xref: /freebsd/sys/kern/kern_thread.c (revision 87b759f0fa1f7554d50ce640c40138512bbded44)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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/systm.h>
35 #include <sys/asan.h>
36 #include <sys/kernel.h>
37 #include <sys/lock.h>
38 #include <sys/msan.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/bitstring.h>
42 #include <sys/epoch.h>
43 #include <sys/rangelock.h>
44 #include <sys/resourcevar.h>
45 #include <sys/sdt.h>
46 #include <sys/smp.h>
47 #include <sys/sched.h>
48 #include <sys/sleepqueue.h>
49 #include <sys/selinfo.h>
50 #include <sys/syscallsubr.h>
51 #include <sys/dtrace_bsd.h>
52 #include <sys/sysent.h>
53 #include <sys/turnstile.h>
54 #include <sys/taskqueue.h>
55 #include <sys/ktr.h>
56 #include <sys/rwlock.h>
57 #include <sys/umtxvar.h>
58 #include <sys/vmmeter.h>
59 #include <sys/cpuset.h>
60 #ifdef	HWPMC_HOOKS
61 #include <sys/pmckern.h>
62 #endif
63 #include <sys/priv.h>
64 
65 #include <security/audit/audit.h>
66 
67 #include <vm/pmap.h>
68 #include <vm/vm.h>
69 #include <vm/vm_extern.h>
70 #include <vm/uma.h>
71 #include <vm/vm_phys.h>
72 #include <sys/eventhandler.h>
73 
74 /*
75  * Asserts below verify the stability of struct thread and struct proc
76  * layout, as exposed by KBI to modules.  On head, the KBI is allowed
77  * to drift, change to the structures must be accompanied by the
78  * assert update.
79  *
80  * On the stable branches after KBI freeze, conditions must not be
81  * violated.  Typically new fields are moved to the end of the
82  * structures.
83  */
84 #ifdef __amd64__
85 _Static_assert(offsetof(struct thread, td_flags) == 0x108,
86     "struct thread KBI td_flags");
87 _Static_assert(offsetof(struct thread, td_pflags) == 0x114,
88     "struct thread KBI td_pflags");
89 _Static_assert(offsetof(struct thread, td_frame) == 0x4b8,
90     "struct thread KBI td_frame");
91 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6c0,
92     "struct thread KBI td_emuldata");
93 _Static_assert(offsetof(struct proc, p_flag) == 0xb8,
94     "struct proc KBI p_flag");
95 _Static_assert(offsetof(struct proc, p_pid) == 0xc4,
96     "struct proc KBI p_pid");
97 _Static_assert(offsetof(struct proc, p_filemon) == 0x3c8,
98     "struct proc KBI p_filemon");
99 _Static_assert(offsetof(struct proc, p_comm) == 0x3e0,
100     "struct proc KBI p_comm");
101 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4d0,
102     "struct proc KBI p_emuldata");
103 #endif
104 #ifdef __i386__
105 _Static_assert(offsetof(struct thread, td_flags) == 0x9c,
106     "struct thread KBI td_flags");
107 _Static_assert(offsetof(struct thread, td_pflags) == 0xa8,
108     "struct thread KBI td_pflags");
109 _Static_assert(offsetof(struct thread, td_frame) == 0x318,
110     "struct thread KBI td_frame");
111 _Static_assert(offsetof(struct thread, td_emuldata) == 0x35c,
112     "struct thread KBI td_emuldata");
113 _Static_assert(offsetof(struct proc, p_flag) == 0x6c,
114     "struct proc KBI p_flag");
115 _Static_assert(offsetof(struct proc, p_pid) == 0x78,
116     "struct proc KBI p_pid");
117 _Static_assert(offsetof(struct proc, p_filemon) == 0x270,
118     "struct proc KBI p_filemon");
119 _Static_assert(offsetof(struct proc, p_comm) == 0x284,
120     "struct proc KBI p_comm");
121 _Static_assert(offsetof(struct proc, p_emuldata) == 0x318,
122     "struct proc KBI p_emuldata");
123 #endif
124 
125 SDT_PROVIDER_DECLARE(proc);
126 SDT_PROBE_DEFINE(proc, , , lwp__exit);
127 
128 /*
129  * thread related storage.
130  */
131 static uma_zone_t thread_zone;
132 
133 struct thread_domain_data {
134 	struct thread	*tdd_zombies;
135 	int		tdd_reapticks;
136 } __aligned(CACHE_LINE_SIZE);
137 
138 static struct thread_domain_data thread_domain_data[MAXMEMDOM];
139 
140 static struct task	thread_reap_task;
141 static struct callout  	thread_reap_callout;
142 
143 static void thread_zombie(struct thread *);
144 static void thread_reap(void);
145 static void thread_reap_all(void);
146 static void thread_reap_task_cb(void *, int);
147 static void thread_reap_callout_cb(void *);
148 static void thread_unsuspend_one(struct thread *td, struct proc *p,
149     bool boundary);
150 static void thread_free_batched(struct thread *td);
151 
152 static __exclusive_cache_line struct mtx tid_lock;
153 static bitstr_t *tid_bitmap;
154 
155 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
156 
157 static int maxthread;
158 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
159     &maxthread, 0, "Maximum number of threads");
160 
161 static __exclusive_cache_line int nthreads;
162 
163 static LIST_HEAD(tidhashhead, thread) *tidhashtbl;
164 static u_long	tidhash;
165 static u_long	tidhashlock;
166 static struct	rwlock *tidhashtbl_lock;
167 #define	TIDHASH(tid)		(&tidhashtbl[(tid) & tidhash])
168 #define	TIDHASHLOCK(tid)	(&tidhashtbl_lock[(tid) & tidhashlock])
169 
170 EVENTHANDLER_LIST_DEFINE(thread_ctor);
171 EVENTHANDLER_LIST_DEFINE(thread_dtor);
172 EVENTHANDLER_LIST_DEFINE(thread_init);
173 EVENTHANDLER_LIST_DEFINE(thread_fini);
174 
175 static bool
176 thread_count_inc_try(void)
177 {
178 	int nthreads_new;
179 
180 	nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1;
181 	if (nthreads_new >= maxthread - 100) {
182 		if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
183 		    nthreads_new >= maxthread) {
184 			atomic_subtract_int(&nthreads, 1);
185 			return (false);
186 		}
187 	}
188 	return (true);
189 }
190 
191 static bool
192 thread_count_inc(void)
193 {
194 	static struct timeval lastfail;
195 	static int curfail;
196 
197 	thread_reap();
198 	if (thread_count_inc_try()) {
199 		return (true);
200 	}
201 
202 	thread_reap_all();
203 	if (thread_count_inc_try()) {
204 		return (true);
205 	}
206 
207 	if (ppsratecheck(&lastfail, &curfail, 1)) {
208 		printf("maxthread limit exceeded by uid %u "
209 		    "(pid %d); consider increasing kern.maxthread\n",
210 		    curthread->td_ucred->cr_ruid, curproc->p_pid);
211 	}
212 	return (false);
213 }
214 
215 static void
216 thread_count_sub(int n)
217 {
218 
219 	atomic_subtract_int(&nthreads, n);
220 }
221 
222 static void
223 thread_count_dec(void)
224 {
225 
226 	thread_count_sub(1);
227 }
228 
229 static lwpid_t
230 tid_alloc(void)
231 {
232 	static lwpid_t trytid;
233 	lwpid_t tid;
234 
235 	mtx_lock(&tid_lock);
236 	/*
237 	 * It is an invariant that the bitmap is big enough to hold maxthread
238 	 * IDs. If we got to this point there has to be at least one free.
