434 		td->td_ucred = crhold(p->p_ucred);
435 	}
436 	if (td->td_limit != p->p_limit) {
437 		oldlimit = td->td_limit;
438 		td->td_limit = lim_hold(p->p_limit);
439 	}
440 	td->td_cowgen = p->p_cowgen;
441 	PROC_UNLOCK(p);
442 	if (oldcred != NULL)
443 		crfree(oldcred);
444 	if (oldlimit != NULL)
445 		lim_free(oldlimit);
446 }
447 
448 /*
449  * Discard the current thread and exit from its context.
450  * Always called with scheduler locked.
451  *
452  * Because we can't free a thread while we're operating under its context,
453  * push the current thread into our CPU's deadthread holder. This means
454  * we needn't worry about someone else grabbing our context before we
455  * do a cpu_throw().
456  */
457 void
458 thread_exit(void)
459 {
460 	uint64_t runtime, new_switchtime;
461 	struct thread *td;
462 	struct thread *td2;
463 	struct proc *p;
464 	int wakeup_swapper;
465 
466 	td = curthread;
467 	p = td->td_proc;
468 
469 	PROC_SLOCK_ASSERT(p, MA_OWNED);
470 	mtx_assert(&Giant, MA_NOTOWNED);
471 
472 	PROC_LOCK_ASSERT(p, MA_OWNED);
473 	KASSERT(p != NULL, ("thread exiting without a process"));
474 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
475 	    (long)p->p_pid, td->td_name);
476 	SDT_PROBE0(proc, , , lwp__exit);
477 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
478 
479 #ifdef AUDIT
480 	AUDIT_SYSCALL_EXIT(0, td);
481 #endif
482 	/*
483 	 * drop FPU & debug register state storage, or any other
484 	 * architecture specific resources that
485 	 * would not be on a new untouched process.
486 	 */
487 	cpu_thread_exit(td);
488 
489 	/*
490 	 * The last thread is left attached to the process
491 	 * So that the whole bundle gets recycled. Skip
492 	 * all this stuff if we never had threads.
493 	 * EXIT clears all sign of other threads when
494 	 * it goes to single threading, so the last thread always
495 	 * takes the short path.
496 	 */
497 	if (p->p_flag & P_HADTHREADS) {
498 		if (p->p_numthreads > 1) {
499 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
500 			thread_unlink(td);
501 			td2 = FIRST_THREAD_IN_PROC(p);
502 			sched_exit_thread(td2, td);
503 
504 			/*
505 			 * The test below is NOT true if we are the
506 			 * sole exiting thread. P_STOPPED_SINGLE is unset
507 			 * in exit1() after it is the only survivor.
508 			 */
509 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
510 				if (p->p_numthreads == p->p_suspcount) {
511 					thread_lock(p->p_singlethread);
512 					wakeup_swapper = thread_unsuspend_one(
513 						p->p_singlethread, p, false);
514 					thread_unlock(p->p_singlethread);
515 					if (wakeup_swapper)
516 						kick_proc0();
517 				}
518 			}
519 
520 			PCPU_SET(deadthread, td);
521 		} else {
522 			/*
523 			 * The last thread is exiting.. but not through exit()
524 			 */
525 			panic ("thread_exit: Last thread exiting on its own");
526 		}
527 	}
528 #ifdef	HWPMC_HOOKS
529 	/*
530 	 * If this thread is part of a process that is being tracked by hwpmc(4),
531 	 * inform the module of the thread's impending exit.
532 	 */
533 	if (PMC_PROC_IS_USING_PMCS(td->td_proc))
534 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
535 #endif
536 	PROC_UNLOCK(p);
537 	PROC_STATLOCK(p);
538 	thread_lock(td);
539 	PROC_SUNLOCK(p);
540 
541 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
542 	new_switchtime = cpu_ticks();
543 	runtime = new_switchtime - PCPU_GET(switchtime);
544 	td->td_runtime += runtime;
545 	td->td_incruntime += runtime;
546 	PCPU_SET(switchtime, new_switchtime);
547 	PCPU_SET(switchticks, ticks);
548 	PCPU_INC(cnt.v_swtch);
549 
550 	/* Save our resource usage in our process. */
551 	td->td_ru.ru_nvcsw++;
552 	ruxagg(p, td);
553 	rucollect(&p->p_ru, &td->td_ru);
554 	PROC_STATUNLOCK(p);
555 
556 	td->td_state = TDS_INACTIVE;
557 #ifdef WITNESS
558 	witness_thread_exit(td);
559 #endif
560 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
561 	sched_throw(td);
562 	panic("I'm a teapot!");
563 	/* NOTREACHED */
564 }
565 
566 /*
567  * Do any thread specific cleanups that may be needed in wait()
568  * called with Giant, proc and schedlock not held.
