xref: /freebsd/sys/kern/kern_thread.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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/mutex.h>
42 #include <sys/proc.h>
43 #include <sys/bitstring.h>
44 #include <sys/epoch.h>
45 #include <sys/rangelock.h>
46 #include <sys/resourcevar.h>
47 #include <sys/sdt.h>
48 #include <sys/smp.h>
49 #include <sys/sched.h>
50 #include <sys/sleepqueue.h>
51 #include <sys/selinfo.h>
52 #include <sys/syscallsubr.h>
53 #include <sys/dtrace_bsd.h>
54 #include <sys/sysent.h>
55 #include <sys/turnstile.h>
56 #include <sys/taskqueue.h>
57 #include <sys/ktr.h>
58 #include <sys/rwlock.h>
59 #include <sys/umtx.h>
60 #include <sys/vmmeter.h>
61 #include <sys/cpuset.h>
62 #ifdef	HWPMC_HOOKS
63 #include <sys/pmckern.h>
64 #endif
65 #include <sys/priv.h>
66 
67 #include <security/audit/audit.h>
68 
69 #include <vm/pmap.h>
70 #include <vm/vm.h>
71 #include <vm/vm_extern.h>
72 #include <vm/uma.h>
73 #include <vm/vm_phys.h>
74 #include <sys/eventhandler.h>
75 
76 /*
77  * Asserts below verify the stability of struct thread and struct proc
78  * layout, as exposed by KBI to modules.  On head, the KBI is allowed
79  * to drift, change to the structures must be accompanied by the
80  * assert update.
81  *
82  * On the stable branches after KBI freeze, conditions must not be
83  * violated.  Typically new fields are moved to the end of the
84  * structures.
85  */
86 #ifdef __amd64__
87 _Static_assert(offsetof(struct thread, td_flags) == 0xfc,
88     "struct thread KBI td_flags");
89 _Static_assert(offsetof(struct thread, td_pflags) == 0x104,
90     "struct thread KBI td_pflags");
91 _Static_assert(offsetof(struct thread, td_frame) == 0x4a0,
92     "struct thread KBI td_frame");
93 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6b0,
94     "struct thread KBI td_emuldata");
95 _Static_assert(offsetof(struct proc, p_flag) == 0xb8,
96     "struct proc KBI p_flag");
97 _Static_assert(offsetof(struct proc, p_pid) == 0xc4,
98     "struct proc KBI p_pid");
99 _Static_assert(offsetof(struct proc, p_filemon) == 0x3b8,
100     "struct proc KBI p_filemon");
101 _Static_assert(offsetof(struct proc, p_comm) == 0x3d0,
102     "struct proc KBI p_comm");
103 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4b0,
104     "struct proc KBI p_emuldata");
105 #endif
106 #ifdef __i386__
107 _Static_assert(offsetof(struct thread, td_flags) == 0x98,
108     "struct thread KBI td_flags");
109 _Static_assert(offsetof(struct thread, td_pflags) == 0xa0,
110     "struct thread KBI td_pflags");
111 _Static_assert(offsetof(struct thread, td_frame) == 0x300,
112     "struct thread KBI td_frame");
113 _Static_assert(offsetof(struct thread, td_emuldata) == 0x344,
114     "struct thread KBI td_emuldata");
115 _Static_assert(offsetof(struct proc, p_flag) == 0x6c,
116     "struct proc KBI p_flag");
117 _Static_assert(offsetof(struct proc, p_pid) == 0x78,
118     "struct proc KBI p_pid");
119 _Static_assert(offsetof(struct proc, p_filemon) == 0x268,
120     "struct proc KBI p_filemon");
121 _Static_assert(offsetof(struct proc, p_comm) == 0x27c,
122     "struct proc KBI p_comm");
123 _Static_assert(offsetof(struct proc, p_emuldata) == 0x308,
124     "struct proc KBI p_emuldata");
125 #endif
126 
127 SDT_PROVIDER_DECLARE(proc);
128 SDT_PROBE_DEFINE(proc, , , lwp__exit);
129 
130 /*
131  * thread related storage.
132  */
133 static uma_zone_t thread_zone;
134 
135 struct thread_domain_data {
136 	struct thread	*tdd_zombies;
137 	int		tdd_reapticks;
138 } __aligned(CACHE_LINE_SIZE);
139 
140 static struct thread_domain_data thread_domain_data[MAXMEMDOM];
141 
142 static struct task	thread_reap_task;
143 static struct callout  	thread_reap_callout;
144 
145 static void thread_zombie(struct thread *);
146 static void thread_reap(void);
147 static void thread_reap_all(void);
148 static void thread_reap_task_cb(void *, int);
149 static void thread_reap_callout_cb(void *);
150 static int thread_unsuspend_one(struct thread *td, struct proc *p,
151     bool boundary);
152 static void thread_free_batched(struct thread *td);
153 
154 static __exclusive_cache_line struct mtx tid_lock;
155 static bitstr_t *tid_bitmap;
156 
157 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
158 
159 static int maxthread;
160 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
161     &maxthread, 0, "Maximum number of threads");
162 
163 static __exclusive_cache_line int nthreads;
164 
165 static LIST_HEAD(tidhashhead, thread) *tidhashtbl;
166 static u_long	tidhash;
167 static u_long	tidhashlock;
168 static struct	rwlock *tidhashtbl_lock;
169 #define	TIDHASH(tid)		(&tidhashtbl[(tid) & tidhash])
170 #define	TIDHASHLOCK(tid)	(&tidhashtbl_lock[(tid) & tidhashlock])
171 
172 EVENTHANDLER_LIST_DEFINE(thread_ctor);
173 EVENTHANDLER_LIST_DEFINE(thread_dtor);
174 EVENTHANDLER_LIST_DEFINE(thread_init);
175 EVENTHANDLER_LIST_DEFINE(thread_fini);
176 
177 static bool
178 thread_count_inc_try(void)
179 {
180 	int nthreads_new;
181 
182 	nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1;
183 	if (nthreads_new >= maxthread - 100) {
184 		if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
185 		    nthreads_new >= maxthread) {
186 			atomic_subtract_int(&nthreads, 1);
187 			return (false);
188 		}
189 	}
190 	return (true);
191 }
192 
193 static bool
194 thread_count_inc(void)
195 {
196 	static struct timeval lastfail;
197 	static int curfail;
198 
199 	thread_reap();
200 	if (thread_count_inc_try()) {
201 		return (true);
202 	}
203 
204 	thread_reap_all();
205 	if (thread_count_inc_try()) {
206 		return (true);
207 	}
208 
209 	if (ppsratecheck(&lastfail, &curfail, 1)) {
210 		printf("maxthread limit exceeded by uid %u "
211 		    "(pid %d); consider increasing kern.maxthread\n",
212 		    curthread->td_ucred->cr_ruid, curproc->p_pid);
213 	}
214 	return (false);
215 }
216 
217 static void
218 thread_count_sub(int n)
219 {
220 
221 	atomic_subtract_int(&nthreads, n);
222 }
223 
224 static void
225 thread_count_dec(void)
226 {
227 
228 	thread_count_sub(1);
229 }
230 
231 static lwpid_t
232 tid_alloc(void)
233 {
234 	static lwpid_t trytid;
235 	lwpid_t tid;
236 
237 	mtx_lock(&tid_lock);
238 	/*
239 	 * It is an invariant that the bitmap is big enough to hold maxthread
240 	 * IDs. If we got to this point there has to be at least one free.