239 	 */
240 	if (trytid >= maxthread)
241 		trytid = 0;
242 	bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
243 	if (tid == -1) {
244 		KASSERT(trytid != 0, ("unexpectedly ran out of IDs"));
245 		trytid = 0;
246 		bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
247 		KASSERT(tid != -1, ("unexpectedly ran out of IDs"));
248 	}
249 	bit_set(tid_bitmap, tid);
250 	trytid = tid + 1;
251 	mtx_unlock(&tid_lock);
252 	return (tid + NO_PID);
253 }
254 
255 static void
256 tid_free_locked(lwpid_t rtid)
257 {
258 	lwpid_t tid;
259 
260 	mtx_assert(&tid_lock, MA_OWNED);
261 	KASSERT(rtid >= NO_PID,
262 	    ("%s: invalid tid %d\n", __func__, rtid));
263 	tid = rtid - NO_PID;
264 	KASSERT(bit_test(tid_bitmap, tid) != 0,
265 	    ("thread ID %d not allocated\n", rtid));
266 	bit_clear(tid_bitmap, tid);
267 }
268 
269 static void
270 tid_free(lwpid_t rtid)
271 {
272 
273 	mtx_lock(&tid_lock);
274 	tid_free_locked(rtid);
275 	mtx_unlock(&tid_lock);
276 }
277 
278 static void
279 tid_free_batch(lwpid_t *batch, int n)
280 {
281 	int i;
282 
283 	mtx_lock(&tid_lock);
284 	for (i = 0; i < n; i++) {
285 		tid_free_locked(batch[i]);
286 	}
287 	mtx_unlock(&tid_lock);
288 }
289 
290 /*
291  * Batching for thread reapping.
292  */
293 struct tidbatch {
294 	lwpid_t tab[16];
295 	int n;
296 };
297 
298 static void
299 tidbatch_prep(struct tidbatch *tb)
300 {
301 
302 	tb->n = 0;
303 }
304 
305 static void
306 tidbatch_add(struct tidbatch *tb, struct thread *td)
307 {
308 
309 	KASSERT(tb->n < nitems(tb->tab),
310 	    ("%s: count too high %d", __func__, tb->n));
311 	tb->tab[tb->n] = td->td_tid;
312 	tb->n++;
313 }
314 
315 static void
316 tidbatch_process(struct tidbatch *tb)
317 {
318 
319 	KASSERT(tb->n <= nitems(tb->tab),
320 	    ("%s: count too high %d", __func__, tb->n));
321 	if (tb->n == nitems(tb->tab)) {
322 		tid_free_batch(tb->tab, tb->n);
323 		tb->n = 0;
324 	}
325 }
326 
327 static void
328 tidbatch_final(struct tidbatch *tb)
329 {
330 
331 	KASSERT(tb->n <= nitems(tb->tab),
332 	    ("%s: count too high %d", __func__, tb->n));
333 	if (tb->n != 0) {
334 		tid_free_batch(tb->tab, tb->n);
335 	}
336 }
337 
338 /*
339  * Batching thread count free, for consistency
340  */
341 struct tdcountbatch {
342 	int n;
343 };
344 
345 static void
346 tdcountbatch_prep(struct tdcountbatch *tb)
347 {
348 
349 	tb->n = 0;
350 }
351 
352 static void
353 tdcountbatch_add(struct tdcountbatch *tb, struct thread *td __unused)
354 {
355 
356 	tb->n++;
357 }
358 
359 static void
360 tdcountbatch_process(struct tdcountbatch *tb)
361 {
362 
363 	if (tb->n == 32) {
364 		thread_count_sub(tb->n);
365 		tb->n = 0;
366 	}
367 }
368 
369 static void
370 tdcountbatch_final(struct tdcountbatch *tb)
371 {
372 
373 	if (tb->n != 0) {
374 		thread_count_sub(tb->n);
375 	}
376 }
377 
378 /*
379  * Prepare a thread for use.
380  */
381 static int
382 thread_ctor(void *mem, int size, void *arg, int flags)
383 {
384 	struct thread	*td;
385 
386 	td = (struct thread *)mem;
387 	TD_SET_STATE(td, TDS_INACTIVE);
388 	td->td_lastcpu = td->td_oncpu = NOCPU;
389 
390 	/*
391 	 * Note that td_critnest begins life as 1 because the thread is not
392 	 * running and is thereby implicitly waiting to be on the receiving
393 	 * end of a context switch.
394 	 */
395 	td->td_critnest = 1;
396 	td->td_lend_user_pri = PRI_MAX;
397 #ifdef AUDIT
398 	audit_thread_alloc(td);
399 #endif
400 #ifdef KDTRACE_HOOKS
401 	kdtrace_thread_ctor(td);
402 #endif
403 	umtx_thread_alloc(td);
404 	MPASS(td->td_sel == NULL);
405 	return (0);
406 }
407 
408 /*
409  * Reclaim a thread after use.
410  */
411 static void
412 thread_dtor(void *mem, int size, void *arg)
413 {
414 	struct thread *td;
415 
416 	td = (struct thread *)mem;
417 
418 #ifdef INVARIANTS
419 	/* Verify that this thread is in a safe state to free. */
420 	switch (TD_GET_STATE(td)) {
421 	case TDS_INHIBITED:
422 	case TDS_RUNNING:
423 	case TDS_CAN_RUN:
424 	case TDS_RUNQ:
425 		/*
426 		 * We must never unlink a thread that is in one of
427 		 * these states, because it is currently active.
428 		 */
429 		panic("bad state for thread unlinking");
430 		/* NOTREACHED */
431 	case TDS_INACTIVE:
432 		break;
433 	default:
434 		panic("bad thread state");
435 		/* NOTREACHED */
436 	}
437 #endif
438 #ifdef AUDIT
439 	audit_thread_free(td);
440 #endif
441 #ifdef KDTRACE_HOOKS
442 	kdtrace_thread_dtor(td);
443 #endif
444 	/* Free all OSD associated to this thread. */
445 	osd_thread_exit(td);
446 	ast_kclear(td);
447 	seltdfini(td);
448 }
449 
450 /*
451  * Initialize type-stable parts of a thread (when newly created).
452  */
453 static int
454 thread_init(void *mem, int size, int flags)
455 {
456 	struct thread *td;
457 
458 	td = (struct thread *)mem;
459 
460 	td->td_allocdomain = vm_phys_domain(vtophys(td));
461 	td->td_sleepqueue = sleepq_alloc();
462 	td->td_turnstile = turnstile_alloc();
463 	EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
464 	umtx_thread_init(td);
465 	td->td_kstack = 0;
466 	td->td_sel = NULL;
467 	return (0);
468 }
469 
470 /*
471  * Tear down type-stable parts of a thread (just before being discarded).
472  */
473 static void
474 thread_fini(void *mem, int size)
475 {
476 	struct thread *td;
477 
478 	td = (struct thread *)mem;
479 	EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
480 	turnstile_free(td->td_turnstile);
481 	sleepq_free(td->td_sleepqueue);
482 	umtx_thread_fini(td);
483 	MPASS(td->td_sel == NULL);
484 }
485 
486 /*
487  * For a newly created process,
488  * link up all the structures and its initial threads etc.