569  */
570 void
571 thread_wait(struct proc *p)
572 {
573 	struct thread *td;
574 
575 	mtx_assert(&Giant, MA_NOTOWNED);
576 	KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
577 	KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
578 	td = FIRST_THREAD_IN_PROC(p);
579 	/* Lock the last thread so we spin until it exits cpu_throw(). */
580 	thread_lock(td);
581 	thread_unlock(td);
582 	lock_profile_thread_exit(td);
583 	cpuset_rel(td->td_cpuset);
584 	td->td_cpuset = NULL;
585 	cpu_thread_clean(td);
586 	thread_cow_free(td);
587 	callout_drain(&td->td_slpcallout);
588 	thread_reap();	/* check for zombie threads etc. */
589 }
590 
591 /*
592  * Link a thread to a process.
593  * set up anything that needs to be initialized for it to
594  * be used by the process.
595  */
596 void
597 thread_link(struct thread *td, struct proc *p)
598 {
599 
600 	/*
601 	 * XXX This can't be enabled because it's called for proc0 before
602 	 * its lock has been created.
603 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
604 	 */
605 	td->td_state    = TDS_INACTIVE;
606 	td->td_proc     = p;
607 	td->td_flags    = TDF_INMEM;
608 
609 	LIST_INIT(&td->td_contested);
610 	LIST_INIT(&td->td_lprof[0]);
611 	LIST_INIT(&td->td_lprof[1]);
612 	sigqueue_init(&td->td_sigqueue, p);
613 	callout_init(&td->td_slpcallout, 1);
614 	TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
615 	p->p_numthreads++;
616 }
617 
618 /*
619  * Called from:
620  *  thread_exit()
621  */
622 void
623 thread_unlink(struct thread *td)
624 {
625 	struct proc *p = td->td_proc;
626 
627 	PROC_LOCK_ASSERT(p, MA_OWNED);
628 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
629 	p->p_numthreads--;
630 	/* could clear a few other things here */
631 	/* Must  NOT clear links to proc! */
632 }
633 
634 static int
635 calc_remaining(struct proc *p, int mode)
636 {
637 	int remaining;
638 
639 	PROC_LOCK_ASSERT(p, MA_OWNED);
640 	PROC_SLOCK_ASSERT(p, MA_OWNED);
641 	if (mode == SINGLE_EXIT)
642 		remaining = p->p_numthreads;
643 	else if (mode == SINGLE_BOUNDARY)
644 		remaining = p->p_numthreads - p->p_boundary_count;
645 	else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
646 		remaining = p->p_numthreads - p->p_suspcount;
647 	else
648 		panic("calc_remaining: wrong mode %d", mode);
649 	return (remaining);
650 }
651 
652 static int
653 remain_for_mode(int mode)
654 {
655 
656 	return (mode == SINGLE_ALLPROC ? 0 : 1);
657 }
658 
659 static int
660 weed_inhib(int mode, struct thread *td2, struct proc *p)
661 {
662 	int wakeup_swapper;
663 
664 	PROC_LOCK_ASSERT(p, MA_OWNED);
665 	PROC_SLOCK_ASSERT(p, MA_OWNED);
666 	THREAD_LOCK_ASSERT(td2, MA_OWNED);
667 
668 	wakeup_swapper = 0;
669 	switch (mode) {
670 	case SINGLE_EXIT:
671 		if (TD_IS_SUSPENDED(td2))
672 			wakeup_swapper |= thread_unsuspend_one(td2, p, true);
673 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0)
674 			wakeup_swapper |= sleepq_abort(td2, EINTR);
675 		break;
676 	case SINGLE_BOUNDARY:
677 	case SINGLE_NO_EXIT:
678 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & TDF_BOUNDARY) == 0)
679 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
680 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0)
681 			wakeup_swapper |= sleepq_abort(td2, ERESTART);
682 		break;
683 	case SINGLE_ALLPROC:
684 		/*
685 		 * ALLPROC suspend tries to avoid spurious EINTR for
686 		 * threads sleeping interruptable, by suspending the
687 		 * thread directly, similarly to sig_suspend_threads().