241 	 */
242 	if (trytid >= maxthread)
243 		trytid = 0;
244 	bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
245 	if (tid == -1) {
246 		KASSERT(trytid != 0, ("unexpectedly ran out of IDs"));
247 		trytid = 0;
248 		bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
249 		KASSERT(tid != -1, ("unexpectedly ran out of IDs"));
250 	}
251 	bit_set(tid_bitmap, tid);
252 	trytid = tid + 1;
253 	mtx_unlock(&tid_lock);
254 	return (tid + NO_PID);
255 }
256 
257 static void
258 tid_free_locked(lwpid_t rtid)
259 {
260 	lwpid_t tid;
261 
262 	mtx_assert(&tid_lock, MA_OWNED);
263 	KASSERT(rtid >= NO_PID,
264 	    ("%s: invalid tid %d\n", __func__, rtid));
265 	tid = rtid - NO_PID;
266 	KASSERT(bit_test(tid_bitmap, tid) != 0,
267 	    ("thread ID %d not allocated\n", rtid));
268 	bit_clear(tid_bitmap, tid);
269 }
270 
271 static void
272 tid_free(lwpid_t rtid)
273 {
274 
275 	mtx_lock(&tid_lock);
276 	tid_free_locked(rtid);
277 	mtx_unlock(&tid_lock);
278 }
279 
280 static void
281 tid_free_batch(lwpid_t *batch, int n)
282 {
283 	int i;
284 
285 	mtx_lock(&tid_lock);
286 	for (i = 0; i < n; i++) {
287 		tid_free_locked(batch[i]);
288 	}
289 	mtx_unlock(&tid_lock);
290 }
291 
292 /*
293  * Batching for thread reapping.
294  */
295 struct tidbatch {
296 	lwpid_t tab[16];
297 	int n;
298 };
299 
300 static void
301 tidbatch_prep(struct tidbatch *tb)
302 {
303 
304 	tb->n = 0;
305 }
306 
307 static void
308 tidbatch_add(struct tidbatch *tb, struct thread *td)
309 {
310 
311 	KASSERT(tb->n < nitems(tb->tab),
312 	    ("%s: count too high %d", __func__, tb->n));
313 	tb->tab[tb->n] = td->td_tid;
314 	tb->n++;
315 }
316 
317 static void
318 tidbatch_process(struct tidbatch *tb)
319 {
320 
321 	KASSERT(tb->n <= nitems(tb->tab),
322 	    ("%s: count too high %d", __func__, tb->n));
323 	if (tb->n == nitems(tb->tab)) {
324 		tid_free_batch(tb->tab, tb->n);
325 		tb->n = 0;
326 	}
327 }
328 
329 static void
330 tidbatch_final(struct tidbatch *tb)
331 {
332 
333 	KASSERT(tb->n <= nitems(tb->tab),
334 	    ("%s: count too high %d", __func__, tb->n));
335 	if (tb->n != 0) {
336 		tid_free_batch(tb->tab, tb->n);
337 	}
338 }
339 
340 /*
341  * Prepare a thread for use.
342  */
343 static int
344 thread_ctor(void *mem, int size, void *arg, int flags)
345 {
346 	struct thread	*td;
347 
348 	td = (struct thread *)mem;
349 	TD_SET_STATE(td, TDS_INACTIVE);
350 	td->td_lastcpu = td->td_oncpu = NOCPU;
351 
352 	/*
353 	 * Note that td_critnest begins life as 1 because the thread is not
354 	 * running and is thereby implicitly waiting to be on the receiving
355 	 * end of a context switch.
356 	 */
357 	td->td_critnest = 1;
358 	td->td_lend_user_pri = PRI_MAX;
359 #ifdef AUDIT
360 	audit_thread_alloc(td);
361 #endif
362 #ifdef KDTRACE_HOOKS
363 	kdtrace_thread_ctor(td);
364 #endif
365 	umtx_thread_alloc(td);
366 	MPASS(td->td_sel == NULL);
367 	return (0);
368 }
369 
370 /*
371  * Reclaim a thread after use.
372  */
373 static void
374 thread_dtor(void *mem, int size, void *arg)
375 {
376 	struct thread *td;
377 
378 	td = (struct thread *)mem;
379 
380 #ifdef INVARIANTS
381 	/* Verify that this thread is in a safe state to free. */
382 	switch (TD_GET_STATE(td)) {
383 	case TDS_INHIBITED:
384 	case TDS_RUNNING:
385 	case TDS_CAN_RUN:
386 	case TDS_RUNQ:
387 		/*
388 		 * We must never unlink a thread that is in one of
389 		 * these states, because it is currently active.
390 		 */
391 		panic("bad state for thread unlinking");
392 		/* NOTREACHED */
393 	case TDS_INACTIVE:
394 		break;
395 	default:
396 		panic("bad thread state");
397 		/* NOTREACHED */
398 	}
399 #endif
400 #ifdef AUDIT
401 	audit_thread_free(td);
402 #endif
403 #ifdef KDTRACE_HOOKS
404 	kdtrace_thread_dtor(td);
405 #endif
406 	/* Free all OSD associated to this thread. */
407 	osd_thread_exit(td);
408 	td_softdep_cleanup(td);
409 	MPASS(td->td_su == NULL);
410 	seltdfini(td);
411 }
412 
413 /*
414  * Initialize type-stable parts of a thread (when newly created).
415  */
416 static int
417 thread_init(void *mem, int size, int flags)
418 {
419 	struct thread *td;
420 
421 	td = (struct thread *)mem;
422 
423 	td->td_allocdomain = vm_phys_domain(vtophys(td));
424 	td->td_sleepqueue = sleepq_alloc();
425 	td->td_turnstile = turnstile_alloc();
426 	td->td_rlqe = NULL;
427 	EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
428 	umtx_thread_init(td);
429 	td->td_kstack = 0;
430 	td->td_sel = NULL;
431 	return (0);
432 }
433 
434 /*
435  * Tear down type-stable parts of a thread (just before being discarded).
436  */
437 static void
438 thread_fini(void *mem, int size)
439 {
440 	struct thread *td;
441 
442 	td = (struct thread *)mem;
443 	EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
444 	rlqentry_free(td->td_rlqe);
445 	turnstile_free(td->td_turnstile);
446 	sleepq_free(td->td_sleepqueue);
447 	umtx_thread_fini(td);
448 	MPASS(td->td_sel == NULL);
449 }
450 
451 /*
452  * For a newly created process,
453  * link up all the structures and its initial threads etc.
454  * called from:
455  * {arch}/{arch}/machdep.c   {arch}_init(), init386() etc.
456  * proc_dtor() (should go away)
457  * proc_init()
458  */
459 void
460 proc_linkup0(struct proc *p, struct thread *td)
461 {
462 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
463 	proc_linkup(p, td);
464 }
465 
466 void
467 proc_linkup(struct proc *p, struct thread *td)
468 {
469 
470 	sigqueue_init(&p->p_sigqueue, p);
471 	p->p_ksi = ksiginfo_alloc(1);
472 	if (p->p_ksi != NULL) {
473 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
474 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
475 	}
476 	LIST_INIT(&p->p_mqnotifier);
477 	p->p_numthreads = 0;
478 	thread_link(td, p);
479 }
480 
481 extern int max_threads_per_proc;
482 
483 /*
484  * Initialize global thread allocation resources.