489  * called from:
490  * {arch}/{arch}/machdep.c   {arch}_init(), init386() etc.
491  * proc_dtor() (should go away)
492  * proc_init()
493  */
494 void
495 proc_linkup0(struct proc *p, struct thread *td)
496 {
497 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
498 	proc_linkup(p, td);
499 }
500 
501 void
502 proc_linkup(struct proc *p, struct thread *td)
503 {
504 
505 	sigqueue_init(&p->p_sigqueue, p);
506 	p->p_ksi = ksiginfo_alloc(M_WAITOK);
507 	if (p->p_ksi != NULL) {
508 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
509 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
510 	}
511 	LIST_INIT(&p->p_mqnotifier);
512 	p->p_numthreads = 0;
513 	thread_link(td, p);
514 }
515 
516 static void
517 ast_suspend(struct thread *td, int tda __unused)
518 {
519 	struct proc *p;
520 
521 	p = td->td_proc;
522 	/*
523 	 * We need to check to see if we have to exit or wait due to a
524 	 * single threading requirement or some other STOP condition.
525 	 */
526 	PROC_LOCK(p);
527 	thread_suspend_check(0);
528 	PROC_UNLOCK(p);
529 }
530 
531 extern int max_threads_per_proc;
532 
533 /*
534  * Initialize global thread allocation resources.
535  */
536 void
537 threadinit(void)
538 {
539 	u_long i;
540 	lwpid_t tid0;
541 
542 	/*
543 	 * Place an upper limit on threads which can be allocated.
544 	 *
545 	 * Note that other factors may make the de facto limit much lower.
546 	 *
547 	 * Platform limits are somewhat arbitrary but deemed "more than good
548 	 * enough" for the foreseable future.
549 	 */
550 	if (maxthread == 0) {
551 #ifdef _LP64
552 		maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
553 #else
554 		maxthread = MIN(maxproc * max_threads_per_proc, 100000);
555 #endif
556 	}
557 
558 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
559 	tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
560 	/*
561 	 * Handle thread0.
562 	 */
563 	thread_count_inc();
564 	tid0 = tid_alloc();
565 	if (tid0 != THREAD0_TID)
566 		panic("tid0 %d != %d\n", tid0, THREAD0_TID);
567 
568 	/*
569 	 * Thread structures are specially aligned so that (at least) the
570 	 * 5 lower bits of a pointer to 'struct thead' must be 0.  These bits
571 	 * are used by synchronization primitives to store flags in pointers to
572 	 * such structures.
573 	 */
574 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
575 	    thread_ctor, thread_dtor, thread_init, thread_fini,
576 	    UMA_ALIGN_CACHE_AND_MASK(32 - 1), UMA_ZONE_NOFREE);
577 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
578 	tidhashlock = (tidhash + 1) / 64;
579 	if (tidhashlock > 0)
580 		tidhashlock--;
581 	tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
582 	    M_TIDHASH, M_WAITOK | M_ZERO);
583 	for (i = 0; i < tidhashlock + 1; i++)
584 		rw_init(&tidhashtbl_lock[i], "tidhash");
585 
586 	TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL);
587 	callout_init(&thread_reap_callout, 1);
588 	callout_reset(&thread_reap_callout, 5 * hz,
589 	    thread_reap_callout_cb, NULL);
590 	ast_register(TDA_SUSPEND, ASTR_ASTF_REQUIRED, 0, ast_suspend);
591 }
592 
593 /*
594  * Place an unused thread on the zombie list.
595  */
596 void
597 thread_zombie(struct thread *td)
598 {
599 	struct thread_domain_data *tdd;
600 	struct thread *ztd;
601 
602 	tdd = &thread_domain_data[td->td_allocdomain];
603 	ztd = atomic_load_ptr(&tdd->tdd_zombies);
604 	for (;;) {
605 		td->td_zombie = ztd;
606 		if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
607 		    (uintptr_t *)&ztd, (uintptr_t)td))
608 			break;
609 		continue;
610 	}
611 }
612 
613 /*
614  * Release a thread that has exited after cpu_throw().
615  */
616 void
617 thread_stash(struct thread *td)
618 {
619 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
620 	thread_zombie(td);
621 }
622 
623 /*
624  * Reap zombies from passed domain.
625  */
626 static void
627 thread_reap_domain(struct thread_domain_data *tdd)
628 {
629 	struct thread *itd, *ntd;
630 	struct tidbatch tidbatch;
631 	struct credbatch credbatch;
632 	struct limbatch limbatch;
633 	struct tdcountbatch tdcountbatch;
634 
635 	/*
636 	 * Reading upfront is pessimal if followed by concurrent atomic_swap,
637 	 * but most of the time the list is empty.
638 	 */
639 	if (tdd->tdd_zombies == NULL)
640 		return;
641 
642 	itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
643 	    (uintptr_t)NULL);
644 	if (itd == NULL)
645 		return;
646 
647 	/*
648 	 * Multiple CPUs can get here, the race is fine as ticks is only
649 	 * advisory.
650 	 */
651 	tdd->tdd_reapticks = ticks;
652 
653 	tidbatch_prep(&tidbatch);
654 	credbatch_prep(&credbatch);
655 	limbatch_prep(&limbatch);
656 	tdcountbatch_prep(&tdcountbatch);
657 
658 	while (itd != NULL) {
659 		ntd = itd->td_zombie;
660 		EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd);
661 
662 		tidbatch_add(&tidbatch, itd);
663 		credbatch_add(&credbatch, itd);
664 		limbatch_add(&limbatch, itd);
665 		tdcountbatch_add(&tdcountbatch, itd);
666 
667 		thread_free_batched(itd);
668 
669 		tidbatch_process(&tidbatch);
670 		credbatch_process(&credbatch);
671 		limbatch_process(&limbatch);
672 		tdcountbatch_process(&tdcountbatch);
673 
674 		itd = ntd;
675 	}
676 
677 	tidbatch_final(&tidbatch);
678 	credbatch_final(&credbatch);
679 	limbatch_final(&limbatch);
680 	tdcountbatch_final(&tdcountbatch);
681 }
682 
683 /*
684  * Reap zombies from all domains.
685  */
686 static void
687 thread_reap_all(void)
688 {
689 	struct thread_domain_data *tdd;
690 	int i, domain;
691 
692 	domain = PCPU_GET(domain);
693 	for (i = 0; i < vm_ndomains; i++) {
694 		tdd = &thread_domain_data[(i + domain) % vm_ndomains];
695 		thread_reap_domain(tdd);
696 	}
697 }
698 
699 /*
700  * Reap zombies from local domain.
701  */
702 static void
703 thread_reap(void)
704 {
705 	struct thread_domain_data *tdd;
706 	int domain;
707 
708 	domain = PCPU_GET(domain);
709 	tdd = &thread_domain_data[domain];
710 
711 	thread_reap_domain(tdd);
712 }
713 
714 static void
715 thread_reap_task_cb(void *arg __unused, int pending __unused)
716 {
717 
718 	thread_reap_all();
719 }
720 
721 static void
722 thread_reap_callout_cb(void *arg __unused)
723 {
724 	struct thread_domain_data *tdd;
725 	int i, cticks, lticks;
726 	bool wantreap;
727 
728 	wantreap = false;
729 	cticks = atomic_load_int(&ticks);
730 	for (i = 0; i < vm_ndomains; i++) {
731 		tdd = &thread_domain_data[i];
732 		lticks = tdd->tdd_reapticks;
733 		if (tdd->tdd_zombies != NULL &&
734 		    (u_int)(cticks - lticks) > 5 * hz) {
735 			wantreap = true;
736 			break;
737 		}
738 	}
739 
740 	if (wantreap)
741 		taskqueue_enqueue(taskqueue_thread, &thread_reap_task);
742 	callout_reset(&thread_reap_callout, 5 * hz,
743 	    thread_reap_callout_cb, NULL);
744 }
745 
746 /*
747  * Calling this function guarantees that any thread that exited before
748  * the call is reaped when the function returns.  By 'exited' we mean
749  * a thread removed from the process linkage with thread_unlink().