688 		 * Since such sleep is not performed at the user
689 		 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
690 		 * is used to avoid immediate un-suspend.
691 		 */
692 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
693 		    TDF_ALLPROCSUSP)) == 0)
694 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
695 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0) {
696 			if ((td2->td_flags & TDF_SBDRY) == 0) {
697 				thread_suspend_one(td2);
698 				td2->td_flags |= TDF_ALLPROCSUSP;
699 			} else {
700 				wakeup_swapper |= sleepq_abort(td2, ERESTART);
701 			}
702 		}
703 		break;
704 	}
705 	return (wakeup_swapper);
706 }
707 
708 /*
709  * Enforce single-threading.
710  *
711  * Returns 1 if the caller must abort (another thread is waiting to
712  * exit the process or similar). Process is locked!
713  * Returns 0 when you are successfully the only thread running.
714  * A process has successfully single threaded in the suspend mode when
715  * There are no threads in user mode. Threads in the kernel must be
716  * allowed to continue until they get to the user boundary. They may even
717  * copy out their return values and data before suspending. They may however be
718  * accelerated in reaching the user boundary as we will wake up
719  * any sleeping threads that are interruptable. (PCATCH).
720  */
721 int
722 thread_single(struct proc *p, int mode)
723 {
724 	struct thread *td;
725 	struct thread *td2;
726 	int remaining, wakeup_swapper;
727 
728 	td = curthread;
729 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
730 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
731 	    ("invalid mode %d", mode));
732 	/*
733 	 * If allowing non-ALLPROC singlethreading for non-curproc
734 	 * callers, calc_remaining() and remain_for_mode() should be
735 	 * adjusted to also account for td->td_proc != p.  For now
736 	 * this is not implemented because it is not used.
737 	 */
738 	KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
739 	    (mode != SINGLE_ALLPROC && td->td_proc == p),
740 	    ("mode %d proc %p curproc %p", mode, p, td->td_proc));
741 	mtx_assert(&Giant, MA_NOTOWNED);
742 	PROC_LOCK_ASSERT(p, MA_OWNED);
743 
744 	if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
745 		return (0);
746 
747 	/* Is someone already single threading? */
748 	if (p->p_singlethread != NULL && p->p_singlethread != td)
749 		return (1);
750 
751 	if (mode == SINGLE_EXIT) {
752 		p->p_flag |= P_SINGLE_EXIT;
753 		p->p_flag &= ~P_SINGLE_BOUNDARY;
754 	} else {
755 		p->p_flag &= ~P_SINGLE_EXIT;
756 		if (mode == SINGLE_BOUNDARY)
757 			p->p_flag |= P_SINGLE_BOUNDARY;
758 		else
759 			p->p_flag &= ~P_SINGLE_BOUNDARY;
760 	}
761 	if (mode == SINGLE_ALLPROC)
762 		p->p_flag |= P_TOTAL_STOP;
763 	p->p_flag |= P_STOPPED_SINGLE;
764 	PROC_SLOCK(p);
765 	p->p_singlethread = td;
766 	remaining = calc_remaining(p, mode);
767 	while (remaining != remain_for_mode(mode)) {
768 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
769 			goto stopme;
770 		wakeup_swapper = 0;
771 		FOREACH_THREAD_IN_PROC(p, td2) {
772 			if (td2 == td)
773 				continue;
774 			thread_lock(td2);
775 			td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
776 			if (TD_IS_INHIBITED(td2)) {
777 				wakeup_swapper |= weed_inhib(mode, td2, p);
778 #ifdef SMP
779 			} else if (TD_IS_RUNNING(td2) && td != td2) {
780 				forward_signal(td2);
781 #endif
782 			}
783 			thread_unlock(td2);
784 		}
785 		if (wakeup_swapper)
786 			kick_proc0();
787 		remaining = calc_remaining(p, mode);
788 
789 		/*
790 		 * Maybe we suspended some threads.. was it enough?
791 		 */
792 		if (remaining == remain_for_mode(mode))
793 			break;
794 
795 stopme:
796 		/*
797 		 * Wake us up when everyone else has suspended.
798 		 * In the mean time we suspend as well.
799 		 */
800 		thread_suspend_switch(td, p);
801 		remaining = calc_remaining(p, mode);
802 	}
803 	if (mode == SINGLE_EXIT) {
804 		/*
805 		 * Convert the process to an unthreaded process.  The
806 		 * SINGLE_EXIT is called by exit1() or execve(), in
807 		 * both cases other threads must be retired.
