xref: /freebsd/sys/kern/kern_thread.c (revision b2bf0c7e5f4037d63458def91a026592468afd2f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
5  *  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice(s), this list of conditions and the following disclaimer as
12  *    the first lines of this file unmodified other than the possible
13  *    addition of one or more copyright notices.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice(s), this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
19  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
22  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
25  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
28  * DAMAGE.
29  */
30 
31 #include "opt_witness.h"
32 #include "opt_hwpmc_hooks.h"
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/lock.h>
41 #include <sys/msan.h>
42 #include <sys/mutex.h>
43 #include <sys/proc.h>
44 #include <sys/bitstring.h>
45 #include <sys/epoch.h>
46 #include <sys/rangelock.h>
47 #include <sys/resourcevar.h>
48 #include <sys/sdt.h>
49 #include <sys/smp.h>
50 #include <sys/sched.h>
51 #include <sys/sleepqueue.h>
52 #include <sys/selinfo.h>
53 #include <sys/syscallsubr.h>
54 #include <sys/dtrace_bsd.h>
55 #include <sys/sysent.h>
56 #include <sys/turnstile.h>
57 #include <sys/taskqueue.h>
58 #include <sys/ktr.h>
59 #include <sys/rwlock.h>
60 #include <sys/umtxvar.h>
61 #include <sys/vmmeter.h>
62 #include <sys/cpuset.h>
63 #ifdef	HWPMC_HOOKS
64 #include <sys/pmckern.h>
65 #endif
66 #include <sys/priv.h>
67 
68 #include <security/audit/audit.h>
69 
70 #include <vm/pmap.h>
71 #include <vm/vm.h>
72 #include <vm/vm_extern.h>
73 #include <vm/uma.h>
74 #include <vm/vm_phys.h>
75 #include <sys/eventhandler.h>
76 
77 /*
78  * Asserts below verify the stability of struct thread and struct proc
79  * layout, as exposed by KBI to modules.  On head, the KBI is allowed
80  * to drift, change to the structures must be accompanied by the
81  * assert update.
82  *
83  * On the stable branches after KBI freeze, conditions must not be
84  * violated.  Typically new fields are moved to the end of the
85  * structures.
86  */
87 #ifdef __amd64__
88 _Static_assert(offsetof(struct thread, td_flags) == 0x108,
89     "struct thread KBI td_flags");
90 _Static_assert(offsetof(struct thread, td_pflags) == 0x110,
91     "struct thread KBI td_pflags");
92 _Static_assert(offsetof(struct thread, td_frame) == 0x4a8,
93     "struct thread KBI td_frame");
94 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6b0,
95     "struct thread KBI td_emuldata");
96 _Static_assert(offsetof(struct proc, p_flag) == 0xb8,
97     "struct proc KBI p_flag");
98 _Static_assert(offsetof(struct proc, p_pid) == 0xc4,
99     "struct proc KBI p_pid");
100 _Static_assert(offsetof(struct proc, p_filemon) == 0x3c8,
101     "struct proc KBI p_filemon");
102 _Static_assert(offsetof(struct proc, p_comm) == 0x3e0,
103     "struct proc KBI p_comm");
104 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4c8,
105     "struct proc KBI p_emuldata");
106 #endif
107 #ifdef __i386__
108 _Static_assert(offsetof(struct thread, td_flags) == 0x9c,
109     "struct thread KBI td_flags");
110 _Static_assert(offsetof(struct thread, td_pflags) == 0xa4,
111     "struct thread KBI td_pflags");
112 _Static_assert(offsetof(struct thread, td_frame) == 0x308,
113     "struct thread KBI td_frame");
114 _Static_assert(offsetof(struct thread, td_emuldata) == 0x34c,
115     "struct thread KBI td_emuldata");
116 _Static_assert(offsetof(struct proc, p_flag) == 0x6c,
117     "struct proc KBI p_flag");
118 _Static_assert(offsetof(struct proc, p_pid) == 0x78,
119     "struct proc KBI p_pid");
120 _Static_assert(offsetof(struct proc, p_filemon) == 0x270,
121     "struct proc KBI p_filemon");
122 _Static_assert(offsetof(struct proc, p_comm) == 0x284,
123     "struct proc KBI p_comm");
124 _Static_assert(offsetof(struct proc, p_emuldata) == 0x310,
125     "struct proc KBI p_emuldata");
126 #endif
127 
128 SDT_PROVIDER_DECLARE(proc);
129 SDT_PROBE_DEFINE(proc, , , lwp__exit);
130 
131 /*
132  * thread related storage.
133  */
134 static uma_zone_t thread_zone;
135 
136 struct thread_domain_data {
137 	struct thread	*tdd_zombies;
138 	int		tdd_reapticks;
139 } __aligned(CACHE_LINE_SIZE);
140 
141 static struct thread_domain_data thread_domain_data[MAXMEMDOM];
142 
143 static struct task	thread_reap_task;
144 static struct callout  	thread_reap_callout;
145 
146 static void thread_zombie(struct thread *);
147 static void thread_reap(void);
148 static void thread_reap_all(void);
149 static void thread_reap_task_cb(void *, int);
150 static void thread_reap_callout_cb(void *);
151 static int thread_unsuspend_one(struct thread *td, struct proc *p,
152     bool boundary);
153 static void thread_free_batched(struct thread *td);
154 
155 static __exclusive_cache_line struct mtx tid_lock;
156 static bitstr_t *tid_bitmap;
157 
158 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
159 
160 static int maxthread;
161 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
162     &maxthread, 0, "Maximum number of threads");
163 
164 static __exclusive_cache_line int nthreads;
165 
166 static LIST_HEAD(tidhashhead, thread) *tidhashtbl;
167 static u_long	tidhash;
168 static u_long	tidhashlock;
169 static struct	rwlock *tidhashtbl_lock;
170 #define	TIDHASH(tid)		(&tidhashtbl[(tid) & tidhash])
171 #define	TIDHASHLOCK(tid)	(&tidhashtbl_lock[(tid) & tidhashlock])
172 
173 EVENTHANDLER_LIST_DEFINE(thread_ctor);
174 EVENTHANDLER_LIST_DEFINE(thread_dtor);
175 EVENTHANDLER_LIST_DEFINE(thread_init);
176 EVENTHANDLER_LIST_DEFINE(thread_fini);
177 
178 static bool
179 thread_count_inc_try(void)
180 {
181 	int nthreads_new;
182 
183 	nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1;
184 	if (nthreads_new >= maxthread - 100) {
185 		if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
186 		    nthreads_new >= maxthread) {
187 			atomic_subtract_int(&nthreads, 1);
188 			return (false);
189 		}
190 	}
191 	return (true);
192 }
193 
194 static bool
195 thread_count_inc(void)
196 {
197 	static struct timeval lastfail;
198 	static int curfail;
199 
200 	thread_reap();
201 	if (thread_count_inc_try()) {
202 		return (true);
203 	}
204 
205 	thread_reap_all();
206 	if (thread_count_inc_try()) {
207 		return (true);
208 	}
209 
210 	if (ppsratecheck(&lastfail, &curfail, 1)) {
211 		printf("maxthread limit exceeded by uid %u "
212 		    "(pid %d); consider increasing kern.maxthread\n",
213 		    curthread->td_ucred->cr_ruid, curproc->p_pid);
214 	}
215 	return (false);
216 }
217 
218 static void
219 thread_count_sub(int n)
220 {
221 
222 	atomic_subtract_int(&nthreads, n);
223 }
224 
225 static void
226 thread_count_dec(void)
227 {
228 
229 	thread_count_sub(1);
230 }
231 
232 static lwpid_t
233 tid_alloc(void)
234 {
235 	static lwpid_t trytid;
236 	lwpid_t tid;
237 
238 	mtx_lock(&tid_lock);
239 	/*
240 	 * It is an invariant that the bitmap is big enough to hold maxthread
241 	 * IDs. If we got to this point there has to be at least one free.