750  * Practically this means that caller must lock/unlock corresponding
751  * process lock before the call, to synchronize with thread_exit().
752  */
753 void
754 thread_reap_barrier(void)
755 {
756 	struct task *t;
757 
758 	/*
759 	 * First do context switches to each CPU to ensure that all
760 	 * PCPU pc_deadthreads are moved to zombie list.
761 	 */
762 	quiesce_all_cpus("", PDROP);
763 
764 	/*
765 	 * Second, fire the task in the same thread as normal
766 	 * thread_reap() is done, to serialize reaping.
767 	 */
768 	t = malloc(sizeof(*t), M_TEMP, M_WAITOK);
769 	TASK_INIT(t, 0, thread_reap_task_cb, t);
770 	taskqueue_enqueue(taskqueue_thread, t);
771 	taskqueue_drain(taskqueue_thread, t);
772 	free(t, M_TEMP);
773 }
774 
775 /*
776  * Allocate a thread.
777  */
778 struct thread *
779 thread_alloc(int pages)
780 {
781 	struct thread *td;
782 	lwpid_t tid;
783 
784 	if (!thread_count_inc()) {
785 		return (NULL);
786 	}
787 
788 	tid = tid_alloc();
789 	td = uma_zalloc(thread_zone, M_WAITOK);
790 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
791 	if (!vm_thread_new(td, pages)) {
792 		uma_zfree(thread_zone, td);
793 		tid_free(tid);
794 		thread_count_dec();
795 		return (NULL);
796 	}
797 	td->td_tid = tid;
798 	bzero(&td->td_sa.args, sizeof(td->td_sa.args));
799 	kasan_thread_alloc(td);
800 	kmsan_thread_alloc(td);
801 	cpu_thread_alloc(td);
802 	EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
803 	return (td);
804 }
805 
806 int
807 thread_recycle(struct thread *td, int pages)
808 {
809 	if (td->td_kstack == 0 || td->td_kstack_pages != pages) {
810 		if (td->td_kstack != 0)
811 			vm_thread_dispose(td);
812 		if (!vm_thread_new(td, pages))
813 			return (ENOMEM);
814 		cpu_thread_alloc(td);
815 	}
816 	kasan_thread_alloc(td);
817 	kmsan_thread_alloc(td);
818 	return (0);
819 }
820 
821 /*
822  * Deallocate a thread.
823  */
824 static void
825 thread_free_batched(struct thread *td)
826 {
827 
828 	lock_profile_thread_exit(td);
829 	if (td->td_cpuset)
830 		cpuset_rel(td->td_cpuset);
831 	td->td_cpuset = NULL;
832 	cpu_thread_free(td);
833 	if (td->td_kstack != 0)
834 		vm_thread_dispose(td);
835 	callout_drain(&td->td_slpcallout);
836 	/*
837 	 * Freeing handled by the caller.
838 	 */
839 	td->td_tid = -1;
840 	kmsan_thread_free(td);
841 	uma_zfree(thread_zone, td);
842 }
843 
844 void
845 thread_free(struct thread *td)
846 {
847 	lwpid_t tid;
848 
849 	EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
850 	tid = td->td_tid;
851 	thread_free_batched(td);
852 	tid_free(tid);
853 	thread_count_dec();
854 }
855 
856 void
857 thread_cow_get_proc(struct thread *newtd, struct proc *p)
858 {
859 
860 	PROC_LOCK_ASSERT(p, MA_OWNED);
861 	newtd->td_realucred = crcowget(p->p_ucred);
862 	newtd->td_ucred = newtd->td_realucred;
863 	newtd->td_limit = lim_hold(p->p_limit);
864 	newtd->td_cowgen = p->p_cowgen;
865 }
866 
867 void
868 thread_cow_get(struct thread *newtd, struct thread *td)
869 {
870 
871 	MPASS(td->td_realucred == td->td_ucred);
872 	newtd->td_realucred = crcowget(td->td_realucred);
873 	newtd->td_ucred = newtd->td_realucred;
874 	newtd->td_limit = lim_hold(td->td_limit);
875 	newtd->td_cowgen = td->td_cowgen;
876 }
877 
878 void
879 thread_cow_free(struct thread *td)
880 {
881 
882 	if (td->td_realucred != NULL)
883 		crcowfree(td);
884 	if (td->td_limit != NULL)
885 		lim_free(td->td_limit);
886 }
887 
888 void
889 thread_cow_update(struct thread *td)
890 {
891 	struct proc *p;
892 	struct ucred *oldcred;
893 	struct plimit *oldlimit;
894 
895 	p = td->td_proc;
896 	PROC_LOCK(p);
897 	oldcred = crcowsync();
898 	oldlimit = lim_cowsync();
899 	td->td_cowgen = p->p_cowgen;
900 	PROC_UNLOCK(p);
901 	if (oldcred != NULL)
902 		crfree(oldcred);
903 	if (oldlimit != NULL)
904 		lim_free(oldlimit);
905 }
906 
907 void
908 thread_cow_synced(struct thread *td)
909 {
910 	struct proc *p;
911 
912 	p = td->td_proc;
913 	PROC_LOCK_ASSERT(p, MA_OWNED);
914 	MPASS(td->td_cowgen != p->p_cowgen);
915 	MPASS(td->td_ucred == p->p_ucred);
916 	MPASS(td->td_limit == p->p_limit);
917 	td->td_cowgen = p->p_cowgen;
918 }
919 
920 /*
921  * Discard the current thread and exit from its context.
922  * Always called with scheduler locked.
923  *
924  * Because we can't free a thread while we're operating under its context,
925  * push the current thread into our CPU's deadthread holder. This means
926  * we needn't worry about someone else grabbing our context before we
927  * do a cpu_throw().
928  */
929 void
930 thread_exit(void)
931 {
932 	uint64_t runtime, new_switchtime;
933 	struct thread *td;
934 	struct thread *td2;
935 	struct proc *p;
936 
937 	td = curthread;
938 	p = td->td_proc;
939 
940 	PROC_SLOCK_ASSERT(p, MA_OWNED);
941 	mtx_assert(&Giant, MA_NOTOWNED);
942 
943 	PROC_LOCK_ASSERT(p, MA_OWNED);
944 	KASSERT(p != NULL, ("thread exiting without a process"));
945 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
946 	    (long)p->p_pid, td->td_name);
947 	SDT_PROBE0(proc, , , lwp__exit);
948 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
949 	MPASS(td->td_realucred == td->td_ucred);
950 
951 	/*
952 	 * drop FPU & debug register state storage, or any other
953 	 * architecture specific resources that
954 	 * would not be on a new untouched process.
955 	 */
956 	cpu_thread_exit(td);
957 
958 	/*
959 	 * The last thread is left attached to the process
960 	 * So that the whole bundle gets recycled. Skip
961 	 * all this stuff if we never had threads.