485  */
486 void
487 threadinit(void)
488 {
489 	u_long i;
490 	lwpid_t tid0;
491 	uint32_t flags;
492 
493 	/*
494 	 * Place an upper limit on threads which can be allocated.
495 	 *
496 	 * Note that other factors may make the de facto limit much lower.
497 	 *
498 	 * Platform limits are somewhat arbitrary but deemed "more than good
499 	 * enough" for the foreseable future.
500 	 */
501 	if (maxthread == 0) {
502 #ifdef _LP64
503 		maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
504 #else
505 		maxthread = MIN(maxproc * max_threads_per_proc, 100000);
506 #endif
507 	}
508 
509 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
510 	tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
511 	/*
512 	 * Handle thread0.
513 	 */
514 	thread_count_inc();
515 	tid0 = tid_alloc();
516 	if (tid0 != THREAD0_TID)
517 		panic("tid0 %d != %d\n", tid0, THREAD0_TID);
518 
519 	flags = UMA_ZONE_NOFREE;
520 #ifdef __aarch64__
521 	/*
522 	 * Force thread structures to be allocated from the direct map.
523 	 * Otherwise, superpage promotions and demotions may temporarily
524 	 * invalidate thread structure mappings.  For most dynamically allocated
525 	 * structures this is not a problem, but translation faults cannot be
526 	 * handled without accessing curthread.
527 	 */
528 	flags |= UMA_ZONE_CONTIG;
529 #endif
530 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
531 	    thread_ctor, thread_dtor, thread_init, thread_fini,
532 	    32 - 1, flags);
533 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
534 	tidhashlock = (tidhash + 1) / 64;
535 	if (tidhashlock > 0)
536 		tidhashlock--;
537 	tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
538 	    M_TIDHASH, M_WAITOK | M_ZERO);
539 	for (i = 0; i < tidhashlock + 1; i++)
540 		rw_init(&tidhashtbl_lock[i], "tidhash");
541 
542 	TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL);
543 	callout_init(&thread_reap_callout, 1);
544 	callout_reset(&thread_reap_callout, 5 * hz,
545 	    thread_reap_callout_cb, NULL);
546 }
547 
548 /*
549  * Place an unused thread on the zombie list.
550  */
551 void
552 thread_zombie(struct thread *td)
553 {
554 	struct thread_domain_data *tdd;
555 	struct thread *ztd;
556 
557 	tdd = &thread_domain_data[td->td_allocdomain];
558 	ztd = atomic_load_ptr(&tdd->tdd_zombies);
559 	for (;;) {
560 		td->td_zombie = ztd;
561 		if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
562 		    (uintptr_t *)&ztd, (uintptr_t)td))
563 			break;
564 		continue;
565 	}
566 }
567 
568 /*
569  * Release a thread that has exited after cpu_throw().
570  */
571 void
572 thread_stash(struct thread *td)
573 {
574 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
575 	thread_zombie(td);
576 }
577 
578 /*
579  * Reap zombies from passed domain.
580  */
581 static void
582 thread_reap_domain(struct thread_domain_data *tdd)
583 {
584 	struct thread *itd, *ntd;
585 	struct tidbatch tidbatch;
586 	struct credbatch credbatch;
587 	int tdcount;
588 	struct plimit *lim;
589 	int limcount;
590 
591 	/*
592 	 * Reading upfront is pessimal if followed by concurrent atomic_swap,
593 	 * but most of the time the list is empty.
594 	 */
595 	if (tdd->tdd_zombies == NULL)
596 		return;
597 
598 	itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
599 	    (uintptr_t)NULL);
600 	if (itd == NULL)
601 		return;
602 
603 	/*
604 	 * Multiple CPUs can get here, the race is fine as ticks is only
605 	 * advisory.
606 	 */
607 	tdd->tdd_reapticks = ticks;
608 
609 	tidbatch_prep(&tidbatch);
610 	credbatch_prep(&credbatch);
611 	tdcount = 0;
612 	lim = NULL;
613 	limcount = 0;
614 
615 	while (itd != NULL) {
616 		ntd = itd->td_zombie;
617 		EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd);
618 		tidbatch_add(&tidbatch, itd);
619 		credbatch_add(&credbatch, itd);
620 		MPASS(itd->td_limit != NULL);
621 		if (lim != itd->td_limit) {
622 			if (limcount != 0) {
623 				lim_freen(lim, limcount);
624 				limcount = 0;
625 			}
626 		}
627 		lim = itd->td_limit;
628 		limcount++;
629 		thread_free_batched(itd);
630 		tidbatch_process(&tidbatch);
631 		credbatch_process(&credbatch);
632 		tdcount++;
633 		if (tdcount == 32) {
634 			thread_count_sub(tdcount);
635 			tdcount = 0;
636 		}
637 		itd = ntd;
638 	}
639 
640 	tidbatch_final(&tidbatch);
641 	credbatch_final(&credbatch);
642 	if (tdcount != 0) {
643 		thread_count_sub(tdcount);
644 	}
645 	MPASS(limcount != 0);
646 	lim_freen(lim, limcount);
647 }
648 
649 /*
650  * Reap zombies from all domains.
651  */
652 static void
653 thread_reap_all(void)
654 {
655 	struct thread_domain_data *tdd;
656 	int i, domain;
657 
658 	domain = PCPU_GET(domain);
659 	for (i = 0; i < vm_ndomains; i++) {
660 		tdd = &thread_domain_data[(i + domain) % vm_ndomains];
661 		thread_reap_domain(tdd);
662 	}
663 }
664 
665 /*
666  * Reap zombies from local domain.
667  */
668 static void
669 thread_reap(void)
670 {
671 	struct thread_domain_data *tdd;
672 	int domain;
673 
674 	domain = PCPU_GET(domain);
675 	tdd = &thread_domain_data[domain];
676 
677 	thread_reap_domain(tdd);
678 }
679 
680 static void
681 thread_reap_task_cb(void *arg __unused, int pending __unused)
682 {
683 
684 	thread_reap_all();
685 }
686 
687 static void
688 thread_reap_callout_cb(void *arg __unused)
689 {
690 	struct thread_domain_data *tdd;
691 	int i, cticks, lticks;
692 	bool wantreap;
693 
694 	wantreap = false;
695 	cticks = atomic_load_int(&ticks);
696 	for (i = 0; i < vm_ndomains; i++) {
697 		tdd = &thread_domain_data[i];
698 		lticks = tdd->tdd_reapticks;
699 		if (tdd->tdd_zombies != NULL &&
700 		    (u_int)(cticks - lticks) > 5 * hz) {
701 			wantreap = true;
702 			break;
703 		}
704 	}
705 
706 	if (wantreap)
707 		taskqueue_enqueue(taskqueue_thread, &thread_reap_task);
708 	callout_reset(&thread_reap_callout, 5 * hz,
709 	    thread_reap_callout_cb, NULL);
710 }
711 
712 /*
713  * Calling this function guarantees that any thread that exited before
714  * the call is reaped when the function returns.  By 'exited' we mean
715  * a thread removed from the process linkage with thread_unlink().