242 	 */
243 	if (trytid >= maxthread)
244 		trytid = 0;
245 	bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
246 	if (tid == -1) {
247 		KASSERT(trytid != 0, ("unexpectedly ran out of IDs"));
248 		trytid = 0;
249 		bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
250 		KASSERT(tid != -1, ("unexpectedly ran out of IDs"));
251 	}
252 	bit_set(tid_bitmap, tid);
253 	trytid = tid + 1;
254 	mtx_unlock(&tid_lock);
255 	return (tid + NO_PID);
256 }
257 
258 static void
259 tid_free_locked(lwpid_t rtid)
260 {
261 	lwpid_t tid;
262 
263 	mtx_assert(&tid_lock, MA_OWNED);
264 	KASSERT(rtid >= NO_PID,
265 	    ("%s: invalid tid %d\n", __func__, rtid));
266 	tid = rtid - NO_PID;
267 	KASSERT(bit_test(tid_bitmap, tid) != 0,
268 	    ("thread ID %d not allocated\n", rtid));
269 	bit_clear(tid_bitmap, tid);
270 }
271 
272 static void
273 tid_free(lwpid_t rtid)
274 {
275 
276 	mtx_lock(&tid_lock);
277 	tid_free_locked(rtid);
278 	mtx_unlock(&tid_lock);
279 }
280 
281 static void
282 tid_free_batch(lwpid_t *batch, int n)
283 {
284 	int i;
285 
286 	mtx_lock(&tid_lock);
287 	for (i = 0; i < n; i++) {
288 		tid_free_locked(batch[i]);
289 	}
290 	mtx_unlock(&tid_lock);
291 }
292 
293 /*
294  * Batching for thread reapping.
295  */
296 struct tidbatch {
297 	lwpid_t tab[16];
298 	int n;
299 };
300 
301 static void
302 tidbatch_prep(struct tidbatch *tb)
303 {
304 
305 	tb->n = 0;
306 }
307 
308 static void
309 tidbatch_add(struct tidbatch *tb, struct thread *td)
310 {
311 
312 	KASSERT(tb->n < nitems(tb->tab),
313 	    ("%s: count too high %d", __func__, tb->n));
314 	tb->tab[tb->n] = td->td_tid;
315 	tb->n++;
316 }
317 
318 static void
319 tidbatch_process(struct tidbatch *tb)
320 {
321 
322 	KASSERT(tb->n <= nitems(tb->tab),
323 	    ("%s: count too high %d", __func__, tb->n));
324 	if (tb->n == nitems(tb->tab)) {
325 		tid_free_batch(tb->tab, tb->n);
326 		tb->n = 0;
327 	}
328 }
329 
330 static void
331 tidbatch_final(struct tidbatch *tb)
332 {
333 
334 	KASSERT(tb->n <= nitems(tb->tab),
335 	    ("%s: count too high %d", __func__, tb->n));
336 	if (tb->n != 0) {
337 		tid_free_batch(tb->tab, tb->n);
338 	}
339 }
340 
341 /*
342  * Prepare a thread for use.
343  */
344 static int
345 thread_ctor(void *mem, int size, void *arg, int flags)
346 {
347 	struct thread	*td;
348 
349 	td = (struct thread *)mem;
350 	TD_SET_STATE(td, TDS_INACTIVE);
351 	td->td_lastcpu = td->td_oncpu = NOCPU;
352 
353 	/*
354 	 * Note that td_critnest begins life as 1 because the thread is not
355 	 * running and is thereby implicitly waiting to be on the receiving
356 	 * end of a context switch.
357 	 */
358 	td->td_critnest = 1;
359 	td->td_lend_user_pri = PRI_MAX;
360 #ifdef AUDIT
361 	audit_thread_alloc(td);
362 #endif
363 #ifdef KDTRACE_HOOKS
364 	kdtrace_thread_ctor(td);
365 #endif
366 	umtx_thread_alloc(td);
367 	MPASS(td->td_sel == NULL);
368 	return (0);
369 }
370 
371 /*
372  * Reclaim a thread after use.
373  */
374 static void
375 thread_dtor(void *mem, int size, void *arg)
376 {
377 	struct thread *td;
378 
379 	td = (struct thread *)mem;
380 
381 #ifdef INVARIANTS
382 	/* Verify that this thread is in a safe state to free. */
383 	switch (TD_GET_STATE(td)) {
384 	case TDS_INHIBITED:
385 	case TDS_RUNNING:
386 	case TDS_CAN_RUN:
387 	case TDS_RUNQ:
388 		/*
389 		 * We must never unlink a thread that is in one of
390 		 * these states, because it is currently active.
391 		 */
392 		panic("bad state for thread unlinking");
393 		/* NOTREACHED */
394 	case TDS_INACTIVE:
395 		break;
396 	default:
397 		panic("bad thread state");
398 		/* NOTREACHED */
399 	}
400 #endif
401 #ifdef AUDIT
402 	audit_thread_free(td);
403 #endif
404 #ifdef KDTRACE_HOOKS
405 	kdtrace_thread_dtor(td);
406 #endif
407 	/* Free all OSD associated to this thread. */
408 	osd_thread_exit(td);
409 	td_softdep_cleanup(td);
410 	MPASS(td->td_su == NULL);
411 	seltdfini(td);
412 }
413 
414 /*
415  * Initialize type-stable parts of a thread (when newly created).
416  */
417 static int
418 thread_init(void *mem, int size, int flags)
419 {
420 	struct thread *td;
421 
422 	td = (struct thread *)mem;
423 
424 	td->td_allocdomain = vm_phys_domain(vtophys(td));
425 	td->td_sleepqueue = sleepq_alloc();
426 	td->td_turnstile = turnstile_alloc();
427 	td->td_rlqe = NULL;
428 	EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
429 	umtx_thread_init(td);
430 	td->td_kstack = 0;
431 	td->td_sel = NULL;
432 	return (0);
433 }
434 
435 /*
436  * Tear down type-stable parts of a thread (just before being discarded).
437  */
438 static void
439 thread_fini(void *mem, int size)
440 {
441 	struct thread *td;
442 
443 	td = (struct thread *)mem;
444 	EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
445 	rlqentry_free(td->td_rlqe);
446 	turnstile_free(td->td_turnstile);
447 	sleepq_free(td->td_sleepqueue);
448 	umtx_thread_fini(td);
449 	MPASS(td->td_sel == NULL);
450 }
451 
452 /*
453  * For a newly created process,
454  * link up all the structures and its initial threads etc.
455  * called from:
456  * {arch}/{arch}/machdep.c   {arch}_init(), init386() etc.
