xref: /freebsd/sys/kern/kern_thread.c (revision 18849b5da0c5eaa88500b457be05b038813b51b1)
1 /*-
2  * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
3  *  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice(s), this list of conditions and the following disclaimer as
10  *    the first lines of this file unmodified other than the possible
11  *    addition of one or more copyright notices.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice(s), this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
26  * DAMAGE.
27  */
28 
29 #include "opt_witness.h"
30 #include "opt_hwpmc_hooks.h"
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/rangelock.h>
42 #include <sys/resourcevar.h>
43 #include <sys/sdt.h>
44 #include <sys/smp.h>
45 #include <sys/sched.h>
46 #include <sys/sleepqueue.h>
47 #include <sys/selinfo.h>
48 #include <sys/syscallsubr.h>
49 #include <sys/sysent.h>
50 #include <sys/turnstile.h>
51 #include <sys/ktr.h>
52 #include <sys/rwlock.h>
53 #include <sys/umtx.h>
54 #include <sys/cpuset.h>
55 #ifdef	HWPMC_HOOKS
56 #include <sys/pmckern.h>
57 #endif
58 
59 #include <security/audit/audit.h>
60 
61 #include <vm/vm.h>
62 #include <vm/vm_extern.h>
63 #include <vm/uma.h>
64 #include <vm/vm_domain.h>
65 #include <sys/eventhandler.h>
66 
67 SDT_PROVIDER_DECLARE(proc);
68 SDT_PROBE_DEFINE(proc, , , lwp__exit);
69 
70 /*
71  * thread related storage.
72  */
73 static uma_zone_t thread_zone;
74 
75 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
76 static struct mtx zombie_lock;
77 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
78 
79 static void thread_zombie(struct thread *);
80 static int thread_unsuspend_one(struct thread *td, struct proc *p,
81     bool boundary);
82 
83 #define TID_BUFFER_SIZE	1024
84 
85 struct mtx tid_lock;
86 static struct unrhdr *tid_unrhdr;
87 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
88 static int tid_head, tid_tail;
89 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
90 
91 struct	tidhashhead *tidhashtbl;
92 u_long	tidhash;
93 struct	rwlock tidhash_lock;
94 
95 static lwpid_t
96 tid_alloc(void)
97 {
98 	lwpid_t	tid;
99 
100 	tid = alloc_unr(tid_unrhdr);
101 	if (tid != -1)
102 		return (tid);
103 	mtx_lock(&tid_lock);
104 	if (tid_head == tid_tail) {
105 		mtx_unlock(&tid_lock);
106 		return (-1);
107 	}
108 	tid = tid_buffer[tid_head];
109 	tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
110 	mtx_unlock(&tid_lock);
111 	return (tid);
112 }
113 
114 static void
115 tid_free(lwpid_t tid)
116 {
117 	lwpid_t tmp_tid = -1;
118 
119 	mtx_lock(&tid_lock);
120 	if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
121 		tmp_tid = tid_buffer[tid_head];
122 		tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
123 	}
124 	tid_buffer[tid_tail] = tid;
125 	tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE;
126 	mtx_unlock(&tid_lock);
127 	if (tmp_tid != -1)
128 		free_unr(tid_unrhdr, tmp_tid);
129 }
130 
131 /*
132  * Prepare a thread for use.
133  */
134 static int
135 thread_ctor(void *mem, int size, void *arg, int flags)
136 {
137 	struct thread	*td;
138 
139 	td = (struct thread *)mem;
140 	td->td_state = TDS_INACTIVE;
141 	td->td_oncpu = NOCPU;
142 
143 	td->td_tid = tid_alloc();
144 
145 	/*
146 	 * Note that td_critnest begins life as 1 because the thread is not
147 	 * running and is thereby implicitly waiting to be on the receiving
148 	 * end of a context switch.
149 	 */
150 	td->td_critnest = 1;
151 	td->td_lend_user_pri = PRI_MAX;
152 	EVENTHANDLER_INVOKE(thread_ctor, td);
153 #ifdef AUDIT
154 	audit_thread_alloc(td);
155 #endif
156 	umtx_thread_alloc(td);
157 	return (0);
158 }
159 
160 /*
161  * Reclaim a thread after use.
162  */
163 static void
164 thread_dtor(void *mem, int size, void *arg)
165 {
166 	struct thread *td;
167 
168 	td = (struct thread *)mem;
169 
170 #ifdef INVARIANTS
171 	/* Verify that this thread is in a safe state to free. */
172 	switch (td->td_state) {
173 	case TDS_INHIBITED:
174 	case TDS_RUNNING:
175 	case TDS_CAN_RUN:
176 	case TDS_RUNQ:
177 		/*
178 		 * We must never unlink a thread that is in one of
179 		 * these states, because it is currently active.