716  * Practically this means that caller must lock/unlock corresponding
717  * process lock before the call, to synchronize with thread_exit().
718  */
719 void
720 thread_reap_barrier(void)
721 {
722 	struct task *t;
723 
724 	/*
725 	 * First do context switches to each CPU to ensure that all
726 	 * PCPU pc_deadthreads are moved to zombie list.
727 	 */
728 	quiesce_all_cpus("", PDROP);
729 
730 	/*
731 	 * Second, fire the task in the same thread as normal
732 	 * thread_reap() is done, to serialize reaping.
733 	 */
734 	t = malloc(sizeof(*t), M_TEMP, M_WAITOK);
735 	TASK_INIT(t, 0, thread_reap_task_cb, t);
736 	taskqueue_enqueue(taskqueue_thread, t);
737 	taskqueue_drain(taskqueue_thread, t);
738 	free(t, M_TEMP);
739 }
740 
741 /*
742  * Allocate a thread.
743  */
744 struct thread *
745 thread_alloc(int pages)
746 {
747 	struct thread *td;
748 	lwpid_t tid;
749 
750 	if (!thread_count_inc()) {
751 		return (NULL);
752 	}
753 
754 	tid = tid_alloc();
755 	td = uma_zalloc(thread_zone, M_WAITOK);
756 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
757 	if (!vm_thread_new(td, pages)) {
758 		uma_zfree(thread_zone, td);
759 		tid_free(tid);
760 		thread_count_dec();
761 		return (NULL);
762 	}
763 	td->td_tid = tid;
764 	cpu_thread_alloc(td);
765 	EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
766 	return (td);
767 }
768 
769 int
770 thread_alloc_stack(struct thread *td, int pages)
771 {
772 
773 	KASSERT(td->td_kstack == 0,
774 	    ("thread_alloc_stack called on a thread with kstack"));
775 	if (!vm_thread_new(td, pages))
776 		return (0);
777 	cpu_thread_alloc(td);
778 	return (1);
779 }
780 
781 /*
782  * Deallocate a thread.
783  */
784 static void
785 thread_free_batched(struct thread *td)
786 {
787 
788 	lock_profile_thread_exit(td);
789 	if (td->td_cpuset)
790 		cpuset_rel(td->td_cpuset);
791 	td->td_cpuset = NULL;
792 	cpu_thread_free(td);
793 	if (td->td_kstack != 0)
794 		vm_thread_dispose(td);
795 	callout_drain(&td->td_slpcallout);
796 	/*
797 	 * Freeing handled by the caller.
798 	 */
799 	td->td_tid = -1;
800 	uma_zfree(thread_zone, td);
801 }
802 
803 void
804 thread_free(struct thread *td)
805 {
806 	lwpid_t tid;
807 
808 	EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
809 	tid = td->td_tid;
810 	thread_free_batched(td);
811 	tid_free(tid);
812 	thread_count_dec();
813 }
814 
815 void
816 thread_cow_get_proc(struct thread *newtd, struct proc *p)
817 {
818 
819 	PROC_LOCK_ASSERT(p, MA_OWNED);
820 	newtd->td_realucred = crcowget(p->p_ucred);
821 	newtd->td_ucred = newtd->td_realucred;
822 	newtd->td_limit = lim_hold(p->p_limit);
823 	newtd->td_cowgen = p->p_cowgen;
824 }
825 
826 void
827 thread_cow_get(struct thread *newtd, struct thread *td)
828 {
829 
830 	MPASS(td->td_realucred == td->td_ucred);
831 	newtd->td_realucred = crcowget(td->td_realucred);
832 	newtd->td_ucred = newtd->td_realucred;
833 	newtd->td_limit = lim_hold(td->td_limit);
834 	newtd->td_cowgen = td->td_cowgen;
835 }
836 
837 void
838 thread_cow_free(struct thread *td)
839 {
840 
841 	if (td->td_realucred != NULL)
842 		crcowfree(td);
843 	if (td->td_limit != NULL)
844 		lim_free(td->td_limit);
845 }
846 
847 void
848 thread_cow_update(struct thread *td)
849 {
850 	struct proc *p;
851 	struct ucred *oldcred;
852 	struct plimit *oldlimit;
853 
854 	p = td->td_proc;
855 	oldlimit = NULL;
856 	PROC_LOCK(p);
857 	oldcred = crcowsync();
858 	if (td->td_limit != p->p_limit) {
859 		oldlimit = td->td_limit;
860 		td->td_limit = lim_hold(p->p_limit);
861 	}
862 	td->td_cowgen = p->p_cowgen;
863 	PROC_UNLOCK(p);
864 	if (oldcred != NULL)
865 		crfree(oldcred);
866 	if (oldlimit != NULL)
867 		lim_free(oldlimit);
868 }
869 
870 /*
871  * Discard the current thread and exit from its context.
872  * Always called with scheduler locked.
873  *
874  * Because we can't free a thread while we're operating under its context,
875  * push the current thread into our CPU's deadthread holder. This means
876  * we needn't worry about someone else grabbing our context before we
877  * do a cpu_throw().
878  */
879 void
880 thread_exit(void)
881 {
882 	uint64_t runtime, new_switchtime;
883 	struct thread *td;
884 	struct thread *td2;
885 	struct proc *p;
886 	int wakeup_swapper;
887 
888 	td = curthread;
889 	p = td->td_proc;
890 
891 	PROC_SLOCK_ASSERT(p, MA_OWNED);
892 	mtx_assert(&Giant, MA_NOTOWNED);
893 
894 	PROC_LOCK_ASSERT(p, MA_OWNED);
895 	KASSERT(p != NULL, ("thread exiting without a process"));
896 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
897 	    (long)p->p_pid, td->td_name);
898 	SDT_PROBE0(proc, , , lwp__exit);
899 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
900 	MPASS(td->td_realucred == td->td_ucred);
901 
902 	/*
903 	 * drop FPU & debug register state storage, or any other
904 	 * architecture specific resources that
905 	 * would not be on a new untouched process.
906 	 */
907 	cpu_thread_exit(td);
908 
909 	/*
910 	 * The last thread is left attached to the process
911 	 * So that the whole bundle gets recycled. Skip
912 	 * all this stuff if we never had threads.
913 	 * EXIT clears all sign of other threads when
914 	 * it goes to single threading, so the last thread always
915 	 * takes the short path.
916 	 */
917 	if (p->p_flag & P_HADTHREADS) {
918 		if (p->p_numthreads > 1) {
919 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
920 			thread_unlink(td);
921 			td2 = FIRST_THREAD_IN_PROC(p);
922 			sched_exit_thread(td2, td);
923 
924 			/*
925 			 * The test below is NOT true if we are the
926 			 * sole exiting thread. P_STOPPED_SINGLE is unset
927 			 * in exit1() after it is the only survivor.