962 	 * EXIT clears all sign of other threads when
963 	 * it goes to single threading, so the last thread always
964 	 * takes the short path.
965 	 */
966 	if (p->p_flag & P_HADTHREADS) {
967 		if (p->p_numthreads > 1) {
968 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
969 			thread_unlink(td);
970 			td2 = FIRST_THREAD_IN_PROC(p);
971 			sched_exit_thread(td2, td);
972 
973 			/*
974 			 * The test below is NOT true if we are the
975 			 * sole exiting thread. P_STOPPED_SINGLE is unset
976 			 * in exit1() after it is the only survivor.
977 			 */
978 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
979 				if (p->p_numthreads == p->p_suspcount) {
980 					thread_lock(p->p_singlethread);
981 					thread_unsuspend_one(p->p_singlethread,
982 					    p, false);
983 				}
984 			}
985 
986 			PCPU_SET(deadthread, td);
987 		} else {
988 			/*
989 			 * The last thread is exiting.. but not through exit()
990 			 */
991 			panic ("thread_exit: Last thread exiting on its own");
992 		}
993 	}
994 #ifdef	HWPMC_HOOKS
995 	/*
996 	 * If this thread is part of a process that is being tracked by hwpmc(4),
997 	 * inform the module of the thread's impending exit.
998 	 */
999 	if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
1000 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1001 		PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
1002 	} else if (PMC_SYSTEM_SAMPLING_ACTIVE())
1003 		PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
1004 #endif
1005 	PROC_UNLOCK(p);
1006 	PROC_STATLOCK(p);
1007 	thread_lock(td);
1008 	PROC_SUNLOCK(p);
1009 
1010 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
1011 	new_switchtime = cpu_ticks();
1012 	runtime = new_switchtime - PCPU_GET(switchtime);
1013 	td->td_runtime += runtime;
1014 	td->td_incruntime += runtime;
1015 	PCPU_SET(switchtime, new_switchtime);
1016 	PCPU_SET(switchticks, ticks);
1017 	VM_CNT_INC(v_swtch);
1018 
1019 	/* Save our resource usage in our process. */
1020 	td->td_ru.ru_nvcsw++;
1021 	ruxagg_locked(p, td);
1022 	rucollect(&p->p_ru, &td->td_ru);
1023 	PROC_STATUNLOCK(p);
1024 
1025 	TD_SET_STATE(td, TDS_INACTIVE);
1026 #ifdef WITNESS
1027 	witness_thread_exit(td);
1028 #endif
1029 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
1030 	sched_throw(td);
1031 	panic("I'm a teapot!");
1032 	/* NOTREACHED */
1033 }
1034 
1035 /*
1036  * Do any thread specific cleanups that may be needed in wait()
1037  * called with Giant, proc and schedlock not held.
1038  */
1039 void
1040 thread_wait(struct proc *p)
1041 {
1042 	struct thread *td;
1043 
1044 	mtx_assert(&Giant, MA_NOTOWNED);
1045 	KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
1046 	KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
1047 	td = FIRST_THREAD_IN_PROC(p);
1048 	/* Lock the last thread so we spin until it exits cpu_throw(). */
1049 	thread_lock(td);
1050 	thread_unlock(td);
1051 	lock_profile_thread_exit(td);
1052 	cpuset_rel(td->td_cpuset);
1053 	td->td_cpuset = NULL;
1054 	cpu_thread_clean(td);
1055 	thread_cow_free(td);
1056 	callout_drain(&td->td_slpcallout);
1057 	thread_reap();	/* check for zombie threads etc. */
1058 }
1059 
1060 /*
1061  * Link a thread to a process.
1062  * set up anything that needs to be initialized for it to
1063  * be used by the process.
1064  */
1065 void
1066 thread_link(struct thread *td, struct proc *p)
1067 {
1068 
1069 	/*
1070 	 * XXX This can't be enabled because it's called for proc0 before
1071 	 * its lock has been created.
1072 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
1073 	 */
1074 	TD_SET_STATE(td, TDS_INACTIVE);
1075 	td->td_proc     = p;
1076 	td->td_flags    = TDF_INMEM;
1077 
1078 	LIST_INIT(&td->td_contested);
1079 	LIST_INIT(&td->td_lprof[0]);
1080 	LIST_INIT(&td->td_lprof[1]);
1081 #ifdef EPOCH_TRACE
1082 	SLIST_INIT(&td->td_epochs);
1083 #endif
1084 	sigqueue_init(&td->td_sigqueue, p);
1085 	callout_init(&td->td_slpcallout, 1);
1086 	TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
1087 	p->p_numthreads++;
1088 }
1089 
1090 /*
1091  * Called from:
1092  *  thread_exit()
1093  */
1094 void
1095 thread_unlink(struct thread *td)
1096 {
1097 	struct proc *p = td->td_proc;
1098 
1099 	PROC_LOCK_ASSERT(p, MA_OWNED);
1100 #ifdef EPOCH_TRACE
1101 	MPASS(SLIST_EMPTY(&td->td_epochs));
1102 #endif
1103 
1104 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
1105 	p->p_numthreads--;
1106 	/* could clear a few other things here */
1107 	/* Must  NOT clear links to proc! */
1108 }
1109 
1110 static int
1111 calc_remaining(struct proc *p, int mode)
1112 {
1113 	int remaining;
1114 
1115 	PROC_LOCK_ASSERT(p, MA_OWNED);
1116 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1117 	if (mode == SINGLE_EXIT)
1118 		remaining = p->p_numthreads;
1119 	else if (mode == SINGLE_BOUNDARY)
1120 		remaining = p->p_numthreads - p->p_boundary_count;
1121 	else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
1122 		remaining = p->p_numthreads - p->p_suspcount;
1123 	else
1124 		panic("calc_remaining: wrong mode %d", mode);
1125 	return (remaining);
1126 }
1127 
1128 static int
1129 remain_for_mode(int mode)
1130 {
1131 
1132 	return (mode == SINGLE_ALLPROC ? 0 : 1);
1133 }
1134 
1135 static void
1136 weed_inhib(int mode, struct thread *td2, struct proc *p)
1137 {
1138 	PROC_LOCK_ASSERT(p, MA_OWNED);
1139 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1140 	THREAD_LOCK_ASSERT(td2, MA_OWNED);
1141 
1142 	/*
1143 	 * Since the thread lock is dropped by the scheduler we have
1144 	 * to retry to check for races.
1145 	 */
1146 restart:
1147 	switch (mode) {
1148 	case SINGLE_EXIT:
1149 		if (TD_IS_SUSPENDED(td2)) {
1150 			thread_unsuspend_one(td2, p, true);
1151 			thread_lock(td2);
1152 			goto restart;
1153 		}
1154 		if (TD_CAN_ABORT(td2)) {
1155 			sleepq_abort(td2, EINTR);
1156 			return;
1157 		}
1158 		break;
1159 	case SINGLE_BOUNDARY:
1160 	case SINGLE_NO_EXIT:
1161 		if (TD_IS_SUSPENDED(td2) &&
1162 		    (td2->td_flags & TDF_BOUNDARY) == 0) {
1163 			thread_unsuspend_one(td2, p, false);
1164 			thread_lock(td2);
1165 			goto restart;
1166 		}
1167 		if (TD_CAN_ABORT(td2)) {
1168 			sleepq_abort(td2, ERESTART);
1169 			return;
1170 		}
1171 		break;
1172 	case SINGLE_ALLPROC:
1173 		/*
1174 		 * ALLPROC suspend tries to avoid spurious EINTR for
1175 		 * threads sleeping interruptable, by suspending the
1176 		 * thread directly, similarly to sig_suspend_threads().