180 		 */
181 		panic("bad state for thread unlinking");
182 		/* NOTREACHED */
183 	case TDS_INACTIVE:
184 		break;
185 	default:
186 		panic("bad thread state");
187 		/* NOTREACHED */
188 	}
189 #endif
190 #ifdef AUDIT
191 	audit_thread_free(td);
192 #endif
193 	/* Free all OSD associated to this thread. */
194 	osd_thread_exit(td);
195 
196 	EVENTHANDLER_INVOKE(thread_dtor, td);
197 	tid_free(td->td_tid);
198 }
199 
200 /*
201  * Initialize type-stable parts of a thread (when newly created).
202  */
203 static int
204 thread_init(void *mem, int size, int flags)
205 {
206 	struct thread *td;
207 
208 	td = (struct thread *)mem;
209 
210 	td->td_sleepqueue = sleepq_alloc();
211 	td->td_turnstile = turnstile_alloc();
212 	td->td_rlqe = NULL;
213 	EVENTHANDLER_INVOKE(thread_init, td);
214 	td->td_sched = (struct td_sched *)&td[1];
215 	umtx_thread_init(td);
216 	td->td_kstack = 0;
217 	td->td_sel = NULL;
218 	return (0);
219 }
220 
221 /*
222  * Tear down type-stable parts of a thread (just before being discarded).
223  */
224 static void
225 thread_fini(void *mem, int size)
226 {
227 	struct thread *td;
228 
229 	td = (struct thread *)mem;
230 	EVENTHANDLER_INVOKE(thread_fini, td);
231 	rlqentry_free(td->td_rlqe);
232 	turnstile_free(td->td_turnstile);
233 	sleepq_free(td->td_sleepqueue);
234 	umtx_thread_fini(td);
235 	seltdfini(td);
236 }
237 
238 /*
239  * For a newly created process,
240  * link up all the structures and its initial threads etc.
241  * called from:
242  * {arch}/{arch}/machdep.c   {arch}_init(), init386() etc.
243  * proc_dtor() (should go away)
244  * proc_init()
245  */
246 void
247 proc_linkup0(struct proc *p, struct thread *td)
248 {
249 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
250 	proc_linkup(p, td);
251 }
252 
253 void
254 proc_linkup(struct proc *p, struct thread *td)
255 {
256 
257 	sigqueue_init(&p->p_sigqueue, p);
258 	p->p_ksi = ksiginfo_alloc(1);
259 	if (p->p_ksi != NULL) {
260 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
261 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
262 	}
263 	LIST_INIT(&p->p_mqnotifier);
264 	p->p_numthreads = 0;
265 	thread_link(td, p);
266 }
267 
268 /*
269  * Initialize global thread allocation resources.
270  */
271 void
272 threadinit(void)
273 {
274 
275 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
276 
277 	/*
278 	 * pid_max cannot be greater than PID_MAX.
279 	 * leave one number for thread0.
280 	 */
281 	tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
282 
283 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
284 	    thread_ctor, thread_dtor, thread_init, thread_fini,
285 	    16 - 1, UMA_ZONE_NOFREE);
286 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
287 	rw_init(&tidhash_lock, "tidhash");
288 }
289 
290 /*
291  * Place an unused thread on the zombie list.
292  * Use the slpq as that must be unused by now.
293  */
294 void
295 thread_zombie(struct thread *td)
296 {
297 	mtx_lock_spin(&zombie_lock);
298 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
299 	mtx_unlock_spin(&zombie_lock);
300 }
301 
302 /*
303  * Release a thread that has exited after cpu_throw().
304  */
305 void
306 thread_stash(struct thread *td)
307 {
308 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
309 	thread_zombie(td);
310 }
311 
312 /*
313  * Reap zombie resources.
314  */
315 void
316 thread_reap(void)
317 {
318 	struct thread *td_first, *td_next;
319 
320 	/*
321 	 * Don't even bother to lock if none at this instant,
322 	 * we really don't care about the next instant..
323 	 */
324 	if (!TAILQ_EMPTY(&zombie_threads)) {
325 		mtx_lock_spin(&zombie_lock);
326 		td_first = TAILQ_FIRST(&zombie_threads);
327 		if (td_first)
328 			TAILQ_INIT(&zombie_threads);
329 		mtx_unlock_spin(&zombie_lock);
330 		while (td_first) {
331 			td_next = TAILQ_NEXT(td_first, td_slpq);
332 			thread_cow_free(td_first);
333 			thread_free(td_first);
334 			td_first = td_next;
335 		}
336 	}
337 }
338 
339 /*
340  * Allocate a thread.