808 		 */
809 		KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
810 		p->p_singlethread = NULL;
811 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
812 
813 		/*
814 		 * Wait for any remaining threads to exit cpu_throw().
815 		 */
816 		while (p->p_exitthreads != 0) {
817 			PROC_SUNLOCK(p);
818 			PROC_UNLOCK(p);
819 			sched_relinquish(td);
820 			PROC_LOCK(p);
821 			PROC_SLOCK(p);
822 		}
823 	} else if (mode == SINGLE_BOUNDARY) {
824 		/*
825 		 * Wait until all suspended threads are removed from
826 		 * the processors.  The thread_suspend_check()
827 		 * increments p_boundary_count while it is still
828 		 * running, which makes it possible for the execve()
829 		 * to destroy vmspace while our other threads are
830 		 * still using the address space.
831 		 *
832 		 * We lock the thread, which is only allowed to
833 		 * succeed after context switch code finished using
834 		 * the address space.
835 		 */
836 		FOREACH_THREAD_IN_PROC(p, td2) {
837 			if (td2 == td)
838 				continue;
839 			thread_lock(td2);
840 			KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
841 			    ("td %p not on boundary", td2));
842 			KASSERT(TD_IS_SUSPENDED(td2),
843 			    ("td %p is not suspended", td2));
844 			thread_unlock(td2);
845 		}
846 	}
847 	PROC_SUNLOCK(p);
848 	return (0);
849 }
850 
851 bool
852 thread_suspend_check_needed(void)
853 {
854 	struct proc *p;
855 	struct thread *td;
856 
857 	td = curthread;
858 	p = td->td_proc;
859 	PROC_LOCK_ASSERT(p, MA_OWNED);
860 	return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
861 	    (td->td_dbgflags & TDB_SUSPEND) != 0));
862 }
863 
864 /*
865  * Called in from locations that can safely check to see
866  * whether we have to suspend or at least throttle for a
867  * single-thread event (e.g. fork).
868  *
869  * Such locations include userret().
870  * If the "return_instead" argument is non zero, the thread must be able to
871  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
872  *
873  * The 'return_instead' argument tells the function if it may do a
874  * thread_exit() or suspend, or whether the caller must abort and back
875  * out instead.
876  *
877  * If the thread that set the single_threading request has set the
878  * P_SINGLE_EXIT bit in the process flags then this call will never return
879  * if 'return_instead' is false, but will exit.
880  *
881  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
882  *---------------+--------------------+---------------------
883  *       0       | returns 0          |   returns 0 or 1
884  *               | when ST ends       |   immediately
885  *---------------+--------------------+---------------------
886  *       1       | thread exits       |   returns 1
887  *               |                    |  immediately
888  * 0 = thread_exit() or suspension ok,
889  * other = return error instead of stopping the thread.
890  *
891  * While a full suspension is under effect, even a single threading
892  * thread would be suspended if it made this call (but it shouldn't).
893  * This call should only be made from places where
894  * thread_exit() would be safe as that may be the outcome unless
895  * return_instead is set.
896  */
897 int
898 thread_suspend_check(int return_instead)
899 {
900 	struct thread *td;
901 	struct proc *p;
902 	int wakeup_swapper;
903 
904 	td = curthread;
905 	p = td->td_proc;
906 	mtx_assert(&Giant, MA_NOTOWNED);
907 	PROC_LOCK_ASSERT(p, MA_OWNED);
908 	while (thread_suspend_check_needed()) {
909 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
910 			KASSERT(p->p_singlethread != NULL,
911 			    ("singlethread not set"));
912 			/*
913 			 * The only suspension in action is a
914 			 * single-threading. Single threader need not stop.
915 			 * It is safe to access p->p_singlethread unlocked
916 			 * because it can only be set to our address by us.
917 			 */
918 			if (p->p_singlethread == td)
919 				return (0);	/* Exempt from stopping. */
920 		}
921 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
922 			return (EINTR);
923 
924 		/* Should we goto user boundary if we didn't come from there? */
925 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
926 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
927 			return (ERESTART);
928 
929 		/*
930 		 * Ignore suspend requests if they are deferred.
931 		 */
932 		if ((td->td_flags & TDF_SBDRY) != 0) {
933 			KASSERT(return_instead,
934 			    ("TDF_SBDRY set for unsafe thread_suspend_check"));
935 			KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
936 			    (TDF_SEINTR | TDF_SERESTART),
937 			    ("both TDF_SEINTR and TDF_SERESTART"));
938 			return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
939 		}
940 
941 		/*
942 		 * If the process is waiting for us to exit,
943 		 * this thread should just suicide.