1177 		 * Since such sleep is not neccessary performed at the user
1178 		 * boundary, TDF_ALLPROCSUSP is used to avoid immediate
1179 		 * un-suspend.
1180 		 */
1181 		if (TD_IS_SUSPENDED(td2) &&
1182 		    (td2->td_flags & TDF_ALLPROCSUSP) == 0) {
1183 			thread_unsuspend_one(td2, p, false);
1184 			thread_lock(td2);
1185 			goto restart;
1186 		}
1187 		if (TD_CAN_ABORT(td2)) {
1188 			td2->td_flags |= TDF_ALLPROCSUSP;
1189 			sleepq_abort(td2, ERESTART);
1190 			return;
1191 		}
1192 		break;
1193 	default:
1194 		break;
1195 	}
1196 	thread_unlock(td2);
1197 }
1198 
1199 /*
1200  * Enforce single-threading.
1201  *
1202  * Returns 1 if the caller must abort (another thread is waiting to
1203  * exit the process or similar). Process is locked!
1204  * Returns 0 when you are successfully the only thread running.
1205  * A process has successfully single threaded in the suspend mode when
1206  * There are no threads in user mode. Threads in the kernel must be
1207  * allowed to continue until they get to the user boundary. They may even
1208  * copy out their return values and data before suspending. They may however be
1209  * accelerated in reaching the user boundary as we will wake up
1210  * any sleeping threads that are interruptable. (PCATCH).
1211  */
1212 int
1213 thread_single(struct proc *p, int mode)
1214 {
1215 	struct thread *td;
1216 	struct thread *td2;
1217 	int remaining;
1218 
1219 	td = curthread;
1220 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1221 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1222 	    ("invalid mode %d", mode));
1223 	/*
1224 	 * If allowing non-ALLPROC singlethreading for non-curproc
1225 	 * callers, calc_remaining() and remain_for_mode() should be
1226 	 * adjusted to also account for td->td_proc != p.  For now
1227 	 * this is not implemented because it is not used.
1228 	 */
1229 	KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
1230 	    (mode != SINGLE_ALLPROC && td->td_proc == p),
1231 	    ("mode %d proc %p curproc %p", mode, p, td->td_proc));
1232 	mtx_assert(&Giant, MA_NOTOWNED);
1233 	PROC_LOCK_ASSERT(p, MA_OWNED);
1234 
1235 	/*
1236 	 * Is someone already single threading?
1237 	 * Or may be singlethreading is not needed at all.
1238 	 */
1239 	if (mode == SINGLE_ALLPROC) {
1240 		while ((p->p_flag & P_STOPPED_SINGLE) != 0) {
1241 			if ((p->p_flag2 & P2_WEXIT) != 0)
1242 				return (1);
1243 			msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0);
1244 		}
1245 		if ((p->p_flag & (P_STOPPED_SIG | P_TRACED)) != 0 ||
1246 		    (p->p_flag2 & P2_WEXIT) != 0)
1247 			return (1);
1248 	} else if ((p->p_flag & P_HADTHREADS) == 0)
1249 		return (0);
1250 	if (p->p_singlethread != NULL && p->p_singlethread != td)
1251 		return (1);
1252 
1253 	if (mode == SINGLE_EXIT) {
1254 		p->p_flag |= P_SINGLE_EXIT;
1255 		p->p_flag &= ~P_SINGLE_BOUNDARY;
1256 	} else {
1257 		p->p_flag &= ~P_SINGLE_EXIT;
1258 		if (mode == SINGLE_BOUNDARY)
1259 			p->p_flag |= P_SINGLE_BOUNDARY;
1260 		else
1261 			p->p_flag &= ~P_SINGLE_BOUNDARY;
1262 	}
1263 	if (mode == SINGLE_ALLPROC)
1264 		p->p_flag |= P_TOTAL_STOP;
1265 	p->p_flag |= P_STOPPED_SINGLE;
1266 	PROC_SLOCK(p);
1267 	p->p_singlethread = td;
1268 	remaining = calc_remaining(p, mode);
1269 	while (remaining != remain_for_mode(mode)) {
1270 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
1271 			goto stopme;
1272 		FOREACH_THREAD_IN_PROC(p, td2) {
1273 			if (td2 == td)
1274 				continue;
1275 			thread_lock(td2);
1276 			ast_sched_locked(td2, TDA_SUSPEND);
1277 			if (TD_IS_INHIBITED(td2)) {
1278 				weed_inhib(mode, td2, p);
1279 #ifdef SMP
1280 			} else if (TD_IS_RUNNING(td2)) {
1281 				forward_signal(td2);
1282 				thread_unlock(td2);
1283 #endif
1284 			} else
1285 				thread_unlock(td2);
1286 		}
1287 		remaining = calc_remaining(p, mode);
1288 
1289 		/*
1290 		 * Maybe we suspended some threads.. was it enough?
1291 		 */
1292 		if (remaining == remain_for_mode(mode))
1293 			break;
1294 
1295 stopme:
1296 		/*
1297 		 * Wake us up when everyone else has suspended.
1298 		 * In the mean time we suspend as well.
1299 		 */
1300 		thread_suspend_switch(td, p);
1301 		remaining = calc_remaining(p, mode);
1302 	}
1303 	if (mode == SINGLE_EXIT) {
1304 		/*
1305 		 * Convert the process to an unthreaded process.  The
1306 		 * SINGLE_EXIT is called by exit1() or execve(), in
1307 		 * both cases other threads must be retired.
1308 		 */
1309 		KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1310 		p->p_singlethread = NULL;
1311 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1312 
1313 		/*
1314 		 * Wait for any remaining threads to exit cpu_throw().
1315 		 */
1316 		while (p->p_exitthreads != 0) {
1317 			PROC_SUNLOCK(p);
1318 			PROC_UNLOCK(p);
1319 			sched_relinquish(td);
1320 			PROC_LOCK(p);
1321 			PROC_SLOCK(p);
1322 		}
1323 	} else if (mode == SINGLE_BOUNDARY) {
1324 		/*
1325 		 * Wait until all suspended threads are removed from
1326 		 * the processors.  The thread_suspend_check()
1327 		 * increments p_boundary_count while it is still
1328 		 * running, which makes it possible for the execve()
1329 		 * to destroy vmspace while our other threads are
1330 		 * still using the address space.
1331 		 *
1332 		 * We lock the thread, which is only allowed to
1333 		 * succeed after context switch code finished using
1334 		 * the address space.
1335 		 */
1336 		FOREACH_THREAD_IN_PROC(p, td2) {
1337 			if (td2 == td)
1338 				continue;
1339 			thread_lock(td2);
1340 			KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1341 			    ("td %p not on boundary", td2));
1342 			KASSERT(TD_IS_SUSPENDED(td2),
1343 			    ("td %p is not suspended", td2));
1344 			thread_unlock(td2);
1345 		}
1346 	}
1347 	PROC_SUNLOCK(p);
1348 	return (0);
1349 }
1350 
1351 bool
1352 thread_suspend_check_needed(void)
1353 {
1354 	struct proc *p;
1355 	struct thread *td;
1356 
1357 	td = curthread;
1358 	p = td->td_proc;
1359 	PROC_LOCK_ASSERT(p, MA_OWNED);
1360 	return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1361 	    (td->td_dbgflags & TDB_SUSPEND) != 0));
1362 }
1363 
1364 /*
1365  * Called in from locations that can safely check to see
1366  * whether we have to suspend or at least throttle for a
1367  * single-thread event (e.g. fork).