341  */
342 struct thread *
343 thread_alloc(int pages)
344 {
345 	struct thread *td;
346 
347 	thread_reap(); /* check if any zombies to get */
348 
349 	td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
350 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
351 	if (!vm_thread_new(td, pages)) {
352 		uma_zfree(thread_zone, td);
353 		return (NULL);
354 	}
355 	cpu_thread_alloc(td);
356 	vm_domain_policy_init(&td->td_vm_dom_policy);
357 	return (td);
358 }
359 
360 int
361 thread_alloc_stack(struct thread *td, int pages)
362 {
363 
364 	KASSERT(td->td_kstack == 0,
365 	    ("thread_alloc_stack called on a thread with kstack"));
366 	if (!vm_thread_new(td, pages))
367 		return (0);
368 	cpu_thread_alloc(td);
369 	return (1);
370 }
371 
372 /*
373  * Deallocate a thread.
374  */
375 void
376 thread_free(struct thread *td)
377 {
378 
379 	lock_profile_thread_exit(td);
380 	if (td->td_cpuset)
381 		cpuset_rel(td->td_cpuset);
382 	td->td_cpuset = NULL;
383 	cpu_thread_free(td);
384 	if (td->td_kstack != 0)
385 		vm_thread_dispose(td);
386 	vm_domain_policy_cleanup(&td->td_vm_dom_policy);
387 	uma_zfree(thread_zone, td);
388 }
389 
390 void
391 thread_cow_get_proc(struct thread *newtd, struct proc *p)
392 {
393 
394 	PROC_LOCK_ASSERT(p, MA_OWNED);
395 	newtd->td_ucred = crhold(p->p_ucred);
396 	newtd->td_limit = lim_hold(p->p_limit);
397 	newtd->td_cowgen = p->p_cowgen;
398 }
399 
400 void
401 thread_cow_get(struct thread *newtd, struct thread *td)
402 {
403 
404 	newtd->td_ucred = crhold(td->td_ucred);
405 	newtd->td_limit = lim_hold(td->td_limit);
406 	newtd->td_cowgen = td->td_cowgen;
407 }
408 
409 void
410 thread_cow_free(struct thread *td)
411 {
412 
413 	if (td->td_ucred != NULL)
414 		crfree(td->td_ucred);
415 	if (td->td_limit != NULL)
416 		lim_free(td->td_limit);
417 }
418 
419 void
420 thread_cow_update(struct thread *td)
421 {
422 	struct proc *p;
423 	struct ucred *oldcred;
424 	struct plimit *oldlimit;
425 
426 	p = td->td_proc;
427 	oldcred = NULL;
428 	oldlimit = NULL;
429 	PROC_LOCK(p);
430 	if (td->td_ucred != p->p_ucred) {
431 		oldcred = td->td_ucred;
432 		td->td_ucred = crhold(p->p_ucred);
433 	}
434 	if (td->td_limit != p->p_limit) {
435 		oldlimit = td->td_limit;
436 		td->td_limit = lim_hold(p->p_limit);
437 	}
438 	td->td_cowgen = p->p_cowgen;
439 	PROC_UNLOCK(p);
440 	if (oldcred != NULL)
441 		crfree(oldcred);
442 	if (oldlimit != NULL)
443 		lim_free(oldlimit);
444 }
445 
446 /*
447  * Discard the current thread and exit from its context.
448  * Always called with scheduler locked.
449  *
450  * Because we can't free a thread while we're operating under its context,
451  * push the current thread into our CPU's deadthread holder. This means
452  * we needn't worry about someone else grabbing our context before we
453  * do a cpu_throw().
454  */
455 void
456 thread_exit(void)
457 {
458 	uint64_t runtime, new_switchtime;
459 	struct thread *td;
460 	struct thread *td2;
461 	struct proc *p;
462 	int wakeup_swapper;
463 
464 	td = curthread;
465 	p = td->td_proc;
466 
467 	PROC_SLOCK_ASSERT(p, MA_OWNED);
468 	mtx_assert(&Giant, MA_NOTOWNED);
469 
470 	PROC_LOCK_ASSERT(p, MA_OWNED);
471 	KASSERT(p != NULL, ("thread exiting without a process"));
472 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
473 	    (long)p->p_pid, td->td_name);
474 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
475 
476 #ifdef AUDIT
477 	AUDIT_SYSCALL_EXIT(0, td);
478 #endif
479 	/*
480 	 * drop FPU & debug register state storage, or any other
481 	 * architecture specific resources that
482 	 * would not be on a new untouched process.
483 	 */
484 	cpu_thread_exit(td);	/* XXXSMP */
485 
486 	/*
487 	 * The last thread is left attached to the process
488 	 * So that the whole bundle gets recycled. Skip
489 	 * all this stuff if we never had threads.
490 	 * EXIT clears all sign of other threads when
491 	 * it goes to single threading, so the last thread always
492 	 * takes the short path.
493 	 */
494 	if (p->p_flag & P_HADTHREADS) {
495 		if (p->p_numthreads > 1) {
496 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
497 			thread_unlink(td);
498 			td2 = FIRST_THREAD_IN_PROC(p);
499 			sched_exit_thread(td2, td);
500 
501 			/*
502 			 * The test below is NOT true if we are the
503 			 * sole exiting thread. P_STOPPED_SINGLE is unset
504 			 * in exit1() after it is the only survivor.