944 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
945 		 */
946 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
947 			PROC_UNLOCK(p);
948 
949 			/*
950 			 * Allow Linux emulation layer to do some work
951 			 * before thread suicide.
952 			 */
953 			if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
954 				(p->p_sysent->sv_thread_detach)(td);
955 			umtx_thread_exit(td);
956 			kern_thr_exit(td);
957 			panic("stopped thread did not exit");
958 		}
959 
960 		PROC_SLOCK(p);
961 		thread_stopped(p);
962 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
963 			if (p->p_numthreads == p->p_suspcount + 1) {
964 				thread_lock(p->p_singlethread);
965 				wakeup_swapper = thread_unsuspend_one(
966 				    p->p_singlethread, p, false);
967 				thread_unlock(p->p_singlethread);
968 				if (wakeup_swapper)
969 					kick_proc0();
970 			}
971 		}
972 		PROC_UNLOCK(p);
973 		thread_lock(td);
974 		/*
975 		 * When a thread suspends, it just
976 		 * gets taken off all queues.
977 		 */
978 		thread_suspend_one(td);
979 		if (return_instead == 0) {
980 			p->p_boundary_count++;
981 			td->td_flags |= TDF_BOUNDARY;
982 		}
983 		PROC_SUNLOCK(p);
984 		mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
985 		thread_unlock(td);
986 		PROC_LOCK(p);
987 	}
988 	return (0);
989 }
990 
991 void
992 thread_suspend_switch(struct thread *td, struct proc *p)
993 {
994 
995 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
996 	PROC_LOCK_ASSERT(p, MA_OWNED);
997 	PROC_SLOCK_ASSERT(p, MA_OWNED);
998 	/*
999 	 * We implement thread_suspend_one in stages here to avoid
1000 	 * dropping the proc lock while the thread lock is owned.
1001 	 */
1002 	if (p == td->td_proc) {
1003 		thread_stopped(p);
1004 		p->p_suspcount++;
1005 	}
1006 	PROC_UNLOCK(p);
1007 	thread_lock(td);
1008 	td->td_flags &= ~TDF_NEEDSUSPCHK;
1009 	TD_SET_SUSPENDED(td);
1010 	sched_sleep(td, 0);
1011 	PROC_SUNLOCK(p);
1012 	DROP_GIANT();
1013 	mi_switch(SW_VOL | SWT_SUSPEND, NULL);
1014 	thread_unlock(td);
1015 	PICKUP_GIANT();
1016 	PROC_LOCK(p);
1017 	PROC_SLOCK(p);
1018 }
1019 
1020 void
1021 thread_suspend_one(struct thread *td)
1022 {
1023 	struct proc *p;
1024 
1025 	p = td->td_proc;
1026 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1027 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1028 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1029 	p->p_suspcount++;
1030 	td->td_flags &= ~TDF_NEEDSUSPCHK;
1031 	TD_SET_SUSPENDED(td);
1032 	sched_sleep(td, 0);
1033 }
1034 
1035 static int
1036 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1037 {
1038 
1039 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1040 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1041 	TD_CLR_SUSPENDED(td);
1042 	td->td_flags &= ~TDF_ALLPROCSUSP;
1043 	if (td->td_proc == p) {
1044 		PROC_SLOCK_ASSERT(p, MA_OWNED);
1045 		p->p_suspcount--;
1046 		if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1047 			td->td_flags &= ~TDF_BOUNDARY;
1048 			p->p_boundary_count--;
1049 		}
1050 	}
1051 	return (setrunnable(td));
1052 }
1053 
1054 /*
1055  * Allow all threads blocked by single threading to continue running.
1056  */
1057 void
1058 thread_unsuspend(struct proc *p)
1059 {
1060 	struct thread *td;
1061 	int wakeup_swapper;
1062 
1063 	PROC_LOCK_ASSERT(p, MA_OWNED);
1064 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1065 	wakeup_swapper = 0;
1066 	if (!P_SHOULDSTOP(p)) {
1067                 FOREACH_THREAD_IN_PROC(p, td) {
1068 			thread_lock(td);
1069 			if (TD_IS_SUSPENDED(td)) {
1070 				wakeup_swapper |= thread_unsuspend_one(td, p,
1071 				    true);
1072 			}
1073 			thread_unlock(td);
1074 		}
1075 	} else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1076 	    p->p_numthreads == p->p_suspcount) {
1077 		/*
1078 		 * Stopping everything also did the job for the single
1079 		 * threading request. Now we've downgraded to single-threaded,
1080 		 * let it continue.