1368  *
1369  * Such locations include userret().
1370  * If the "return_instead" argument is non zero, the thread must be able to
1371  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1372  *
1373  * The 'return_instead' argument tells the function if it may do a
1374  * thread_exit() or suspend, or whether the caller must abort and back
1375  * out instead.
1376  *
1377  * If the thread that set the single_threading request has set the
1378  * P_SINGLE_EXIT bit in the process flags then this call will never return
1379  * if 'return_instead' is false, but will exit.
1380  *
1381  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1382  *---------------+--------------------+---------------------
1383  *       0       | returns 0          |   returns 0 or 1
1384  *               | when ST ends       |   immediately
1385  *---------------+--------------------+---------------------
1386  *       1       | thread exits       |   returns 1
1387  *               |                    |  immediately
1388  * 0 = thread_exit() or suspension ok,
1389  * other = return error instead of stopping the thread.
1390  *
1391  * While a full suspension is under effect, even a single threading
1392  * thread would be suspended if it made this call (but it shouldn't).
1393  * This call should only be made from places where
1394  * thread_exit() would be safe as that may be the outcome unless
1395  * return_instead is set.
1396  */
1397 int
1398 thread_suspend_check(int return_instead)
1399 {
1400 	struct thread *td;
1401 	struct proc *p;
1402 
1403 	td = curthread;
1404 	p = td->td_proc;
1405 	mtx_assert(&Giant, MA_NOTOWNED);
1406 	PROC_LOCK_ASSERT(p, MA_OWNED);
1407 	while (thread_suspend_check_needed()) {
1408 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1409 			KASSERT(p->p_singlethread != NULL,
1410 			    ("singlethread not set"));
1411 			/*
1412 			 * The only suspension in action is a
1413 			 * single-threading. Single threader need not stop.
1414 			 * It is safe to access p->p_singlethread unlocked
1415 			 * because it can only be set to our address by us.
1416 			 */
1417 			if (p->p_singlethread == td)
1418 				return (0);	/* Exempt from stopping. */
1419 		}
1420 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1421 			return (EINTR);
1422 
1423 		/* Should we goto user boundary if we didn't come from there? */
1424 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1425 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1426 			return (ERESTART);
1427 
1428 		/*
1429 		 * Ignore suspend requests if they are deferred.
1430 		 */
1431 		if ((td->td_flags & TDF_SBDRY) != 0) {
1432 			KASSERT(return_instead,
1433 			    ("TDF_SBDRY set for unsafe thread_suspend_check"));
1434 			KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1435 			    (TDF_SEINTR | TDF_SERESTART),
1436 			    ("both TDF_SEINTR and TDF_SERESTART"));
1437 			return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1438 		}
1439 
1440 		/*
1441 		 * If the process is waiting for us to exit,
1442 		 * this thread should just suicide.
1443 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1444 		 */
1445 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1446 			PROC_UNLOCK(p);
1447 
1448 			/*
1449 			 * Allow Linux emulation layer to do some work
1450 			 * before thread suicide.
1451 			 */
1452 			if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1453 				(p->p_sysent->sv_thread_detach)(td);
1454 			umtx_thread_exit(td);
1455 			kern_thr_exit(td);
1456 			panic("stopped thread did not exit");
1457 		}
1458 
1459 		PROC_SLOCK(p);
1460 		thread_stopped(p);
1461 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1462 			if (p->p_numthreads == p->p_suspcount + 1) {
1463 				thread_lock(p->p_singlethread);
1464 				thread_unsuspend_one(p->p_singlethread, p,
1465 				    false);
1466 			}
1467 		}
1468 		PROC_UNLOCK(p);
1469 		thread_lock(td);
1470 		/*
1471 		 * When a thread suspends, it just
1472 		 * gets taken off all queues.
1473 		 */
1474 		thread_suspend_one(td);
1475 		if (return_instead == 0) {
1476 			p->p_boundary_count++;
1477 			td->td_flags |= TDF_BOUNDARY;
1478 		}
1479 		PROC_SUNLOCK(p);
1480 		mi_switch(SW_INVOL | SWT_SUSPEND);
1481 		PROC_LOCK(p);
1482 	}
1483 	return (0);
1484 }
1485 
1486 /*
1487  * Check for possible stops and suspensions while executing a
1488  * casueword or similar transiently failing operation.
1489  *
1490  * The sleep argument controls whether the function can handle a stop
1491  * request itself or it should return ERESTART and the request is
1492  * proceed at the kernel/user boundary in ast.
1493  *
1494  * Typically, when retrying due to casueword(9) failure (rv == 1), we
1495  * should handle the stop requests there, with exception of cases when
1496  * the thread owns a kernel resource, for instance busied the umtx
1497  * key, or when functions return immediately if thread_check_susp()
1498  * returned non-zero.  On the other hand, retrying the whole lock
1499  * operation, we better not stop there but delegate the handling to
1500  * ast.
1501  *
1502  * If the request is for thread termination P_SINGLE_EXIT, we cannot
1503  * handle it at all, and simply return EINTR.
1504  */
1505 int
1506 thread_check_susp(struct thread *td, bool sleep)
1507 {
1508 	struct proc *p;
1509 	int error;
1510 
1511 	/*
1512 	 * The check for TDA_SUSPEND is racy, but it is enough to
1513 	 * eventually break the lockstep loop.
1514 	 */
1515 	if (!td_ast_pending(td, TDA_SUSPEND))
1516 		return (0);
1517 	error = 0;
1518 	p = td->td_proc;
1519 	PROC_LOCK(p);
1520 	if (p->p_flag & P_SINGLE_EXIT)
1521 		error = EINTR;
1522 	else if (P_SHOULDSTOP(p) ||
1523 	    ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1524 		error = sleep ? thread_suspend_check(0) : ERESTART;
1525 	PROC_UNLOCK(p);
1526 	return (error);
1527 }
1528 
1529 void
1530 thread_suspend_switch(struct thread *td, struct proc *p)
1531 {
1532 
1533 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1534 	PROC_LOCK_ASSERT(p, MA_OWNED);
1535 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1536 	/*
1537 	 * We implement thread_suspend_one in stages here to avoid
1538 	 * dropping the proc lock while the thread lock is owned.