505 			 */
506 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
507 				if (p->p_numthreads == p->p_suspcount) {
508 					thread_lock(p->p_singlethread);
509 					wakeup_swapper = thread_unsuspend_one(
510 						p->p_singlethread, p, false);
511 					thread_unlock(p->p_singlethread);
512 					if (wakeup_swapper)
513 						kick_proc0();
514 				}
515 			}
516 
517 			PCPU_SET(deadthread, td);
518 		} else {
519 			/*
520 			 * The last thread is exiting.. but not through exit()
521 			 */
522 			panic ("thread_exit: Last thread exiting on its own");
523 		}
524 	}
525 #ifdef	HWPMC_HOOKS
526 	/*
527 	 * If this thread is part of a process that is being tracked by hwpmc(4),
528 	 * inform the module of the thread's impending exit.
529 	 */
530 	if (PMC_PROC_IS_USING_PMCS(td->td_proc))
531 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
532 #endif
533 	PROC_UNLOCK(p);
534 	PROC_STATLOCK(p);
535 	thread_lock(td);
536 	PROC_SUNLOCK(p);
537 
538 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
539 	new_switchtime = cpu_ticks();
540 	runtime = new_switchtime - PCPU_GET(switchtime);
541 	td->td_runtime += runtime;
542 	td->td_incruntime += runtime;
543 	PCPU_SET(switchtime, new_switchtime);
544 	PCPU_SET(switchticks, ticks);
545 	PCPU_INC(cnt.v_swtch);
546 
547 	/* Save our resource usage in our process. */
548 	td->td_ru.ru_nvcsw++;
549 	ruxagg(p, td);
550 	rucollect(&p->p_ru, &td->td_ru);
551 	PROC_STATUNLOCK(p);
552 
553 	td->td_state = TDS_INACTIVE;
554 #ifdef WITNESS
555 	witness_thread_exit(td);
556 #endif
557 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
558 	sched_throw(td);
559 	panic("I'm a teapot!");
560 	/* NOTREACHED */
561 }
562 
563 /*
564  * Do any thread specific cleanups that may be needed in wait()
565  * called with Giant, proc and schedlock not held.
566  */
567 void
568 thread_wait(struct proc *p)
569 {
570 	struct thread *td;
571 
572 	mtx_assert(&Giant, MA_NOTOWNED);
573 	KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
574 	KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
575 	td = FIRST_THREAD_IN_PROC(p);
576 	/* Lock the last thread so we spin until it exits cpu_throw(). */
577 	thread_lock(td);
578 	thread_unlock(td);
579 	lock_profile_thread_exit(td);
580 	cpuset_rel(td->td_cpuset);
581 	td->td_cpuset = NULL;
582 	cpu_thread_clean(td);
583 	thread_cow_free(td);
584 	thread_reap();	/* check for zombie threads etc. */
585 }
586 
587 /*
588  * Link a thread to a process.
589  * set up anything that needs to be initialized for it to
590  * be used by the process.
591  */
592 void
593 thread_link(struct thread *td, struct proc *p)
594 {
595 
596 	/*
597 	 * XXX This can't be enabled because it's called for proc0 before
598 	 * its lock has been created.
599 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
600 	 */
601 	td->td_state    = TDS_INACTIVE;
602 	td->td_proc     = p;
603 	td->td_flags    = TDF_INMEM;
604 
605 	LIST_INIT(&td->td_contested);
606 	LIST_INIT(&td->td_lprof[0]);
607 	LIST_INIT(&td->td_lprof[1]);
608 	sigqueue_init(&td->td_sigqueue, p);
609 	callout_init(&td->td_slpcallout, 1);
610 	TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
611 	p->p_numthreads++;
612 }
613 
614 /*
615  * Called from:
616  *  thread_exit()
617  */
618 void
619 thread_unlink(struct thread *td)
620 {
621 	struct proc *p = td->td_proc;
622 
623 	PROC_LOCK_ASSERT(p, MA_OWNED);
624 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
625 	p->p_numthreads--;
626 	/* could clear a few other things here */
627 	/* Must  NOT clear links to proc! */
628 }
629 
630 static int
631 calc_remaining(struct proc *p, int mode)
632 {
633 	int remaining;
634 
635 	PROC_LOCK_ASSERT(p, MA_OWNED);
636 	PROC_SLOCK_ASSERT(p, MA_OWNED);
637 	if (mode == SINGLE_EXIT)
638 		remaining = p->p_numthreads;
639 	else if (mode == SINGLE_BOUNDARY)
640 		remaining = p->p_numthreads - p->p_boundary_count;
641 	else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
642 		remaining = p->p_numthreads - p->p_suspcount;
643 	else
644 		panic("calc_remaining: wrong mode %d", mode);
645 	return (remaining);
646 }
647 
648 static int
649 remain_for_mode(int mode)
650 {
651 
652 	return (mode == SINGLE_ALLPROC ? 0 : 1);
653 }
654 
655 static int
656 weed_inhib(int mode, struct thread *td2, struct proc *p)
657 {
658 	int wakeup_swapper;
659 
660 	PROC_LOCK_ASSERT(p, MA_OWNED);
661 	PROC_SLOCK_ASSERT(p, MA_OWNED);
662 	THREAD_LOCK_ASSERT(td2, MA_OWNED);
663 
664 	wakeup_swapper = 0;
665 	switch (mode) {
666 	case SINGLE_EXIT:
667 		if (TD_IS_SUSPENDED(td2))