928 			 */
929 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
930 				if (p->p_numthreads == p->p_suspcount) {
931 					thread_lock(p->p_singlethread);
932 					wakeup_swapper = thread_unsuspend_one(
933 						p->p_singlethread, p, false);
934 					if (wakeup_swapper)
935 						kick_proc0();
936 				}
937 			}
938 
939 			PCPU_SET(deadthread, td);
940 		} else {
941 			/*
942 			 * The last thread is exiting.. but not through exit()
943 			 */
944 			panic ("thread_exit: Last thread exiting on its own");
945 		}
946 	}
947 #ifdef	HWPMC_HOOKS
948 	/*
949 	 * If this thread is part of a process that is being tracked by hwpmc(4),
950 	 * inform the module of the thread's impending exit.
951 	 */
952 	if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
953 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
954 		PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
955 	} else if (PMC_SYSTEM_SAMPLING_ACTIVE())
956 		PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
957 #endif
958 	PROC_UNLOCK(p);
959 	PROC_STATLOCK(p);
960 	thread_lock(td);
961 	PROC_SUNLOCK(p);
962 
963 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
964 	new_switchtime = cpu_ticks();
965 	runtime = new_switchtime - PCPU_GET(switchtime);
966 	td->td_runtime += runtime;
967 	td->td_incruntime += runtime;
968 	PCPU_SET(switchtime, new_switchtime);
969 	PCPU_SET(switchticks, ticks);
970 	VM_CNT_INC(v_swtch);
971 
972 	/* Save our resource usage in our process. */
973 	td->td_ru.ru_nvcsw++;
974 	ruxagg_locked(p, td);
975 	rucollect(&p->p_ru, &td->td_ru);
976 	PROC_STATUNLOCK(p);
977 
978 	TD_SET_STATE(td, TDS_INACTIVE);
979 #ifdef WITNESS
980 	witness_thread_exit(td);
981 #endif
982 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
983 	sched_throw(td);
984 	panic("I'm a teapot!");
985 	/* NOTREACHED */
986 }
987 
988 /*
989  * Do any thread specific cleanups that may be needed in wait()
990  * called with Giant, proc and schedlock not held.
991  */
992 void
993 thread_wait(struct proc *p)
994 {
995 	struct thread *td;
996 
997 	mtx_assert(&Giant, MA_NOTOWNED);
998 	KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
999 	KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
1000 	td = FIRST_THREAD_IN_PROC(p);
1001 	/* Lock the last thread so we spin until it exits cpu_throw(). */
1002 	thread_lock(td);
1003 	thread_unlock(td);
1004 	lock_profile_thread_exit(td);
1005 	cpuset_rel(td->td_cpuset);
1006 	td->td_cpuset = NULL;
1007 	cpu_thread_clean(td);
1008 	thread_cow_free(td);
1009 	callout_drain(&td->td_slpcallout);
1010 	thread_reap();	/* check for zombie threads etc. */
1011 }
1012 
1013 /*
1014  * Link a thread to a process.
1015  * set up anything that needs to be initialized for it to
1016  * be used by the process.
1017  */
1018 void
1019 thread_link(struct thread *td, struct proc *p)
1020 {
1021 
1022 	/*
1023 	 * XXX This can't be enabled because it's called for proc0 before
1024 	 * its lock has been created.
1025 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
1026 	 */
1027 	TD_SET_STATE(td, TDS_INACTIVE);
1028 	td->td_proc     = p;
1029 	td->td_flags    = TDF_INMEM;
1030 
1031 	LIST_INIT(&td->td_contested);
1032 	LIST_INIT(&td->td_lprof[0]);
1033 	LIST_INIT(&td->td_lprof[1]);
1034 #ifdef EPOCH_TRACE
1035 	SLIST_INIT(&td->td_epochs);
1036 #endif
1037 	sigqueue_init(&td->td_sigqueue, p);
1038 	callout_init(&td->td_slpcallout, 1);
1039 	TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
1040 	p->p_numthreads++;
1041 }
1042 
1043 /*
1044  * Called from:
1045  *  thread_exit()
1046  */
1047 void
1048 thread_unlink(struct thread *td)
1049 {
1050 	struct proc *p = td->td_proc;
1051 
1052 	PROC_LOCK_ASSERT(p, MA_OWNED);
1053 #ifdef EPOCH_TRACE
1054 	MPASS(SLIST_EMPTY(&td->td_epochs));
1055 #endif
1056 
1057 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
1058 	p->p_numthreads--;
1059 	/* could clear a few other things here */
1060 	/* Must  NOT clear links to proc! */
1061 }
1062 
1063 static int
1064 calc_remaining(struct proc *p, int mode)
1065 {
1066 	int remaining;
1067 
1068 	PROC_LOCK_ASSERT(p, MA_OWNED);
1069 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1070 	if (mode == SINGLE_EXIT)
1071 		remaining = p->p_numthreads;
1072 	else if (mode == SINGLE_BOUNDARY)
1073 		remaining = p->p_numthreads - p->p_boundary_count;
1074 	else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
1075 		remaining = p->p_numthreads - p->p_suspcount;
1076 	else
1077 		panic("calc_remaining: wrong mode %d", mode);
1078 	return (remaining);
1079 }
1080 
1081 static int
1082 remain_for_mode(int mode)
1083 {
1084 
1085 	return (mode == SINGLE_ALLPROC ? 0 : 1);
1086 }
1087 
1088 static int
1089 weed_inhib(int mode, struct thread *td2, struct proc *p)
1090 {
1091 	int wakeup_swapper;
1092 
1093 	PROC_LOCK_ASSERT(p, MA_OWNED);
1094 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1095 	THREAD_LOCK_ASSERT(td2, MA_OWNED);
1096 
1097 	wakeup_swapper = 0;
1098 
1099 	/*
1100 	 * Since the thread lock is dropped by the scheduler we have
1101 	 * to retry to check for races.
1102 	 */
1103 restart:
1104 	switch (mode) {
1105 	case SINGLE_EXIT:
1106 		if (TD_IS_SUSPENDED(td2)) {
1107 			wakeup_swapper |= thread_unsuspend_one(td2, p, true);
1108 			thread_lock(td2);
1109 			goto restart;
1110 		}
1111 		if (TD_CAN_ABORT(td2)) {
1112 			wakeup_swapper |= sleepq_abort(td2, EINTR);
1113 			return (wakeup_swapper);
1114 		}
1115 		break;
1116 	case SINGLE_BOUNDARY:
1117 	case SINGLE_NO_EXIT:
1118 		if (TD_IS_SUSPENDED(td2) &&
1119 		    (td2->td_flags & TDF_BOUNDARY) == 0) {
1120 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1121 			thread_lock(td2);
1122 			goto restart;
1123 		}
1124 		if (TD_CAN_ABORT(td2)) {
1125 			wakeup_swapper |= sleepq_abort(td2, ERESTART);
1126 			return (wakeup_swapper);
1127 		}
1128 		break;
1129 	case SINGLE_ALLPROC:
1130 		/*
1131 		 * ALLPROC suspend tries to avoid spurious EINTR for
1132 		 * threads sleeping interruptable, by suspending the
1133 		 * thread directly, similarly to sig_suspend_threads().
1134 		 * Since such sleep is not performed at the user
1135 		 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
1136 		 * is used to avoid immediate un-suspend.