457  * proc_dtor() (should go away)
458  * proc_init()
459  */
460 void
461 proc_linkup0(struct proc *p, struct thread *td)
462 {
463 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
464 	proc_linkup(p, td);
465 }
466 
467 void
468 proc_linkup(struct proc *p, struct thread *td)
469 {
470 
471 	sigqueue_init(&p->p_sigqueue, p);
472 	p->p_ksi = ksiginfo_alloc(1);
473 	if (p->p_ksi != NULL) {
474 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
475 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
476 	}
477 	LIST_INIT(&p->p_mqnotifier);
478 	p->p_numthreads = 0;
479 	thread_link(td, p);
480 }
481 
482 extern int max_threads_per_proc;
483 
484 /*
485  * Initialize global thread allocation resources.
486  */
487 void
488 threadinit(void)
489 {
490 	u_long i;
491 	lwpid_t tid0;
492 	uint32_t flags;
493 
494 	/*
495 	 * Place an upper limit on threads which can be allocated.
496 	 *
497 	 * Note that other factors may make the de facto limit much lower.
498 	 *
499 	 * Platform limits are somewhat arbitrary but deemed "more than good
500 	 * enough" for the foreseable future.
501 	 */
502 	if (maxthread == 0) {
503 #ifdef _LP64
504 		maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
505 #else
506 		maxthread = MIN(maxproc * max_threads_per_proc, 100000);
507 #endif
508 	}
509 
510 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
511 	tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
512 	/*
513 	 * Handle thread0.
514 	 */
515 	thread_count_inc();
516 	tid0 = tid_alloc();
517 	if (tid0 != THREAD0_TID)
518 		panic("tid0 %d != %d\n", tid0, THREAD0_TID);
519 
520 	flags = UMA_ZONE_NOFREE;
521 #ifdef __aarch64__
522 	/*
523 	 * Force thread structures to be allocated from the direct map.
524 	 * Otherwise, superpage promotions and demotions may temporarily
525 	 * invalidate thread structure mappings.  For most dynamically allocated
526 	 * structures this is not a problem, but translation faults cannot be
527 	 * handled without accessing curthread.
528 	 */
529 	flags |= UMA_ZONE_CONTIG;
530 #endif
531 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
532 	    thread_ctor, thread_dtor, thread_init, thread_fini,
533 	    32 - 1, flags);
534 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
535 	tidhashlock = (tidhash + 1) / 64;
536 	if (tidhashlock > 0)
537 		tidhashlock--;
538 	tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
539 	    M_TIDHASH, M_WAITOK | M_ZERO);
540 	for (i = 0; i < tidhashlock + 1; i++)
541 		rw_init(&tidhashtbl_lock[i], "tidhash");
542 
543 	TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL);
544 	callout_init(&thread_reap_callout, 1);
545 	callout_reset(&thread_reap_callout, 5 * hz,
546 	    thread_reap_callout_cb, NULL);
547 }
548 
549 /*
550  * Place an unused thread on the zombie list.
551  */
552 void
553 thread_zombie(struct thread *td)
554 {
555 	struct thread_domain_data *tdd;
556 	struct thread *ztd;
557 
558 	tdd = &thread_domain_data[td->td_allocdomain];
559 	ztd = atomic_load_ptr(&tdd->tdd_zombies);
560 	for (;;) {
561 		td->td_zombie = ztd;
562 		if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
563 		    (uintptr_t *)&ztd, (uintptr_t)td))
564 			break;
565 		continue;
566 	}
567 }
568 
569 /*
570  * Release a thread that has exited after cpu_throw().
571  */
572 void
573 thread_stash(struct thread *td)
574 {
575 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
576 	thread_zombie(td);
577 }
578 
579 /*
580  * Reap zombies from passed domain.
581  */
582 static void
583 thread_reap_domain(struct thread_domain_data *tdd)
584 {
585 	struct thread *itd, *ntd;
586 	struct tidbatch tidbatch;
587 	struct credbatch credbatch;
588 	int tdcount;
589 	struct plimit *lim;
590 	int limcount;
591 
592 	/*
593 	 * Reading upfront is pessimal if followed by concurrent atomic_swap,
594 	 * but most of the time the list is empty.
595 	 */
596 	if (tdd->tdd_zombies == NULL)
597 		return;
598 
599 	itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
600 	    (uintptr_t)NULL);
601 	if (itd == NULL)
602 		return;
603 
604 	/*
605 	 * Multiple CPUs can get here, the race is fine as ticks is only
606 	 * advisory.
607 	 */
608 	tdd->tdd_reapticks = ticks;
609 
610 	tidbatch_prep(&tidbatch);
611 	credbatch_prep(&credbatch);
612 	tdcount = 0;
613 	lim = NULL;
614 	limcount = 0;
615 
616 	while (itd != NULL) {
617 		ntd = itd->td_zombie;
618 		EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd);
619 		tidbatch_add(&tidbatch, itd);
620 		credbatch_add(&credbatch, itd);
621 		MPASS(itd->td_limit != NULL);
622 		if (lim != itd->td_limit) {
623 			if (limcount != 0) {
624 				lim_freen(lim, limcount);
625 				limcount = 0;
626 			}
627 		}
628 		lim = itd->td_limit;
629 		limcount++;
630 		thread_free_batched(itd);
631 		tidbatch_process(&tidbatch);
632 		credbatch_process(&credbatch);
633 		tdcount++;
634 		if (tdcount == 32) {
635 			thread_count_sub(tdcount);
636 			tdcount = 0;
637 		}
638 		itd = ntd;
639 	}
640 
641 	tidbatch_final(&tidbatch);
642 	credbatch_final(&credbatch);
643 	if (tdcount != 0) {
644 		thread_count_sub(tdcount);
645 	}
646 	MPASS(limcount != 0);
647 	lim_freen(lim, limcount);
648 }
649 
650 /*
651  * Reap zombies from all domains.
652  */
653 static void
654 thread_reap_all(void)
655 {
656 	struct thread_domain_data *tdd;
657 	int i, domain;
658 
659 	domain = PCPU_GET(domain);
660 	for (i = 0; i < vm_ndomains; i++) {
661 		tdd = &thread_domain_data[(i + domain) % vm_ndomains];
662 		thread_reap_domain(tdd);
663 	}
664 }
665 
666 /*
667  * Reap zombies from local domain.
668  */
669 static void
670 thread_reap(void)
671 {
672 	struct thread_domain_data *tdd;
673 	int domain;
674 
675 	domain = PCPU_GET(domain);
676 	tdd = &thread_domain_data[domain];
677 
678 	thread_reap_domain(tdd);
679 }
680 
681 static void
682 thread_reap_task_cb(void *arg __unused, int pending __unused)
683 {
684 
685 	thread_reap_all();
686 }
687 
688 static void
689 thread_reap_callout_cb(void *arg __unused)
690 {
691 	struct thread_domain_data *tdd;
692 	int i, cticks, lticks;
693 	bool wantreap;
694 
695 	wantreap = false;
696 	cticks = atomic_load_int(&ticks);
697 	for (i = 0; i < vm_ndomains; i++) {
698 		tdd = &thread_domain_data[i];
699 		lticks = tdd->tdd_reapticks;
700 		if (tdd->tdd_zombies != NULL &&
701 		    (u_int)(cticks - lticks) > 5 * hz) {
702 			wantreap = true;
703 			break;
704 		}
705 	}
706 
707 	if (wantreap)
708 		taskqueue_enqueue(taskqueue_thread, &thread_reap_task);
709 	callout_reset(&thread_reap_callout, 5 * hz,
710 	    thread_reap_callout_cb, NULL);
711 }
712 
713 /*
714  * Calling this function guarantees that any thread that exited before
715  * the call is reaped when the function returns.  By 'exited' we mean
716  * a thread removed from the process linkage with thread_unlink().