668 			wakeup_swapper |= thread_unsuspend_one(td2, p, true);
669 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0)
670 			wakeup_swapper |= sleepq_abort(td2, EINTR);
671 		break;
672 	case SINGLE_BOUNDARY:
673 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & TDF_BOUNDARY) == 0)
674 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
675 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0)
676 			wakeup_swapper |= sleepq_abort(td2, ERESTART);
677 		break;
678 	case SINGLE_NO_EXIT:
679 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & TDF_BOUNDARY) == 0)
680 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
681 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0)
682 			wakeup_swapper |= sleepq_abort(td2, ERESTART);
683 		break;
684 	case SINGLE_ALLPROC:
685 		/*
686 		 * ALLPROC suspend tries to avoid spurious EINTR for
687 		 * threads sleeping interruptable, by suspending the
688 		 * thread directly, similarly to sig_suspend_threads().
689 		 * Since such sleep is not performed at the user
690 		 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
691 		 * is used to avoid immediate un-suspend.
692 		 */
693 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
694 		    TDF_ALLPROCSUSP)) == 0)
695 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
696 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0) {
697 			if ((td2->td_flags & TDF_SBDRY) == 0) {
698 				thread_suspend_one(td2);
699 				td2->td_flags |= TDF_ALLPROCSUSP;
700 			} else {
701 				wakeup_swapper |= sleepq_abort(td2, ERESTART);
702 			}
703 		}
704 		break;
705 	}
706 	return (wakeup_swapper);
707 }
708 
709 /*
710  * Enforce single-threading.
711  *
712  * Returns 1 if the caller must abort (another thread is waiting to
713  * exit the process or similar). Process is locked!
714  * Returns 0 when you are successfully the only thread running.
715  * A process has successfully single threaded in the suspend mode when
716  * There are no threads in user mode. Threads in the kernel must be
717  * allowed to continue until they get to the user boundary. They may even
718  * copy out their return values and data before suspending. They may however be
719  * accelerated in reaching the user boundary as we will wake up
720  * any sleeping threads that are interruptable. (PCATCH).
721  */
722 int
723 thread_single(struct proc *p, int mode)
724 {
725 	struct thread *td;
726 	struct thread *td2;
727 	int remaining, wakeup_swapper;
728 
729 	td = curthread;
730 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
731 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
732 	    ("invalid mode %d", mode));
733 	/*
734 	 * If allowing non-ALLPROC singlethreading for non-curproc
735 	 * callers, calc_remaining() and remain_for_mode() should be
736 	 * adjusted to also account for td->td_proc != p.  For now
737 	 * this is not implemented because it is not used.
738 	 */
739 	KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
740 	    (mode != SINGLE_ALLPROC && td->td_proc == p),
741 	    ("mode %d proc %p curproc %p", mode, p, td->td_proc));
742 	mtx_assert(&Giant, MA_NOTOWNED);
743 	PROC_LOCK_ASSERT(p, MA_OWNED);
744 
745 	if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
746 		return (0);
747 
748 	/* Is someone already single threading? */
749 	if (p->p_singlethread != NULL && p->p_singlethread != td)
750 		return (1);
751 
752 	if (mode == SINGLE_EXIT) {
753 		p->p_flag |= P_SINGLE_EXIT;
754 		p->p_flag &= ~P_SINGLE_BOUNDARY;
755 	} else {
756 		p->p_flag &= ~P_SINGLE_EXIT;
757 		if (mode == SINGLE_BOUNDARY)
758 			p->p_flag |= P_SINGLE_BOUNDARY;
759 		else
760 			p->p_flag &= ~P_SINGLE_BOUNDARY;
761 	}
762 	if (mode == SINGLE_ALLPROC)
763 		p->p_flag |= P_TOTAL_STOP;
764 	p->p_flag |= P_STOPPED_SINGLE;
765 	PROC_SLOCK(p);
766 	p->p_singlethread = td;
767 	remaining = calc_remaining(p, mode);
768 	while (remaining != remain_for_mode(mode)) {
769 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
770 			goto stopme;
771 		wakeup_swapper = 0;
772 		FOREACH_THREAD_IN_PROC(p, td2) {
773 			if (td2 == td)
774 				continue;
775 			thread_lock(td2);
776 			td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
777 			if (TD_IS_INHIBITED(td2)) {
778 				wakeup_swapper |= weed_inhib(mode, td2, p);
779 #ifdef SMP
780 			} else if (TD_IS_RUNNING(td2) && td != td2) {
781 				forward_signal(td2);
782 #endif
783 			}
784 			thread_unlock(td2);
785 		}
786 		if (wakeup_swapper)
787 			kick_proc0();
788 		remaining = calc_remaining(p, mode);
789 
790 		/*
791 		 * Maybe we suspended some threads.. was it enough?