1137 		 */
1138 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
1139 		    TDF_ALLPROCSUSP)) == 0) {
1140 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1141 			thread_lock(td2);
1142 			goto restart;
1143 		}
1144 		if (TD_CAN_ABORT(td2)) {
1145 			if ((td2->td_flags & TDF_SBDRY) == 0) {
1146 				thread_suspend_one(td2);
1147 				td2->td_flags |= TDF_ALLPROCSUSP;
1148 			} else {
1149 				wakeup_swapper |= sleepq_abort(td2, ERESTART);
1150 				return (wakeup_swapper);
1151 			}
1152 		}
1153 		break;
1154 	default:
1155 		break;
1156 	}
1157 	thread_unlock(td2);
1158 	return (wakeup_swapper);
1159 }
1160 
1161 /*
1162  * Enforce single-threading.
1163  *
1164  * Returns 1 if the caller must abort (another thread is waiting to
1165  * exit the process or similar). Process is locked!
1166  * Returns 0 when you are successfully the only thread running.
1167  * A process has successfully single threaded in the suspend mode when
1168  * There are no threads in user mode. Threads in the kernel must be
1169  * allowed to continue until they get to the user boundary. They may even
1170  * copy out their return values and data before suspending. They may however be
1171  * accelerated in reaching the user boundary as we will wake up
1172  * any sleeping threads that are interruptable. (PCATCH).
1173  */
1174 int
1175 thread_single(struct proc *p, int mode)
1176 {
1177 	struct thread *td;
1178 	struct thread *td2;
1179 	int remaining, wakeup_swapper;
1180 
1181 	td = curthread;
1182 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1183 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1184 	    ("invalid mode %d", mode));
1185 	/*
1186 	 * If allowing non-ALLPROC singlethreading for non-curproc
1187 	 * callers, calc_remaining() and remain_for_mode() should be
1188 	 * adjusted to also account for td->td_proc != p.  For now
1189 	 * this is not implemented because it is not used.
1190 	 */
1191 	KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
1192 	    (mode != SINGLE_ALLPROC && td->td_proc == p),
1193 	    ("mode %d proc %p curproc %p", mode, p, td->td_proc));
1194 	mtx_assert(&Giant, MA_NOTOWNED);
1195 	PROC_LOCK_ASSERT(p, MA_OWNED);
1196 
1197 	if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
1198 		return (0);
1199 
1200 	/* Is someone already single threading? */
1201 	if (p->p_singlethread != NULL && p->p_singlethread != td)
1202 		return (1);
1203 
1204 	if (mode == SINGLE_EXIT) {
1205 		p->p_flag |= P_SINGLE_EXIT;
1206 		p->p_flag &= ~P_SINGLE_BOUNDARY;
1207 	} else {
1208 		p->p_flag &= ~P_SINGLE_EXIT;
1209 		if (mode == SINGLE_BOUNDARY)
1210 			p->p_flag |= P_SINGLE_BOUNDARY;
1211 		else
1212 			p->p_flag &= ~P_SINGLE_BOUNDARY;
1213 	}
1214 	if (mode == SINGLE_ALLPROC)
1215 		p->p_flag |= P_TOTAL_STOP;
1216 	p->p_flag |= P_STOPPED_SINGLE;
1217 	PROC_SLOCK(p);
1218 	p->p_singlethread = td;
1219 	remaining = calc_remaining(p, mode);
1220 	while (remaining != remain_for_mode(mode)) {
1221 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
1222 			goto stopme;
1223 		wakeup_swapper = 0;
1224 		FOREACH_THREAD_IN_PROC(p, td2) {
1225 			if (td2 == td)
1226 				continue;
1227 			thread_lock(td2);
1228 			td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
1229 			if (TD_IS_INHIBITED(td2)) {
1230 				wakeup_swapper |= weed_inhib(mode, td2, p);
1231 #ifdef SMP
1232 			} else if (TD_IS_RUNNING(td2) && td != td2) {
1233 				forward_signal(td2);
1234 				thread_unlock(td2);
1235 #endif
1236 			} else
1237 				thread_unlock(td2);
1238 		}
1239 		if (wakeup_swapper)
1240 			kick_proc0();
1241 		remaining = calc_remaining(p, mode);
1242 
1243 		/*
1244 		 * Maybe we suspended some threads.. was it enough?
1245 		 */
1246 		if (remaining == remain_for_mode(mode))
1247 			break;
1248 
1249 stopme:
1250 		/*
1251 		 * Wake us up when everyone else has suspended.
1252 		 * In the mean time we suspend as well.
1253 		 */
1254 		thread_suspend_switch(td, p);
1255 		remaining = calc_remaining(p, mode);
1256 	}
1257 	if (mode == SINGLE_EXIT) {
1258 		/*
1259 		 * Convert the process to an unthreaded process.  The
1260 		 * SINGLE_EXIT is called by exit1() or execve(), in
1261 		 * both cases other threads must be retired.
1262 		 */
1263 		KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1264 		p->p_singlethread = NULL;
1265 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1266 
1267 		/*
1268 		 * Wait for any remaining threads to exit cpu_throw().
1269 		 */
1270 		while (p->p_exitthreads != 0) {
1271 			PROC_SUNLOCK(p);
1272 			PROC_UNLOCK(p);
1273 			sched_relinquish(td);
1274 			PROC_LOCK(p);
1275 			PROC_SLOCK(p);
1276 		}
1277 	} else if (mode == SINGLE_BOUNDARY) {
1278 		/*
1279 		 * Wait until all suspended threads are removed from
1280 		 * the processors.  The thread_suspend_check()
1281 		 * increments p_boundary_count while it is still
1282 		 * running, which makes it possible for the execve()
1283 		 * to destroy vmspace while our other threads are
1284 		 * still using the address space.
1285 		 *
1286 		 * We lock the thread, which is only allowed to
1287 		 * succeed after context switch code finished using
1288 		 * the address space.
1289 		 */
1290 		FOREACH_THREAD_IN_PROC(p, td2) {
1291 			if (td2 == td)
1292 				continue;
1293 			thread_lock(td2);
1294 			KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1295 			    ("td %p not on boundary", td2));
1296 			KASSERT(TD_IS_SUSPENDED(td2),
1297 			    ("td %p is not suspended", td2));
1298 			thread_unlock(td2);
1299 		}
1300 	}
1301 	PROC_SUNLOCK(p);
1302 	return (0);
1303 }
1304 
1305 bool
1306 thread_suspend_check_needed(void)
1307 {
1308 	struct proc *p;
1309 	struct thread *td;
1310 
1311 	td = curthread;
1312 	p = td->td_proc;
1313 	PROC_LOCK_ASSERT(p, MA_OWNED);
1314 	return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1315 	    (td->td_dbgflags & TDB_SUSPEND) != 0));
1316 }
1317 
1318 /*
1319  * Called in from locations that can safely check to see
1320  * whether we have to suspend or at least throttle for a
1321  * single-thread event (e.g. fork).
1322  *
1323  * Such locations include userret().
1324  * If the "return_instead" argument is non zero, the thread must be able to
1325  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1326  *
1327  * The 'return_instead' argument tells the function if it may do a
1328  * thread_exit() or suspend, or whether the caller must abort and back
1329  * out instead.
1330  *
1331  * If the thread that set the single_threading request has set the
1332  * P_SINGLE_EXIT bit in the process flags then this call will never return
1333  * if 'return_instead' is false, but will exit.