717  * Practically this means that caller must lock/unlock corresponding
718  * process lock before the call, to synchronize with thread_exit().
719  */
720 void
721 thread_reap_barrier(void)
722 {
723 	struct task *t;
724 
725 	/*
726 	 * First do context switches to each CPU to ensure that all
727 	 * PCPU pc_deadthreads are moved to zombie list.
728 	 */
729 	quiesce_all_cpus("", PDROP);
730 
731 	/*
732 	 * Second, fire the task in the same thread as normal
733 	 * thread_reap() is done, to serialize reaping.
734 	 */
735 	t = malloc(sizeof(*t), M_TEMP, M_WAITOK);
736 	TASK_INIT(t, 0, thread_reap_task_cb, t);
737 	taskqueue_enqueue(taskqueue_thread, t);
738 	taskqueue_drain(taskqueue_thread, t);
739 	free(t, M_TEMP);
740 }
741 
742 /*
743  * Allocate a thread.
744  */
745 struct thread *
746 thread_alloc(int pages)
747 {
748 	struct thread *td;
749 	lwpid_t tid;
750 
751 	if (!thread_count_inc()) {
752 		return (NULL);
753 	}
754 
755 	tid = tid_alloc();
756 	td = uma_zalloc(thread_zone, M_WAITOK);
757 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
758 	if (!vm_thread_new(td, pages)) {
759 		uma_zfree(thread_zone, td);
760 		tid_free(tid);
761 		thread_count_dec();
762 		return (NULL);
763 	}
764 	td->td_tid = tid;
765 	bzero(&td->td_sa.args, sizeof(td->td_sa.args));
766 	kmsan_thread_alloc(td);
767 	cpu_thread_alloc(td);
768 	EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
769 	return (td);
770 }
771 
772 int
773 thread_alloc_stack(struct thread *td, int pages)
774 {
775 
776 	KASSERT(td->td_kstack == 0,
777 	    ("thread_alloc_stack called on a thread with kstack"));
778 	if (!vm_thread_new(td, pages))
779 		return (0);
780 	cpu_thread_alloc(td);
781 	return (1);
782 }
783 
784 /*
785  * Deallocate a thread.
786  */
787 static void
788 thread_free_batched(struct thread *td)
789 {
790 
791 	lock_profile_thread_exit(td);
792 	if (td->td_cpuset)
793 		cpuset_rel(td->td_cpuset);
794 	td->td_cpuset = NULL;
795 	cpu_thread_free(td);
796 	if (td->td_kstack != 0)
797 		vm_thread_dispose(td);
798 	callout_drain(&td->td_slpcallout);
799 	/*
800 	 * Freeing handled by the caller.
801 	 */
802 	td->td_tid = -1;
803 	kmsan_thread_free(td);
804 	uma_zfree(thread_zone, td);
805 }
806 
807 void
808 thread_free(struct thread *td)
809 {
810 	lwpid_t tid;
811 
812 	EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
813 	tid = td->td_tid;
814 	thread_free_batched(td);
815 	tid_free(tid);
816 	thread_count_dec();
817 }
818 
819 void
820 thread_cow_get_proc(struct thread *newtd, struct proc *p)
821 {
822 
823 	PROC_LOCK_ASSERT(p, MA_OWNED);
824 	newtd->td_realucred = crcowget(p->p_ucred);
825 	newtd->td_ucred = newtd->td_realucred;
826 	newtd->td_limit = lim_hold(p->p_limit);
827 	newtd->td_cowgen = p->p_cowgen;
828 }
829 
830 void
831 thread_cow_get(struct thread *newtd, struct thread *td)
832 {
833 
834 	MPASS(td->td_realucred == td->td_ucred);
835 	newtd->td_realucred = crcowget(td->td_realucred);
836 	newtd->td_ucred = newtd->td_realucred;
837 	newtd->td_limit = lim_hold(td->td_limit);
838 	newtd->td_cowgen = td->td_cowgen;
839 }
840 
841 void
842 thread_cow_free(struct thread *td)
843 {
844 
845 	if (td->td_realucred != NULL)
846 		crcowfree(td);
847 	if (td->td_limit != NULL)
848 		lim_free(td->td_limit);
849 }
850 
851 void
852 thread_cow_update(struct thread *td)
853 {
854 	struct proc *p;
855 	struct ucred *oldcred;
856 	struct plimit *oldlimit;
857 
858 	p = td->td_proc;
859 	oldlimit = NULL;
860 	PROC_LOCK(p);
861 	oldcred = crcowsync();
862 	if (td->td_limit != p->p_limit) {
863 		oldlimit = td->td_limit;
864 		td->td_limit = lim_hold(p->p_limit);
865 	}
866 	td->td_cowgen = p->p_cowgen;
867 	PROC_UNLOCK(p);
868 	if (oldcred != NULL)
869 		crfree(oldcred);
870 	if (oldlimit != NULL)
871 		lim_free(oldlimit);
872 }
873 
874 /*
875  * Discard the current thread and exit from its context.
876  * Always called with scheduler locked.
877  *
878  * Because we can't free a thread while we're operating under its context,
879  * push the current thread into our CPU's deadthread holder. This means
880  * we needn't worry about someone else grabbing our context before we
881  * do a cpu_throw().
882  */
883 void
884 thread_exit(void)
885 {
886 	uint64_t runtime, new_switchtime;
887 	struct thread *td;
888 	struct thread *td2;
889 	struct proc *p;
890 	int wakeup_swapper;
891 
892 	td = curthread;
893 	p = td->td_proc;
894 
895 	PROC_SLOCK_ASSERT(p, MA_OWNED);
896 	mtx_assert(&Giant, MA_NOTOWNED);
897 
898 	PROC_LOCK_ASSERT(p, MA_OWNED);
899 	KASSERT(p != NULL, ("thread exiting without a process"));
900 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
901 	    (long)p->p_pid, td->td_name);
902 	SDT_PROBE0(proc, , , lwp__exit);
903 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
904 	MPASS(td->td_realucred == td->td_ucred);
905 
906 	/*
907 	 * drop FPU & debug register state storage, or any other
908 	 * architecture specific resources that
909 	 * would not be on a new untouched process.
910 	 */
911 	cpu_thread_exit(td);
912 
913 	/*
914 	 * The last thread is left attached to the process
915 	 * So that the whole bundle gets recycled. Skip
916 	 * all this stuff if we never had threads.
917 	 * EXIT clears all sign of other threads when
918 	 * it goes to single threading, so the last thread always
919 	 * takes the short path.
920 	 */
921 	if (p->p_flag & P_HADTHREADS) {
922 		if (p->p_numthreads > 1) {
923 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
924 			thread_unlink(td);
925 			td2 = FIRST_THREAD_IN_PROC(p);
926 			sched_exit_thread(td2, td);
927 
928 			/*
929 			 * The test below is NOT true if we are the
930 			 * sole exiting thread. P_STOPPED_SINGLE is unset
931 			 * in exit1() after it is the only survivor.