792 		 */
793 		if (remaining == remain_for_mode(mode))
794 			break;
795 
796 stopme:
797 		/*
798 		 * Wake us up when everyone else has suspended.
799 		 * In the mean time we suspend as well.
800 		 */
801 		thread_suspend_switch(td, p);
802 		remaining = calc_remaining(p, mode);
803 	}
804 	if (mode == SINGLE_EXIT) {
805 		/*
806 		 * Convert the process to an unthreaded process.  The
807 		 * SINGLE_EXIT is called by exit1() or execve(), in
808 		 * both cases other threads must be retired.
809 		 */
810 		KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
811 		p->p_singlethread = NULL;
812 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
813 
814 		/*
815 		 * Wait for any remaining threads to exit cpu_throw().
816 		 */
817 		while (p->p_exitthreads != 0) {
818 			PROC_SUNLOCK(p);
819 			PROC_UNLOCK(p);
820 			sched_relinquish(td);
821 			PROC_LOCK(p);
822 			PROC_SLOCK(p);
823 		}
824 	} else if (mode == SINGLE_BOUNDARY) {
825 		/*
826 		 * Wait until all suspended threads are removed from
827 		 * the processors.  The thread_suspend_check()
828 		 * increments p_boundary_count while it is still
829 		 * running, which makes it possible for the execve()
830 		 * to destroy vmspace while our other threads are
831 		 * still using the address space.
832 		 *
833 		 * We lock the thread, which is only allowed to
834 		 * succeed after context switch code finished using
835 		 * the address space.
836 		 */
837 		FOREACH_THREAD_IN_PROC(p, td2) {
838 			if (td2 == td)
839 				continue;
840 			thread_lock(td2);
841 			KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
842 			    ("td %p not on boundary", td2));
843 			KASSERT(TD_IS_SUSPENDED(td2),
844 			    ("td %p is not suspended", td2));
845 			thread_unlock(td2);
846 		}
847 	}
848 	PROC_SUNLOCK(p);
849 	return (0);
850 }
851 
852 bool
853 thread_suspend_check_needed(void)
854 {
855 	struct proc *p;
856 	struct thread *td;
857 
858 	td = curthread;
859 	p = td->td_proc;
860 	PROC_LOCK_ASSERT(p, MA_OWNED);
861 	return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
862 	    (td->td_dbgflags & TDB_SUSPEND) != 0));
863 }
864 
865 /*
866  * Called in from locations that can safely check to see
867  * whether we have to suspend or at least throttle for a
868  * single-thread event (e.g. fork).
869  *
870  * Such locations include userret().
871  * If the "return_instead" argument is non zero, the thread must be able to
872  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
873  *
874  * The 'return_instead' argument tells the function if it may do a
875  * thread_exit() or suspend, or whether the caller must abort and back
876  * out instead.
877  *
878  * If the thread that set the single_threading request has set the
879  * P_SINGLE_EXIT bit in the process flags then this call will never return
880  * if 'return_instead' is false, but will exit.
881  *
882  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
883  *---------------+--------------------+---------------------
884  *       0       | returns 0          |   returns 0 or 1
885  *               | when ST ends       |   immediately
886  *---------------+--------------------+---------------------
887  *       1       | thread exits       |   returns 1
888  *               |                    |  immediately
889  * 0 = thread_exit() or suspension ok,
890  * other = return error instead of stopping the thread.
891  *
892  * While a full suspension is under effect, even a single threading
893  * thread would be suspended if it made this call (but it shouldn't).
894  * This call should only be made from places where
895  * thread_exit() would be safe as that may be the outcome unless
896  * return_instead is set.
897  */
898 int
899 thread_suspend_check(int return_instead)
900 {
901 	struct thread *td;
902 	struct proc *p;
903 	int wakeup_swapper;
904 
905 	td = curthread;
906 	p = td->td_proc;
907 	mtx_assert(&Giant, MA_NOTOWNED);
908 	PROC_LOCK_ASSERT(p, MA_OWNED);
909 	while (thread_suspend_check_needed()) {
910 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
911 			KASSERT(p->p_singlethread != NULL,
912 			    ("singlethread not set"));
913 			/*
914 			 * The only suspension in action is a
915 			 * single-threading. Single threader need not stop.