1334  *
1335  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1336  *---------------+--------------------+---------------------
1337  *       0       | returns 0          |   returns 0 or 1
1338  *               | when ST ends       |   immediately
1339  *---------------+--------------------+---------------------
1340  *       1       | thread exits       |   returns 1
1341  *               |                    |  immediately
1342  * 0 = thread_exit() or suspension ok,
1343  * other = return error instead of stopping the thread.
1344  *
1345  * While a full suspension is under effect, even a single threading
1346  * thread would be suspended if it made this call (but it shouldn't).
1347  * This call should only be made from places where
1348  * thread_exit() would be safe as that may be the outcome unless
1349  * return_instead is set.
1350  */
1351 int
1352 thread_suspend_check(int return_instead)
1353 {
1354 	struct thread *td;
1355 	struct proc *p;
1356 	int wakeup_swapper;
1357 
1358 	td = curthread;
1359 	p = td->td_proc;
1360 	mtx_assert(&Giant, MA_NOTOWNED);
1361 	PROC_LOCK_ASSERT(p, MA_OWNED);
1362 	while (thread_suspend_check_needed()) {
1363 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1364 			KASSERT(p->p_singlethread != NULL,
1365 			    ("singlethread not set"));
1366 			/*
1367 			 * The only suspension in action is a
1368 			 * single-threading. Single threader need not stop.
1369 			 * It is safe to access p->p_singlethread unlocked
1370 			 * because it can only be set to our address by us.
1371 			 */
1372 			if (p->p_singlethread == td)
1373 				return (0);	/* Exempt from stopping. */
1374 		}
1375 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1376 			return (EINTR);
1377 
1378 		/* Should we goto user boundary if we didn't come from there? */
1379 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1380 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1381 			return (ERESTART);
1382 
1383 		/*
1384 		 * Ignore suspend requests if they are deferred.
1385 		 */
1386 		if ((td->td_flags & TDF_SBDRY) != 0) {
1387 			KASSERT(return_instead,
1388 			    ("TDF_SBDRY set for unsafe thread_suspend_check"));
1389 			KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1390 			    (TDF_SEINTR | TDF_SERESTART),
1391 			    ("both TDF_SEINTR and TDF_SERESTART"));
1392 			return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1393 		}
1394 
1395 		/*
1396 		 * If the process is waiting for us to exit,
1397 		 * this thread should just suicide.
1398 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1399 		 */
1400 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1401 			PROC_UNLOCK(p);
1402 
1403 			/*
1404 			 * Allow Linux emulation layer to do some work
1405 			 * before thread suicide.
1406 			 */
1407 			if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1408 				(p->p_sysent->sv_thread_detach)(td);
1409 			umtx_thread_exit(td);
1410 			kern_thr_exit(td);
1411 			panic("stopped thread did not exit");
1412 		}
1413 
1414 		PROC_SLOCK(p);
1415 		thread_stopped(p);
1416 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1417 			if (p->p_numthreads == p->p_suspcount + 1) {
1418 				thread_lock(p->p_singlethread);
1419 				wakeup_swapper = thread_unsuspend_one(
1420 				    p->p_singlethread, p, false);
1421 				if (wakeup_swapper)
1422 					kick_proc0();
1423 			}
1424 		}
1425 		PROC_UNLOCK(p);
1426 		thread_lock(td);
1427 		/*
1428 		 * When a thread suspends, it just
1429 		 * gets taken off all queues.
1430 		 */
1431 		thread_suspend_one(td);
1432 		if (return_instead == 0) {
1433 			p->p_boundary_count++;
1434 			td->td_flags |= TDF_BOUNDARY;
1435 		}
1436 		PROC_SUNLOCK(p);
1437 		mi_switch(SW_INVOL | SWT_SUSPEND);
1438 		PROC_LOCK(p);
1439 	}
1440 	return (0);
1441 }
1442 
1443 /*
1444  * Check for possible stops and suspensions while executing a
1445  * casueword or similar transiently failing operation.
1446  *
1447  * The sleep argument controls whether the function can handle a stop
1448  * request itself or it should return ERESTART and the request is
1449  * proceed at the kernel/user boundary in ast.
1450  *
1451  * Typically, when retrying due to casueword(9) failure (rv == 1), we
1452  * should handle the stop requests there, with exception of cases when
1453  * the thread owns a kernel resource, for instance busied the umtx
1454  * key, or when functions return immediately if thread_check_susp()
1455  * returned non-zero.  On the other hand, retrying the whole lock
1456  * operation, we better not stop there but delegate the handling to
1457  * ast.
1458  *
1459  * If the request is for thread termination P_SINGLE_EXIT, we cannot
1460  * handle it at all, and simply return EINTR.
1461  */
1462 int
1463 thread_check_susp(struct thread *td, bool sleep)
1464 {
1465 	struct proc *p;
1466 	int error;
1467 
1468 	/*
1469 	 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
1470 	 * eventually break the lockstep loop.
1471 	 */
1472 	if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
1473 		return (0);
1474 	error = 0;
1475 	p = td->td_proc;
1476 	PROC_LOCK(p);
1477 	if (p->p_flag & P_SINGLE_EXIT)
1478 		error = EINTR;
1479 	else if (P_SHOULDSTOP(p) ||
1480 	    ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1481 		error = sleep ? thread_suspend_check(0) : ERESTART;
1482 	PROC_UNLOCK(p);
1483 	return (error);
1484 }
1485 
1486 void
1487 thread_suspend_switch(struct thread *td, struct proc *p)
1488 {
1489 
1490 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1491 	PROC_LOCK_ASSERT(p, MA_OWNED);
1492 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1493 	/*
1494 	 * We implement thread_suspend_one in stages here to avoid
1495 	 * dropping the proc lock while the thread lock is owned.
1496 	 */
1497 	if (p == td->td_proc) {
1498 		thread_stopped(p);
1499 		p->p_suspcount++;
1500 	}
1501 	PROC_UNLOCK(p);
1502 	thread_lock(td);
1503 	td->td_flags &= ~TDF_NEEDSUSPCHK;
1504 	TD_SET_SUSPENDED(td);
1505 	sched_sleep(td, 0);
1506 	PROC_SUNLOCK(p);
1507 	DROP_GIANT();
1508 	mi_switch(SW_VOL | SWT_SUSPEND);
1509 	PICKUP_GIANT();
1510 	PROC_LOCK(p);
1511 	PROC_SLOCK(p);
1512 }
1513 
1514 void
1515 thread_suspend_one(struct thread *td)
1516 {
1517 	struct proc *p;
1518 
1519 	p = td->td_proc;
1520 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1521 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1522 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1523 	p->p_suspcount++;
1524 	td->td_flags &= ~TDF_NEEDSUSPCHK;
1525 	TD_SET_SUSPENDED(td);
1526 	sched_sleep(td, 0);
1527 }
1528 
1529 static int
1530 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1531 {
1532 
1533 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1534 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1535 	TD_CLR_SUSPENDED(td);
1536 	td->td_flags &= ~TDF_ALLPROCSUSP;
1537 	if (td->td_proc == p) {
1538 		PROC_SLOCK_ASSERT(p, MA_OWNED);
1539 		p->p_suspcount--;
1540 		if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1541 			td->td_flags &= ~TDF_BOUNDARY;
1542 			p->p_boundary_count--;
1543 		}
1544 	}
1545 	return (setrunnable(td, 0));
1546 }
1547 
1548 void
1549 thread_run_flash(struct thread *td)
1550 {
1551 	struct proc *p;
1552 
1553 	p = td->td_proc;
1554 	PROC_LOCK_ASSERT(p, MA_OWNED);
1555 
1556 	if (TD_ON_SLEEPQ(td))
1557 		sleepq_remove_nested(td);
1558 	else
1559 		thread_lock(td);
1560 
1561 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1562 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1563 
1564 	TD_CLR_SUSPENDED(td);
1565 	PROC_SLOCK(p);
1566 	MPASS(p->p_suspcount > 0);
1567 	p->p_suspcount--;
1568 	PROC_SUNLOCK(p);
1569 	if (setrunnable(td, 0))
1570 		kick_proc0();
1571 }
1572 
1573 /*
1574  * Allow all threads blocked by single threading to continue running.