932 			 */
933 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
934 				if (p->p_numthreads == p->p_suspcount) {
935 					thread_lock(p->p_singlethread);
936 					wakeup_swapper = thread_unsuspend_one(
937 						p->p_singlethread, p, false);
938 					if (wakeup_swapper)
939 						kick_proc0();
940 				}
941 			}
942 
943 			PCPU_SET(deadthread, td);
944 		} else {
945 			/*
946 			 * The last thread is exiting.. but not through exit()
947 			 */
948 			panic ("thread_exit: Last thread exiting on its own");
949 		}
950 	}
951 #ifdef	HWPMC_HOOKS
952 	/*
953 	 * If this thread is part of a process that is being tracked by hwpmc(4),
954 	 * inform the module of the thread's impending exit.
955 	 */
956 	if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
957 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
958 		PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
959 	} else if (PMC_SYSTEM_SAMPLING_ACTIVE())
960 		PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
961 #endif
962 	PROC_UNLOCK(p);
963 	PROC_STATLOCK(p);
964 	thread_lock(td);
965 	PROC_SUNLOCK(p);
966 
967 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
968 	new_switchtime = cpu_ticks();
969 	runtime = new_switchtime - PCPU_GET(switchtime);
970 	td->td_runtime += runtime;
971 	td->td_incruntime += runtime;
972 	PCPU_SET(switchtime, new_switchtime);
973 	PCPU_SET(switchticks, ticks);
974 	VM_CNT_INC(v_swtch);
975 
976 	/* Save our resource usage in our process. */
977 	td->td_ru.ru_nvcsw++;
978 	ruxagg_locked(p, td);
979 	rucollect(&p->p_ru, &td->td_ru);
980 	PROC_STATUNLOCK(p);
981 
982 	TD_SET_STATE(td, TDS_INACTIVE);
983 #ifdef WITNESS
984 	witness_thread_exit(td);
985 #endif
986 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
987 	sched_throw(td);
988 	panic("I'm a teapot!");
989 	/* NOTREACHED */
990 }
991 
992 /*
993  * Do any thread specific cleanups that may be needed in wait()
994  * called with Giant, proc and schedlock not held.
995  */
996 void
997 thread_wait(struct proc *p)
998 {
999 	struct thread *td;
1000 
1001 	mtx_assert(&Giant, MA_NOTOWNED);
1002 	KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
1003 	KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
1004 	td = FIRST_THREAD_IN_PROC(p);
1005 	/* Lock the last thread so we spin until it exits cpu_throw(). */
1006 	thread_lock(td);
1007 	thread_unlock(td);
1008 	lock_profile_thread_exit(td);
1009 	cpuset_rel(td->td_cpuset);
1010 	td->td_cpuset = NULL;
1011 	cpu_thread_clean(td);
1012 	thread_cow_free(td);
1013 	callout_drain(&td->td_slpcallout);
1014 	thread_reap();	/* check for zombie threads etc. */
1015 }
1016 
1017 /*
1018  * Link a thread to a process.
1019  * set up anything that needs to be initialized for it to
1020  * be used by the process.
1021  */
1022 void
1023 thread_link(struct thread *td, struct proc *p)
1024 {
1025 
1026 	/*
1027 	 * XXX This can't be enabled because it's called for proc0 before
1028 	 * its lock has been created.
1029 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
1030 	 */
1031 	TD_SET_STATE(td, TDS_INACTIVE);
1032 	td->td_proc     = p;
1033 	td->td_flags    = TDF_INMEM;
1034 
1035 	LIST_INIT(&td->td_contested);
1036 	LIST_INIT(&td->td_lprof[0]);
1037 	LIST_INIT(&td->td_lprof[1]);
1038 #ifdef EPOCH_TRACE
1039 	SLIST_INIT(&td->td_epochs);
1040 #endif
1041 	sigqueue_init(&td->td_sigqueue, p);
1042 	callout_init(&td->td_slpcallout, 1);
1043 	TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
1044 	p->p_numthreads++;
1045 }
1046 
1047 /*
1048  * Called from:
1049  *  thread_exit()
1050  */
1051 void
1052 thread_unlink(struct thread *td)
1053 {
1054 	struct proc *p = td->td_proc;
1055 
1056 	PROC_LOCK_ASSERT(p, MA_OWNED);
1057 #ifdef EPOCH_TRACE
1058 	MPASS(SLIST_EMPTY(&td->td_epochs));
1059 #endif
1060 
1061 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
1062 	p->p_numthreads--;
1063 	/* could clear a few other things here */
1064 	/* Must  NOT clear links to proc! */
1065 }
1066 
1067 static int
1068 calc_remaining(struct proc *p, int mode)
1069 {
1070 	int remaining;
1071 
1072 	PROC_LOCK_ASSERT(p, MA_OWNED);
1073 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1074 	if (mode == SINGLE_EXIT)
1075 		remaining = p->p_numthreads;
1076 	else if (mode == SINGLE_BOUNDARY)
1077 		remaining = p->p_numthreads - p->p_boundary_count;
1078 	else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
1079 		remaining = p->p_numthreads - p->p_suspcount;
1080 	else
1081 		panic("calc_remaining: wrong mode %d", mode);
1082 	return (remaining);
1083 }
1084 
1085 static int
1086 remain_for_mode(int mode)
1087 {
1088 
1089 	return (mode == SINGLE_ALLPROC ? 0 : 1);
1090 }
1091 
1092 static int
1093 weed_inhib(int mode, struct thread *td2, struct proc *p)
1094 {
1095 	int wakeup_swapper;
1096 
1097 	PROC_LOCK_ASSERT(p, MA_OWNED);
1098 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1099 	THREAD_LOCK_ASSERT(td2, MA_OWNED);
1100 
1101 	wakeup_swapper = 0;
1102 
1103 	/*
1104 	 * Since the thread lock is dropped by the scheduler we have
1105 	 * to retry to check for races.
1106 	 */
1107 restart:
1108 	switch (mode) {
1109 	case SINGLE_EXIT:
1110 		if (TD_IS_SUSPENDED(td2)) {
1111 			wakeup_swapper |= thread_unsuspend_one(td2, p, true);
1112 			thread_lock(td2);
1113 			goto restart;
1114 		}
1115 		if (TD_CAN_ABORT(td2)) {
1116 			wakeup_swapper |= sleepq_abort(td2, EINTR);
1117 			return (wakeup_swapper);
1118 		}
1119 		break;
1120 	case SINGLE_BOUNDARY:
1121 	case SINGLE_NO_EXIT:
1122 		if (TD_IS_SUSPENDED(td2) &&
1123 		    (td2->td_flags & TDF_BOUNDARY) == 0) {
1124 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1125 			thread_lock(td2);
1126 			goto restart;
1127 		}
1128 		if (TD_CAN_ABORT(td2)) {
1129 			wakeup_swapper |= sleepq_abort(td2, ERESTART);
1130 			return (wakeup_swapper);
1131 		}
1132 		break;
1133 	case SINGLE_ALLPROC:
1134 		/*
1135 		 * ALLPROC suspend tries to avoid spurious EINTR for
1136 		 * threads sleeping interruptable, by suspending the
1137 		 * thread directly, similarly to sig_suspend_threads().