1081 		 */
1082 		if (p->p_singlethread->td_proc == p) {
1083 			thread_lock(p->p_singlethread);
1084 			wakeup_swapper = thread_unsuspend_one(
1085 			    p->p_singlethread, p, false);
1086 			thread_unlock(p->p_singlethread);
1087 		}
1088 	}
1089 	if (wakeup_swapper)
1090 		kick_proc0();
1091 }
1092 
1093 /*
1094  * End the single threading mode..
1095  */
1096 void
1097 thread_single_end(struct proc *p, int mode)
1098 {
1099 	struct thread *td;
1100 	int wakeup_swapper;
1101 
1102 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1103 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1104 	    ("invalid mode %d", mode));
1105 	PROC_LOCK_ASSERT(p, MA_OWNED);
1106 	KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1107 	    (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1108 	    ("mode %d does not match P_TOTAL_STOP", mode));
1109 	KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1110 	    ("thread_single_end from other thread %p %p",
1111 	    curthread, p->p_singlethread));
1112 	KASSERT(mode != SINGLE_BOUNDARY ||
1113 	    (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1114 	    ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1115 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1116 	    P_TOTAL_STOP);
1117 	PROC_SLOCK(p);
1118 	p->p_singlethread = NULL;
1119 	wakeup_swapper = 0;
1120 	/*
1121 	 * If there are other threads they may now run,
1122 	 * unless of course there is a blanket 'stop order'
1123 	 * on the process. The single threader must be allowed
1124 	 * to continue however as this is a bad place to stop.
1125 	 */
1126 	if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1127                 FOREACH_THREAD_IN_PROC(p, td) {
1128 			thread_lock(td);
1129 			if (TD_IS_SUSPENDED(td)) {
1130 				wakeup_swapper |= thread_unsuspend_one(td, p,
1131 				    mode == SINGLE_BOUNDARY);
1132 			}
1133 			thread_unlock(td);
1134 		}
1135 	}
1136 	KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1137 	    ("inconsistent boundary count %d", p->p_boundary_count));
1138 	PROC_SUNLOCK(p);
1139 	if (wakeup_swapper)
1140 		kick_proc0();
1141 }
1142 
1143 struct thread *
1144 thread_find(struct proc *p, lwpid_t tid)
1145 {
1146 	struct thread *td;
1147 
1148 	PROC_LOCK_ASSERT(p, MA_OWNED);
1149 	FOREACH_THREAD_IN_PROC(p, td) {
1150 		if (td->td_tid == tid)
1151 			break;
1152 	}
1153 	return (td);
1154 }
1155 
1156 /* Locate a thread by number; return with proc lock held. */
1157 struct thread *
1158 tdfind(lwpid_t tid, pid_t pid)
1159 {
1160 #define RUN_THRESH	16
1161 	struct thread *td;
1162 	int run = 0;
1163 
1164 	rw_rlock(&tidhash_lock);
1165 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1166 		if (td->td_tid == tid) {
1167 			if (pid != -1 && td->td_proc->p_pid != pid) {
1168 				td = NULL;
1169 				break;
1170 			}
1171 			PROC_LOCK(td->td_proc);
1172 			if (td->td_proc->p_state == PRS_NEW) {
1173 				PROC_UNLOCK(td->td_proc);
1174 				td = NULL;
1175 				break;
1176 			}
1177 			if (run > RUN_THRESH) {
1178 				if (rw_try_upgrade(&tidhash_lock)) {
1179 					LIST_REMOVE(td, td_hash);
1180 					LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1181 						td, td_hash);
1182 					rw_wunlock(&tidhash_lock);
1183 					return (td);
1184 				}
1185 			}
1186 			break;
1187 		}
1188 		run++;
1189 	}
1190 	rw_runlock(&tidhash_lock);
1191 	return (td);
1192 }
1193 
1194 void
1195 tidhash_add(struct thread *td)
1196 {
1197 	rw_wlock(&tidhash_lock);
1198 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1199 	rw_wunlock(&tidhash_lock);
1200 }
1201 
1202 void
1203 tidhash_remove(struct thread *td)
1204 {
1205 	rw_wlock(&tidhash_lock);
1206 	LIST_REMOVE(td, td_hash);
1207 	rw_wunlock(&tidhash_lock);
1208 }
1209