1575  */
1576 void
1577 thread_unsuspend(struct proc *p)
1578 {
1579 	struct thread *td;
1580 	int wakeup_swapper;
1581 
1582 	PROC_LOCK_ASSERT(p, MA_OWNED);
1583 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1584 	wakeup_swapper = 0;
1585 	if (!P_SHOULDSTOP(p)) {
1586                 FOREACH_THREAD_IN_PROC(p, td) {
1587 			thread_lock(td);
1588 			if (TD_IS_SUSPENDED(td)) {
1589 				wakeup_swapper |= thread_unsuspend_one(td, p,
1590 				    true);
1591 			} else
1592 				thread_unlock(td);
1593 		}
1594 	} else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1595 	    p->p_numthreads == p->p_suspcount) {
1596 		/*
1597 		 * Stopping everything also did the job for the single
1598 		 * threading request. Now we've downgraded to single-threaded,
1599 		 * let it continue.
1600 		 */
1601 		if (p->p_singlethread->td_proc == p) {
1602 			thread_lock(p->p_singlethread);
1603 			wakeup_swapper = thread_unsuspend_one(
1604 			    p->p_singlethread, p, false);
1605 		}
1606 	}
1607 	if (wakeup_swapper)
1608 		kick_proc0();
1609 }
1610 
1611 /*
1612  * End the single threading mode..
1613  */
1614 void
1615 thread_single_end(struct proc *p, int mode)
1616 {
1617 	struct thread *td;
1618 	int wakeup_swapper;
1619 
1620 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1621 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1622 	    ("invalid mode %d", mode));
1623 	PROC_LOCK_ASSERT(p, MA_OWNED);
1624 	KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1625 	    (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1626 	    ("mode %d does not match P_TOTAL_STOP", mode));
1627 	KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1628 	    ("thread_single_end from other thread %p %p",
1629 	    curthread, p->p_singlethread));
1630 	KASSERT(mode != SINGLE_BOUNDARY ||
1631 	    (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1632 	    ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1633 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1634 	    P_TOTAL_STOP);
1635 	PROC_SLOCK(p);
1636 	p->p_singlethread = NULL;
1637 	wakeup_swapper = 0;
1638 	/*
1639 	 * If there are other threads they may now run,
1640 	 * unless of course there is a blanket 'stop order'
1641 	 * on the process. The single threader must be allowed
1642 	 * to continue however as this is a bad place to stop.
1643 	 */
1644 	if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1645                 FOREACH_THREAD_IN_PROC(p, td) {
1646 			thread_lock(td);
1647 			if (TD_IS_SUSPENDED(td)) {
1648 				wakeup_swapper |= thread_unsuspend_one(td, p,
1649 				    mode == SINGLE_BOUNDARY);
1650 			} else
1651 				thread_unlock(td);
1652 		}
1653 	}
1654 	KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1655 	    ("inconsistent boundary count %d", p->p_boundary_count));
1656 	PROC_SUNLOCK(p);
1657 	if (wakeup_swapper)
1658 		kick_proc0();
1659 }
1660 
1661 /*
1662  * Locate a thread by number and return with proc lock held.
1663  *
1664  * thread exit establishes proc -> tidhash lock ordering, but lookup
1665  * takes tidhash first and needs to return locked proc.
1666  *
1667  * The problem is worked around by relying on type-safety of both
1668  * structures and doing the work in 2 steps:
1669  * - tidhash-locked lookup which saves both thread and proc pointers
1670  * - proc-locked verification that the found thread still matches
1671  */
1672 static bool
1673 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1674 {
1675 #define RUN_THRESH	16
1676 	struct proc *p;
1677 	struct thread *td;
1678 	int run;
1679 	bool locked;
1680 
1681 	run = 0;
1682 	rw_rlock(TIDHASHLOCK(tid));
1683 	locked = true;
1684 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1685 		if (td->td_tid != tid) {
1686 			run++;
1687 			continue;
1688 		}
1689 		p = td->td_proc;
1690 		if (pid != -1 && p->p_pid != pid) {
1691 			td = NULL;
1692 			break;
1693 		}
1694 		if (run > RUN_THRESH) {
1695 			if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1696 				LIST_REMOVE(td, td_hash);
1697 				LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1698 					td, td_hash);
1699 				rw_wunlock(TIDHASHLOCK(tid));
1700 				locked = false;
1701 				break;
1702 			}
1703 		}
1704 		break;
1705 	}
1706 	if (locked)
1707 		rw_runlock(TIDHASHLOCK(tid));
1708 	if (td == NULL)
1709 		return (false);
1710 	*pp = p;
1711 	*tdp = td;
1712 	return (true);
1713 }
1714 
1715 struct thread *
1716 tdfind(lwpid_t tid, pid_t pid)
1717 {
1718 	struct proc *p;
1719 	struct thread *td;
1720 
1721 	td = curthread;
1722 	if (td->td_tid == tid) {
1723 		if (pid != -1 && td->td_proc->p_pid != pid)
1724 			return (NULL);
1725 		PROC_LOCK(td->td_proc);
1726 		return (td);
1727 	}
1728 
1729 	for (;;) {
1730 		if (!tdfind_hash(tid, pid, &p, &td))
1731 			return (NULL);
1732 		PROC_LOCK(p);
1733 		if (td->td_tid != tid) {
1734 			PROC_UNLOCK(p);
1735 			continue;
1736 		}
1737 		if (td->td_proc != p) {
1738 			PROC_UNLOCK(p);
1739 			continue;
1740 		}
1741 		if (p->p_state == PRS_NEW) {
1742 			PROC_UNLOCK(p);
1743 			return (NULL);
1744 		}
1745 		return (td);
1746 	}
1747 }
1748 
1749 void
1750 tidhash_add(struct thread *td)
1751 {
1752 	rw_wlock(TIDHASHLOCK(td->td_tid));
1753 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1754 	rw_wunlock(TIDHASHLOCK(td->td_tid));
1755 }
1756 
1757 void
1758 tidhash_remove(struct thread *td)
1759 {
1760 
1761 	rw_wlock(TIDHASHLOCK(td->td_tid));
1762 	LIST_REMOVE(td, td_hash);
1763 	rw_wunlock(TIDHASHLOCK(td->td_tid));
1764 }
1765