1138 		 * Since such sleep is not performed at the user
1139 		 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
1140 		 * is used to avoid immediate un-suspend.
1141 		 */
1142 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
1143 		    TDF_ALLPROCSUSP)) == 0) {
1144 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1145 			thread_lock(td2);
1146 			goto restart;
1147 		}
1148 		if (TD_CAN_ABORT(td2)) {
1149 			if ((td2->td_flags & TDF_SBDRY) == 0) {
1150 				thread_suspend_one(td2);
1151 				td2->td_flags |= TDF_ALLPROCSUSP;
1152 			} else {
1153 				wakeup_swapper |= sleepq_abort(td2, ERESTART);
1154 				return (wakeup_swapper);
1155 			}
1156 		}
1157 		break;
1158 	default:
1159 		break;
1160 	}
1161 	thread_unlock(td2);
1162 	return (wakeup_swapper);
1163 }
1164 
1165 /*
1166  * Enforce single-threading.
1167  *
1168  * Returns 1 if the caller must abort (another thread is waiting to
1169  * exit the process or similar). Process is locked!
1170  * Returns 0 when you are successfully the only thread running.
1171  * A process has successfully single threaded in the suspend mode when
1172  * There are no threads in user mode. Threads in the kernel must be
1173  * allowed to continue until they get to the user boundary. They may even
1174  * copy out their return values and data before suspending. They may however be
1175  * accelerated in reaching the user boundary as we will wake up
1176  * any sleeping threads that are interruptable. (PCATCH).
1177  */
1178 int
1179 thread_single(struct proc *p, int mode)
1180 {
1181 	struct thread *td;
1182 	struct thread *td2;
1183 	int remaining, wakeup_swapper;
1184 
1185 	td = curthread;
1186 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1187 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1188 	    ("invalid mode %d", mode));
1189 	/*
1190 	 * If allowing non-ALLPROC singlethreading for non-curproc
1191 	 * callers, calc_remaining() and remain_for_mode() should be
1192 	 * adjusted to also account for td->td_proc != p.  For now
1193 	 * this is not implemented because it is not used.
1194 	 */
1195 	KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
1196 	    (mode != SINGLE_ALLPROC && td->td_proc == p),
1197 	    ("mode %d proc %p curproc %p", mode, p, td->td_proc));
1198 	mtx_assert(&Giant, MA_NOTOWNED);
1199 	PROC_LOCK_ASSERT(p, MA_OWNED);
1200 
1201 	if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
1202 		return (0);
1203 
1204 	/* Is someone already single threading? */
1205 	if (p->p_singlethread != NULL && p->p_singlethread != td)
1206 		return (1);
1207 
1208 	if (mode == SINGLE_EXIT) {
1209 		p->p_flag |= P_SINGLE_EXIT;
1210 		p->p_flag &= ~P_SINGLE_BOUNDARY;
1211 	} else {
1212 		p->p_flag &= ~P_SINGLE_EXIT;
1213 		if (mode == SINGLE_BOUNDARY)
1214 			p->p_flag |= P_SINGLE_BOUNDARY;
1215 		else
1216 			p->p_flag &= ~P_SINGLE_BOUNDARY;
1217 	}
1218 	if (mode == SINGLE_ALLPROC)
1219 		p->p_flag |= P_TOTAL_STOP;
1220 	p->p_flag |= P_STOPPED_SINGLE;
1221 	PROC_SLOCK(p);
1222 	p->p_singlethread = td;
1223 	remaining = calc_remaining(p, mode);
1224 	while (remaining != remain_for_mode(mode)) {
1225 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
1226 			goto stopme;
1227 		wakeup_swapper = 0;
1228 		FOREACH_THREAD_IN_PROC(p, td2) {
1229 			if (td2 == td)
1230 				continue;
1231 			thread_lock(td2);
1232 			td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
1233 			if (TD_IS_INHIBITED(td2)) {
1234 				wakeup_swapper |= weed_inhib(mode, td2, p);
1235 #ifdef SMP
1236 			} else if (TD_IS_RUNNING(td2) && td != td2) {
1237 				forward_signal(td2);
1238 				thread_unlock(td2);
1239 #endif
1240 			} else
1241 				thread_unlock(td2);
1242 		}
1243 		if (wakeup_swapper)
1244 			kick_proc0();
1245 		remaining = calc_remaining(p, mode);
1246 
1247 		/*
1248 		 * Maybe we suspended some threads.. was it enough?
1249 		 */
1250 		if (remaining == remain_for_mode(mode))
1251 			break;
1252 
1253 stopme:
1254 		/*
1255 		 * Wake us up when everyone else has suspended.
1256 		 * In the mean time we suspend as well.
1257 		 */
1258 		thread_suspend_switch(td, p);
1259 		remaining = calc_remaining(p, mode);
1260 	}
1261 	if (mode == SINGLE_EXIT) {
1262 		/*
1263 		 * Convert the process to an unthreaded process.  The
1264 		 * SINGLE_EXIT is called by exit1() or execve(), in
1265 		 * both cases other threads must be retired.
1266 		 */
1267 		KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1268 		p->p_singlethread = NULL;
1269 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1270 
1271 		/*
1272 		 * Wait for any remaining threads to exit cpu_throw().
1273 		 */
1274 		while (p->p_exitthreads != 0) {
1275 			PROC_SUNLOCK(p);
1276 			PROC_UNLOCK(p);
1277 			sched_relinquish(td);
1278 			PROC_LOCK(p);
1279 			PROC_SLOCK(p);
1280 		}
1281 	} else if (mode == SINGLE_BOUNDARY) {
1282 		/*
1283 		 * Wait until all suspended threads are removed from
1284 		 * the processors.  The thread_suspend_check()
1285 		 * increments p_boundary_count while it is still
1286 		 * running, which makes it possible for the execve()
1287 		 * to destroy vmspace while our other threads are
1288 		 * still using the address space.
1289 		 *
1290 		 * We lock the thread, which is only allowed to
1291 		 * succeed after context switch code finished using
1292 		 * the address space.
1293 		 */
1294 		FOREACH_THREAD_IN_PROC(p, td2) {
1295 			if (td2 == td)
1296 				continue;
1297 			thread_lock(td2);
1298 			KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1299 			    ("td %p not on boundary", td2));
1300 			KASSERT(TD_IS_SUSPENDED(td2),
1301 			    ("td %p is not suspended", td2));
1302 			thread_unlock(td2);
1303 		}
1304 	}
1305 	PROC_SUNLOCK(p);
1306 	return (0);
1307 }
1308 
1309 bool
1310 thread_suspend_check_needed(void)
1311 {
1312 	struct proc *p;
1313 	struct thread *td;
1314 
1315 	td = curthread;
1316 	p = td->td_proc;
1317 	PROC_LOCK_ASSERT(p, MA_OWNED);
1318 	return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1319 	    (td->td_dbgflags & TDB_SUSPEND) != 0));
1320 }
1321 
1322 /*
1323  * Called in from locations that can safely check to see
1324  * whether we have to suspend or at least throttle for a
1325  * single-thread event (e.g. fork).
1326  *
1327  * Such locations include userret().
1328  * If the "return_instead" argument is non zero, the thread must be able to
1329  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1330  *
1331  * The 'return_instead' argument tells the function if it may do a
1332  * thread_exit() or suspend, or whether the caller must abort and back
1333  * out instead.