916 			 * XXX Should be safe to access unlocked
917 			 * as it can only be set to be true by us.
918 			 */
919 			if (p->p_singlethread == td)
920 				return (0);	/* Exempt from stopping. */
921 		}
922 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
923 			return (EINTR);
924 
925 		/* Should we goto user boundary if we didn't come from there? */
926 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
927 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
928 			return (ERESTART);
929 
930 		/*
931 		 * Ignore suspend requests if they are deferred.
932 		 */
933 		if ((td->td_flags & TDF_SBDRY) != 0) {
934 			KASSERT(return_instead,
935 			    ("TDF_SBDRY set for unsafe thread_suspend_check"));
936 			return (0);
937 		}
938 
939 		/*
940 		 * If the process is waiting for us to exit,
941 		 * this thread should just suicide.
942 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
943 		 */
944 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
945 			PROC_UNLOCK(p);
946 
947 			/*
948 			 * Allow Linux emulation layer to do some work
949 			 * before thread suicide.
950 			 */
951 			if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
952 				(p->p_sysent->sv_thread_detach)(td);
953 			umtx_thread_exit(td);
954 			kern_thr_exit(td);
955 			panic("stopped thread did not exit");
956 		}
957 
958 		PROC_SLOCK(p);
959 		thread_stopped(p);
960 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
961 			if (p->p_numthreads == p->p_suspcount + 1) {
962 				thread_lock(p->p_singlethread);
963 				wakeup_swapper = thread_unsuspend_one(
964 				    p->p_singlethread, p, false);
965 				thread_unlock(p->p_singlethread);
966 				if (wakeup_swapper)
967 					kick_proc0();
968 			}
969 		}
970 		PROC_UNLOCK(p);
971 		thread_lock(td);
972 		/*
973 		 * When a thread suspends, it just
974 		 * gets taken off all queues.
975 		 */
976 		thread_suspend_one(td);
977 		if (return_instead == 0) {
978 			p->p_boundary_count++;
979 			td->td_flags |= TDF_BOUNDARY;
980 		}
981 		PROC_SUNLOCK(p);
982 		mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
983 		thread_unlock(td);
984 		PROC_LOCK(p);
985 	}
986 	return (0);
987 }
988 
989 void
990 thread_suspend_switch(struct thread *td, struct proc *p)
991 {
992 
993 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
994 	PROC_LOCK_ASSERT(p, MA_OWNED);
995 	PROC_SLOCK_ASSERT(p, MA_OWNED);
996 	/*
997 	 * We implement thread_suspend_one in stages here to avoid
998 	 * dropping the proc lock while the thread lock is owned.
999 	 */
1000 	if (p == td->td_proc) {
1001 		thread_stopped(p);
1002 		p->p_suspcount++;
1003 	}
1004 	PROC_UNLOCK(p);
1005 	thread_lock(td);
1006 	td->td_flags &= ~TDF_NEEDSUSPCHK;
1007 	TD_SET_SUSPENDED(td);
1008 	sched_sleep(td, 0);
1009 	PROC_SUNLOCK(p);
1010 	DROP_GIANT();
1011 	mi_switch(SW_VOL | SWT_SUSPEND, NULL);
1012 	thread_unlock(td);
1013 	PICKUP_GIANT();
1014 	PROC_LOCK(p);
1015 	PROC_SLOCK(p);
1016 }
1017 
1018 void
1019 thread_suspend_one(struct thread *td)
1020 {
1021 	struct proc *p;
1022 
1023 	p = td->td_proc;
1024 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1025 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1026 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1027 	p->p_suspcount++;
1028 	td->td_flags &= ~TDF_NEEDSUSPCHK;
1029 	TD_SET_SUSPENDED(td);
1030 	sched_sleep(td, 0);
1031 }
1032 
1033 static int
1034 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1035 {
1036 
1037 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1038 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1039 	TD_CLR_SUSPENDED(td);
1040 	td->td_flags &= ~TDF_ALLPROCSUSP;
1041 	if (td->td_proc == p) {
1042 		PROC_SLOCK_ASSERT(p, MA_OWNED);
1043 		p->p_suspcount--;
1044 		if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1045 			td->td_flags &= ~TDF_BOUNDARY;
1046 			p->p_boundary_count--;
1047 		}
1048 	}
1049 	return (setrunnable(td));
1050 }
1051 
1052 /*
1053  * Allow all threads blocked by single threading to continue running.