1539 	 */
1540 	if (p == td->td_proc) {
1541 		thread_stopped(p);
1542 		p->p_suspcount++;
1543 	}
1544 	PROC_UNLOCK(p);
1545 	thread_lock(td);
1546 	ast_unsched_locked(td, TDA_SUSPEND);
1547 	TD_SET_SUSPENDED(td);
1548 	sched_sleep(td, 0);
1549 	PROC_SUNLOCK(p);
1550 	DROP_GIANT();
1551 	mi_switch(SW_VOL | SWT_SUSPEND);
1552 	PICKUP_GIANT();
1553 	PROC_LOCK(p);
1554 	PROC_SLOCK(p);
1555 }
1556 
1557 void
1558 thread_suspend_one(struct thread *td)
1559 {
1560 	struct proc *p;
1561 
1562 	p = td->td_proc;
1563 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1564 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1565 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1566 	p->p_suspcount++;
1567 	ast_unsched_locked(td, TDA_SUSPEND);
1568 	TD_SET_SUSPENDED(td);
1569 	sched_sleep(td, 0);
1570 }
1571 
1572 static void
1573 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1574 {
1575 
1576 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1577 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1578 	TD_CLR_SUSPENDED(td);
1579 	td->td_flags &= ~TDF_ALLPROCSUSP;
1580 	if (td->td_proc == p) {
1581 		PROC_SLOCK_ASSERT(p, MA_OWNED);
1582 		p->p_suspcount--;
1583 		if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1584 			td->td_flags &= ~TDF_BOUNDARY;
1585 			p->p_boundary_count--;
1586 		}
1587 	}
1588 	setrunnable(td, 0);
1589 }
1590 
1591 void
1592 thread_run_flash(struct thread *td)
1593 {
1594 	struct proc *p;
1595 
1596 	p = td->td_proc;
1597 	PROC_LOCK_ASSERT(p, MA_OWNED);
1598 
1599 	if (TD_ON_SLEEPQ(td))
1600 		sleepq_remove_nested(td);
1601 	else
1602 		thread_lock(td);
1603 
1604 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1605 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1606 
1607 	TD_CLR_SUSPENDED(td);
1608 	PROC_SLOCK(p);
1609 	MPASS(p->p_suspcount > 0);
1610 	p->p_suspcount--;
1611 	PROC_SUNLOCK(p);
1612 	setrunnable(td, 0);
1613 }
1614 
1615 /*
1616  * Allow all threads blocked by single threading to continue running.
1617  */
1618 void
1619 thread_unsuspend(struct proc *p)
1620 {
1621 	struct thread *td;
1622 
1623 	PROC_LOCK_ASSERT(p, MA_OWNED);
1624 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1625 	if (!P_SHOULDSTOP(p)) {
1626                 FOREACH_THREAD_IN_PROC(p, td) {
1627 			thread_lock(td);
1628 			if (TD_IS_SUSPENDED(td))
1629 				thread_unsuspend_one(td, p, true);
1630 			else
1631 				thread_unlock(td);
1632 		}
1633 	} else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1634 	    p->p_numthreads == p->p_suspcount) {
1635 		/*
1636 		 * Stopping everything also did the job for the single
1637 		 * threading request. Now we've downgraded to single-threaded,
1638 		 * let it continue.
1639 		 */
1640 		if (p->p_singlethread->td_proc == p) {
1641 			thread_lock(p->p_singlethread);
1642 			thread_unsuspend_one(p->p_singlethread, p, false);
1643 		}
1644 	}
1645 }
1646 
1647 /*
1648  * End the single threading mode..
1649  */
1650 void
1651 thread_single_end(struct proc *p, int mode)
1652 {
1653 	struct thread *td;
1654 
1655 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1656 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1657 	    ("invalid mode %d", mode));
1658 	PROC_LOCK_ASSERT(p, MA_OWNED);
1659 	KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1660 	    (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1661 	    ("mode %d does not match P_TOTAL_STOP", mode));
1662 	KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1663 	    ("thread_single_end from other thread %p %p",
1664 	    curthread, p->p_singlethread));
1665 	KASSERT(mode != SINGLE_BOUNDARY ||
1666 	    (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1667 	    ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1668 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1669 	    P_TOTAL_STOP);
1670 	PROC_SLOCK(p);
1671 	p->p_singlethread = NULL;
1672 
1673 	/*
1674 	 * If there are other threads they may now run,
1675 	 * unless of course there is a blanket 'stop order'
1676 	 * on the process. The single threader must be allowed
1677 	 * to continue however as this is a bad place to stop.
1678 	 */
1679 	if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1680                 FOREACH_THREAD_IN_PROC(p, td) {
1681 			thread_lock(td);
1682 			if (TD_IS_SUSPENDED(td))
1683 				thread_unsuspend_one(td, p, true);
1684 			else
1685 				thread_unlock(td);
1686 		}
1687 	}
1688 	KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1689 	    ("inconsistent boundary count %d", p->p_boundary_count));
1690 	PROC_SUNLOCK(p);
1691 	wakeup(&p->p_flag);
1692 }
1693 
1694 /*
1695  * Locate a thread by number and return with proc lock held.
1696  *
1697  * thread exit establishes proc -> tidhash lock ordering, but lookup
1698  * takes tidhash first and needs to return locked proc.
1699  *
1700  * The problem is worked around by relying on type-safety of both
1701  * structures and doing the work in 2 steps:
1702  * - tidhash-locked lookup which saves both thread and proc pointers
1703  * - proc-locked verification that the found thread still matches
1704  */
1705 static bool
1706 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1707 {
1708 #define RUN_THRESH	16
1709 	struct proc *p;
1710 	struct thread *td;
1711 	int run;
1712 	bool locked;
1713 
1714 	run = 0;
1715 	rw_rlock(TIDHASHLOCK(tid));
1716 	locked = true;
1717 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1718 		if (td->td_tid != tid) {
1719 			run++;
1720 			continue;
1721 		}
1722 		p = td->td_proc;
1723 		if (pid != -1 && p->p_pid != pid) {
1724 			td = NULL;
1725 			break;
1726 		}
1727 		if (run > RUN_THRESH) {
1728 			if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1729 				LIST_REMOVE(td, td_hash);
1730 				LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1731 					td, td_hash);
1732 				rw_wunlock(TIDHASHLOCK(tid));
1733 				locked = false;
1734 				break;
1735 			}
1736 		}
1737 		break;
1738 	}
1739 	if (locked)
1740 		rw_runlock(TIDHASHLOCK(tid));
1741 	if (td == NULL)
1742 		return (false);
1743 	*pp = p;
1744 	*tdp = td;
1745 	return (true);
1746 }
1747 
1748 struct thread *
1749 tdfind(lwpid_t tid, pid_t pid)
1750 {
1751 	struct proc *p;
1752 	struct thread *td;
1753 
1754 	td = curthread;
1755 	if (td->td_tid == tid) {
1756 		if (pid != -1 && td->td_proc->p_pid != pid)
1757 			return (NULL);
1758 		PROC_LOCK(td->td_proc);
1759 		return (td);
1760 	}
1761 
1762 	for (;;) {
1763 		if (!tdfind_hash(tid, pid, &p, &td))
1764 			return (NULL);
1765 		PROC_LOCK(p);
1766 		if (td->td_tid != tid) {
1767 			PROC_UNLOCK(p);
1768 			continue;
1769 		}
1770 		if (td->td_proc != p) {
1771 			PROC_UNLOCK(p);
1772 			continue;
1773 		}
1774 		if (p->p_state == PRS_NEW) {
1775 			PROC_UNLOCK(p);
1776 			return (NULL);
1777 		}
1778 		return (td);
1779 	}
1780 }
1781 
1782 void
1783 tidhash_add(struct thread *td)
1784 {
1785 	rw_wlock(TIDHASHLOCK(td->td_tid));
1786 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1787 	rw_wunlock(TIDHASHLOCK(td->td_tid));
1788 }
1789 
1790 void
1791 tidhash_remove(struct thread *td)
1792 {
1793 
1794 	rw_wlock(TIDHASHLOCK(td->td_tid));
1795 	LIST_REMOVE(td, td_hash);
1796 	rw_wunlock(TIDHASHLOCK(td->td_tid));
1797 }
1798