1334  *
1335  * If the thread that set the single_threading request has set the
1336  * P_SINGLE_EXIT bit in the process flags then this call will never return
1337  * if 'return_instead' is false, but will exit.
1338  *
1339  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1340  *---------------+--------------------+---------------------
1341  *       0       | returns 0          |   returns 0 or 1
1342  *               | when ST ends       |   immediately
1343  *---------------+--------------------+---------------------
1344  *       1       | thread exits       |   returns 1
1345  *               |                    |  immediately
1346  * 0 = thread_exit() or suspension ok,
1347  * other = return error instead of stopping the thread.
1348  *
1349  * While a full suspension is under effect, even a single threading
1350  * thread would be suspended if it made this call (but it shouldn't).
1351  * This call should only be made from places where
1352  * thread_exit() would be safe as that may be the outcome unless
1353  * return_instead is set.
1354  */
1355 int
1356 thread_suspend_check(int return_instead)
1357 {
1358 	struct thread *td;
1359 	struct proc *p;
1360 	int wakeup_swapper;
1361 
1362 	td = curthread;
1363 	p = td->td_proc;
1364 	mtx_assert(&Giant, MA_NOTOWNED);
1365 	PROC_LOCK_ASSERT(p, MA_OWNED);
1366 	while (thread_suspend_check_needed()) {
1367 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1368 			KASSERT(p->p_singlethread != NULL,
1369 			    ("singlethread not set"));
1370 			/*
1371 			 * The only suspension in action is a
1372 			 * single-threading. Single threader need not stop.
1373 			 * It is safe to access p->p_singlethread unlocked
1374 			 * because it can only be set to our address by us.
1375 			 */
1376 			if (p->p_singlethread == td)
1377 				return (0);	/* Exempt from stopping. */
1378 		}
1379 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1380 			return (EINTR);
1381 
1382 		/* Should we goto user boundary if we didn't come from there? */
1383 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1384 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1385 			return (ERESTART);
1386 
1387 		/*
1388 		 * Ignore suspend requests if they are deferred.
1389 		 */
1390 		if ((td->td_flags & TDF_SBDRY) != 0) {
1391 			KASSERT(return_instead,
1392 			    ("TDF_SBDRY set for unsafe thread_suspend_check"));
1393 			KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1394 			    (TDF_SEINTR | TDF_SERESTART),
1395 			    ("both TDF_SEINTR and TDF_SERESTART"));
1396 			return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1397 		}
1398 
1399 		/*
1400 		 * If the process is waiting for us to exit,
1401 		 * this thread should just suicide.
1402 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1403 		 */
1404 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1405 			PROC_UNLOCK(p);
1406 
1407 			/*
1408 			 * Allow Linux emulation layer to do some work
1409 			 * before thread suicide.
1410 			 */
1411 			if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1412 				(p->p_sysent->sv_thread_detach)(td);
1413 			umtx_thread_exit(td);
1414 			kern_thr_exit(td);
1415 			panic("stopped thread did not exit");
1416 		}
1417 
1418 		PROC_SLOCK(p);
1419 		thread_stopped(p);
1420 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1421 			if (p->p_numthreads == p->p_suspcount + 1) {
1422 				thread_lock(p->p_singlethread);
1423 				wakeup_swapper = thread_unsuspend_one(
1424 				    p->p_singlethread, p, false);
1425 				if (wakeup_swapper)
1426 					kick_proc0();
1427 			}
1428 		}
1429 		PROC_UNLOCK(p);
1430 		thread_lock(td);
1431 		/*
1432 		 * When a thread suspends, it just
1433 		 * gets taken off all queues.
1434 		 */
1435 		thread_suspend_one(td);
1436 		if (return_instead == 0) {
1437 			p->p_boundary_count++;
1438 			td->td_flags |= TDF_BOUNDARY;
1439 		}
1440 		PROC_SUNLOCK(p);
1441 		mi_switch(SW_INVOL | SWT_SUSPEND);
1442 		PROC_LOCK(p);
1443 	}
1444 	return (0);
1445 }
1446 
1447 /*
1448  * Check for possible stops and suspensions while executing a
1449  * casueword or similar transiently failing operation.
1450  *
1451  * The sleep argument controls whether the function can handle a stop
1452  * request itself or it should return ERESTART and the request is
1453  * proceed at the kernel/user boundary in ast.
1454  *
1455  * Typically, when retrying due to casueword(9) failure (rv == 1), we
1456  * should handle the stop requests there, with exception of cases when
1457  * the thread owns a kernel resource, for instance busied the umtx
1458  * key, or when functions return immediately if thread_check_susp()
1459  * returned non-zero.  On the other hand, retrying the whole lock
1460  * operation, we better not stop there but delegate the handling to
1461  * ast.
1462  *
1463  * If the request is for thread termination P_SINGLE_EXIT, we cannot
1464  * handle it at all, and simply return EINTR.
1465  */
1466 int
1467 thread_check_susp(struct thread *td, bool sleep)
1468 {
1469 	struct proc *p;
1470 	int error;
1471 
1472 	/*
1473 	 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
1474 	 * eventually break the lockstep loop.
1475 	 */
1476 	if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
1477 		return (0);
1478 	error = 0;
1479 	p = td->td_proc;
1480 	PROC_LOCK(p);
1481 	if (p->p_flag & P_SINGLE_EXIT)
1482 		error = EINTR;
1483 	else if (P_SHOULDSTOP(p) ||
1484 	    ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1485 		error = sleep ? thread_suspend_check(0) : ERESTART;
1486 	PROC_UNLOCK(p);
1487 	return (error);
1488 }
1489 
1490 void
1491 thread_suspend_switch(struct thread *td, struct proc *p)
1492 {
1493 
1494 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1495 	PROC_LOCK_ASSERT(p, MA_OWNED);
1496 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1497 	/*
1498 	 * We implement thread_suspend_one in stages here to avoid
1499 	 * dropping the proc lock while the thread lock is owned.
1500 	 */
1501 	if (p == td->td_proc) {
1502 		thread_stopped(p);
1503 		p->p_suspcount++;
1504 	}
1505 	PROC_UNLOCK(p);
1506 	thread_lock(td);
1507 	td->td_flags &= ~TDF_NEEDSUSPCHK;
1508 	TD_SET_SUSPENDED(td);
1509 	sched_sleep(td, 0);
1510 	PROC_SUNLOCK(p);
1511 	DROP_GIANT();
1512 	mi_switch(SW_VOL | SWT_SUSPEND);
1513 	PICKUP_GIANT();
1514 	PROC_LOCK(p);
1515 	PROC_SLOCK(p);
1516 }
1517 
1518 void
1519 thread_suspend_one(struct thread *td)
1520 {
1521 	struct proc *p;
1522 
1523 	p = td->td_proc;
1524 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1525 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1526 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1527 	p->p_suspcount++;
1528 	td->td_flags &= ~TDF_NEEDSUSPCHK;
1529 	TD_SET_SUSPENDED(td);
1530 	sched_sleep(td, 0);
1531 }
1532 
1533 static int
1534 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1535 {
1536 
1537 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1538 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1539 	TD_CLR_SUSPENDED(td);
1540 	td->td_flags &= ~TDF_ALLPROCSUSP;
1541 	if (td->td_proc == p) {
1542 		PROC_SLOCK_ASSERT(p, MA_OWNED);
1543 		p->p_suspcount--;
1544 		if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1545 			td->td_flags &= ~TDF_BOUNDARY;
1546 			p->p_boundary_count--;
1547 		}
1548 	}
1549 	return (setrunnable(td, 0));
1550 }
1551 
1552 void
1553 thread_run_flash(struct thread *td)
1554 {
1555 	struct proc *p;
1556 
1557 	p = td->td_proc;
1558 	PROC_LOCK_ASSERT(p, MA_OWNED);
1559 
1560 	if (TD_ON_SLEEPQ(td))
1561 		sleepq_remove_nested(td);
1562 	else
1563 		thread_lock(td);
1564 
1565 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1566 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1567 
1568 	TD_CLR_SUSPENDED(td);
1569 	PROC_SLOCK(p);
1570 	MPASS(p->p_suspcount > 0);
1571 	p->p_suspcount--;
1572 	PROC_SUNLOCK(p);
1573 	if (setrunnable(td, 0))
1574 		kick_proc0();
1575 }
1576 
1577 /*
1578  * Allow all threads blocked by single threading to continue running.