1054  */
1055 void
1056 thread_unsuspend(struct proc *p)
1057 {
1058 	struct thread *td;
1059 	int wakeup_swapper;
1060 
1061 	PROC_LOCK_ASSERT(p, MA_OWNED);
1062 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1063 	wakeup_swapper = 0;
1064 	if (!P_SHOULDSTOP(p)) {
1065                 FOREACH_THREAD_IN_PROC(p, td) {
1066 			thread_lock(td);
1067 			if (TD_IS_SUSPENDED(td)) {
1068 				wakeup_swapper |= thread_unsuspend_one(td, p,
1069 				    true);
1070 			}
1071 			thread_unlock(td);
1072 		}
1073 	} else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1074 	    p->p_numthreads == p->p_suspcount) {
1075 		/*
1076 		 * Stopping everything also did the job for the single
1077 		 * threading request. Now we've downgraded to single-threaded,
1078 		 * let it continue.
1079 		 */
1080 		if (p->p_singlethread->td_proc == p) {
1081 			thread_lock(p->p_singlethread);
1082 			wakeup_swapper = thread_unsuspend_one(
1083 			    p->p_singlethread, p, false);
1084 			thread_unlock(p->p_singlethread);
1085 		}
1086 	}
1087 	if (wakeup_swapper)
1088 		kick_proc0();
1089 }
1090 
1091 /*
1092  * End the single threading mode..
1093  */
1094 void
1095 thread_single_end(struct proc *p, int mode)
1096 {
1097 	struct thread *td;
1098 	int wakeup_swapper;
1099 
1100 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1101 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1102 	    ("invalid mode %d", mode));
1103 	PROC_LOCK_ASSERT(p, MA_OWNED);
1104 	KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1105 	    (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1106 	    ("mode %d does not match P_TOTAL_STOP", mode));
1107 	KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1108 	    ("thread_single_end from other thread %p %p",
1109 	    curthread, p->p_singlethread));
1110 	KASSERT(mode != SINGLE_BOUNDARY ||
1111 	    (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1112 	    ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1113 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1114 	    P_TOTAL_STOP);
1115 	PROC_SLOCK(p);
1116 	p->p_singlethread = NULL;
1117 	wakeup_swapper = 0;
1118 	/*
1119 	 * If there are other threads they may now run,
1120 	 * unless of course there is a blanket 'stop order'
1121 	 * on the process. The single threader must be allowed
1122 	 * to continue however as this is a bad place to stop.
1123 	 */
1124 	if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1125                 FOREACH_THREAD_IN_PROC(p, td) {
1126 			thread_lock(td);
1127 			if (TD_IS_SUSPENDED(td)) {
1128 				wakeup_swapper |= thread_unsuspend_one(td, p,
1129 				    mode == SINGLE_BOUNDARY);
1130 			}
1131 			thread_unlock(td);
1132 		}
1133 	}
1134 	KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1135 	    ("inconsistent boundary count %d", p->p_boundary_count));
1136 	PROC_SUNLOCK(p);
1137 	if (wakeup_swapper)
1138 		kick_proc0();
1139 }
1140 
1141 struct thread *
1142 thread_find(struct proc *p, lwpid_t tid)
1143 {
1144 	struct thread *td;
1145 
1146 	PROC_LOCK_ASSERT(p, MA_OWNED);
1147 	FOREACH_THREAD_IN_PROC(p, td) {
1148 		if (td->td_tid == tid)
1149 			break;
1150 	}
1151 	return (td);
1152 }
1153 
1154 /* Locate a thread by number; return with proc lock held. */
1155 struct thread *
1156 tdfind(lwpid_t tid, pid_t pid)
1157 {
1158 #define RUN_THRESH	16
1159 	struct thread *td;
1160 	int run = 0;
1161 
1162 	rw_rlock(&tidhash_lock);
1163 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1164 		if (td->td_tid == tid) {
1165 			if (pid != -1 && td->td_proc->p_pid != pid) {
1166 				td = NULL;
1167 				break;
1168 			}
1169 			PROC_LOCK(td->td_proc);
1170 			if (td->td_proc->p_state == PRS_NEW) {
1171 				PROC_UNLOCK(td->td_proc);
1172 				td = NULL;
1173 				break;
1174 			}
1175 			if (run > RUN_THRESH) {
1176 				if (rw_try_upgrade(&tidhash_lock)) {
1177 					LIST_REMOVE(td, td_hash);
1178 					LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1179 						td, td_hash);
1180 					rw_wunlock(&tidhash_lock);
1181 					return (td);
1182 				}
1183 			}
1184 			break;
1185 		}
1186 		run++;
1187 	}
1188 	rw_runlock(&tidhash_lock);
1189 	return (td);
1190 }
1191 
1192 void
1193 tidhash_add(struct thread *td)
1194 {
1195 	rw_wlock(&tidhash_lock);
1196 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1197 	rw_wunlock(&tidhash_lock);
1198 }
1199 
1200 void
1201 tidhash_remove(struct thread *td)
1202 {
1203 	rw_wlock(&tidhash_lock);
1204 	LIST_REMOVE(td, td_hash);
1205 	rw_wunlock(&tidhash_lock);
1206 }
1207