1579  */
1580 void
1581 thread_unsuspend(struct proc *p)
1582 {
1583 	struct thread *td;
1584 	int wakeup_swapper;
1585 
1586 	PROC_LOCK_ASSERT(p, MA_OWNED);
1587 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1588 	wakeup_swapper = 0;
1589 	if (!P_SHOULDSTOP(p)) {
1590                 FOREACH_THREAD_IN_PROC(p, td) {
1591 			thread_lock(td);
1592 			if (TD_IS_SUSPENDED(td)) {
1593 				wakeup_swapper |= thread_unsuspend_one(td, p,
1594 				    true);
1595 			} else
1596 				thread_unlock(td);
1597 		}
1598 	} else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1599 	    p->p_numthreads == p->p_suspcount) {
1600 		/*
1601 		 * Stopping everything also did the job for the single
1602 		 * threading request. Now we've downgraded to single-threaded,
1603 		 * let it continue.
1604 		 */
1605 		if (p->p_singlethread->td_proc == p) {
1606 			thread_lock(p->p_singlethread);
1607 			wakeup_swapper = thread_unsuspend_one(
1608 			    p->p_singlethread, p, false);
1609 		}
1610 	}
1611 	if (wakeup_swapper)
1612 		kick_proc0();
1613 }
1614 
1615 /*
1616  * End the single threading mode..
1617  */
1618 void
1619 thread_single_end(struct proc *p, int mode)
1620 {
1621 	struct thread *td;
1622 	int wakeup_swapper;
1623 
1624 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1625 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1626 	    ("invalid mode %d", mode));
1627 	PROC_LOCK_ASSERT(p, MA_OWNED);
1628 	KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1629 	    (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1630 	    ("mode %d does not match P_TOTAL_STOP", mode));
1631 	KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1632 	    ("thread_single_end from other thread %p %p",
1633 	    curthread, p->p_singlethread));
1634 	KASSERT(mode != SINGLE_BOUNDARY ||
1635 	    (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1636 	    ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1637 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1638 	    P_TOTAL_STOP);
1639 	PROC_SLOCK(p);
1640 	p->p_singlethread = NULL;
1641 	wakeup_swapper = 0;
1642 	/*
1643 	 * If there are other threads they may now run,
1644 	 * unless of course there is a blanket 'stop order'
1645 	 * on the process. The single threader must be allowed
1646 	 * to continue however as this is a bad place to stop.
1647 	 */
1648 	if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1649                 FOREACH_THREAD_IN_PROC(p, td) {
1650 			thread_lock(td);
1651 			if (TD_IS_SUSPENDED(td)) {
1652 				wakeup_swapper |= thread_unsuspend_one(td, p,
1653 				    mode == SINGLE_BOUNDARY);
1654 			} else
1655 				thread_unlock(td);
1656 		}
1657 	}
1658 	KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1659 	    ("inconsistent boundary count %d", p->p_boundary_count));
1660 	PROC_SUNLOCK(p);
1661 	if (wakeup_swapper)
1662 		kick_proc0();
1663 }
1664 
1665 /*
1666  * Locate a thread by number and return with proc lock held.
1667  *
1668  * thread exit establishes proc -> tidhash lock ordering, but lookup
1669  * takes tidhash first and needs to return locked proc.
1670  *
1671  * The problem is worked around by relying on type-safety of both
1672  * structures and doing the work in 2 steps:
1673  * - tidhash-locked lookup which saves both thread and proc pointers
1674  * - proc-locked verification that the found thread still matches
1675  */
1676 static bool
1677 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1678 {
1679 #define RUN_THRESH	16
1680 	struct proc *p;
1681 	struct thread *td;
1682 	int run;
1683 	bool locked;
1684 
1685 	run = 0;
1686 	rw_rlock(TIDHASHLOCK(tid));
1687 	locked = true;
1688 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1689 		if (td->td_tid != tid) {
1690 			run++;
1691 			continue;
1692 		}
1693 		p = td->td_proc;
1694 		if (pid != -1 && p->p_pid != pid) {
1695 			td = NULL;
1696 			break;
1697 		}
1698 		if (run > RUN_THRESH) {
1699 			if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1700 				LIST_REMOVE(td, td_hash);
1701 				LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1702 					td, td_hash);
1703 				rw_wunlock(TIDHASHLOCK(tid));
1704 				locked = false;
1705 				break;
1706 			}
1707 		}
1708 		break;
1709 	}
1710 	if (locked)
1711 		rw_runlock(TIDHASHLOCK(tid));
1712 	if (td == NULL)
1713 		return (false);
1714 	*pp = p;
1715 	*tdp = td;
1716 	return (true);
1717 }
1718 
1719 struct thread *
1720 tdfind(lwpid_t tid, pid_t pid)
1721 {
1722 	struct proc *p;
1723 	struct thread *td;
1724 
1725 	td = curthread;
1726 	if (td->td_tid == tid) {
1727 		if (pid != -1 && td->td_proc->p_pid != pid)
1728 			return (NULL);
1729 		PROC_LOCK(td->td_proc);
1730 		return (td);
1731 	}
1732 
1733 	for (;;) {
1734 		if (!tdfind_hash(tid, pid, &p, &td))
1735 			return (NULL);
1736 		PROC_LOCK(p);
1737 		if (td->td_tid != tid) {
1738 			PROC_UNLOCK(p);
1739 			continue;
1740 		}
1741 		if (td->td_proc != p) {
1742 			PROC_UNLOCK(p);
1743 			continue;
1744 		}
1745 		if (p->p_state == PRS_NEW) {
1746 			PROC_UNLOCK(p);
1747 			return (NULL);
1748 		}
1749 		return (td);
1750 	}
1751 }
1752 
1753 void
1754 tidhash_add(struct thread *td)
1755 {
1756 	rw_wlock(TIDHASHLOCK(td->td_tid));
1757 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1758 	rw_wunlock(TIDHASHLOCK(td->td_tid));
1759 }
1760 
1761 void
1762 tidhash_remove(struct thread *td)
1763 {
1764 
1765 	rw_wlock(TIDHASHLOCK(td->td_tid));
1766 	LIST_REMOVE(td, td_hash);
1767 	rw_wunlock(TIDHASHLOCK(td->td_tid));
1768 }
1769