xref: /freebsd/sys/kern/kern_thread.c (revision ef0cb5db0af0d5d5b75b74f8e534fe601b7176d7)
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/sysent.h>
49 #include <sys/turnstile.h>
50 #include <sys/ktr.h>
51 #include <sys/rwlock.h>
52 #include <sys/umtx.h>
53 #include <sys/cpuset.h>
54 #ifdef	HWPMC_HOOKS
55 #include <sys/pmckern.h>
56 #endif
57 
58 #include <security/audit/audit.h>
59 
60 #include <vm/vm.h>
61 #include <vm/vm_extern.h>
62 #include <vm/uma.h>
63 #include <sys/eventhandler.h>
64 
65 SDT_PROVIDER_DECLARE(proc);
66 SDT_PROBE_DEFINE(proc, , , lwp__exit);
67 
68 /*
69  * thread related storage.
70  */
71 static uma_zone_t thread_zone;
72 
73 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
74 static struct mtx zombie_lock;
75 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
76 
77 static void thread_zombie(struct thread *);
78 static int thread_unsuspend_one(struct thread *td, struct proc *p,
79     bool boundary);
80 
81 #define TID_BUFFER_SIZE	1024
82 
83 struct mtx tid_lock;
84 static struct unrhdr *tid_unrhdr;
85 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
86 static int tid_head, tid_tail;
87 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
88 
89 struct	tidhashhead *tidhashtbl;
90 u_long	tidhash;
91 struct	rwlock tidhash_lock;
92 
93 static lwpid_t
94 tid_alloc(void)
95 {
96 	lwpid_t	tid;
97 
98 	tid = alloc_unr(tid_unrhdr);
99 	if (tid != -1)
100 		return (tid);
101 	mtx_lock(&tid_lock);
102 	if (tid_head == tid_tail) {
103 		mtx_unlock(&tid_lock);
104 		return (-1);
105 	}
106 	tid = tid_buffer[tid_head];
107 	tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
108 	mtx_unlock(&tid_lock);
109 	return (tid);
110 }
111 
112 static void
113 tid_free(lwpid_t tid)
114 {
115 	lwpid_t tmp_tid = -1;
116 
117 	mtx_lock(&tid_lock);
118 	if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
119 		tmp_tid = tid_buffer[tid_head];
120 		tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
121 	}
122 	tid_buffer[tid_tail] = tid;
123 	tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE;
124 	mtx_unlock(&tid_lock);
125 	if (tmp_tid != -1)
126 		free_unr(tid_unrhdr, tmp_tid);
127 }
128 
129 /*
130  * Prepare a thread for use.
131  */
132 static int
133 thread_ctor(void *mem, int size, void *arg, int flags)
134 {
135 	struct thread	*td;
136 
137 	td = (struct thread *)mem;
138 	td->td_state = TDS_INACTIVE;
139 	td->td_oncpu = NOCPU;
140 
141 	td->td_tid = tid_alloc();
142 
143 	/*
144 	 * Note that td_critnest begins life as 1 because the thread is not
145 	 * running and is thereby implicitly waiting to be on the receiving
146 	 * end of a context switch.
147 	 */
148 	td->td_critnest = 1;
149 	td->td_lend_user_pri = PRI_MAX;
150 	EVENTHANDLER_INVOKE(thread_ctor, td);
151 #ifdef AUDIT
152 	audit_thread_alloc(td);
153 #endif
154 	umtx_thread_alloc(td);
155 	return (0);
156 }
157 
158 /*
159  * Reclaim a thread after use.
160  */
161 static void
162 thread_dtor(void *mem, int size, void *arg)
163 {
164 	struct thread *td;
165 
166 	td = (struct thread *)mem;
167 
168 #ifdef INVARIANTS
169 	/* Verify that this thread is in a safe state to free. */
170 	switch (td->td_state) {
171 	case TDS_INHIBITED:
172 	case TDS_RUNNING:
173 	case TDS_CAN_RUN:
174 	case TDS_RUNQ:
175 		/*
176 		 * We must never unlink a thread that is in one of
177 		 * these states, because it is currently active.
178 		 */
179 		panic("bad state for thread unlinking");
180 		/* NOTREACHED */
181 	case TDS_INACTIVE:
182 		break;
183 	default:
184 		panic("bad thread state");
185 		/* NOTREACHED */
186 	}
187 #endif
188 #ifdef AUDIT
189 	audit_thread_free(td);
190 #endif
191 	/* Free all OSD associated to this thread. */
192 	osd_thread_exit(td);
193 
194 	EVENTHANDLER_INVOKE(thread_dtor, td);
195 	tid_free(td->td_tid);
196 }
197 
198 /*
199  * Initialize type-stable parts of a thread (when newly created).
200  */
201 static int
202 thread_init(void *mem, int size, int flags)
203 {
204 	struct thread *td;
205 
206 	td = (struct thread *)mem;
207 
208 	td->td_sleepqueue = sleepq_alloc();
209 	td->td_turnstile = turnstile_alloc();
210 	td->td_rlqe = NULL;
211 	EVENTHANDLER_INVOKE(thread_init, td);
212 	td->td_sched = (struct td_sched *)&td[1];
213 	umtx_thread_init(td);
214 	td->td_kstack = 0;
215 	td->td_sel = NULL;
216 	return (0);
217 }
218 
219 /*
220  * Tear down type-stable parts of a thread (just before being discarded).
221  */
222 static void
223 thread_fini(void *mem, int size)
224 {
225 	struct thread *td;
226 
227 	td = (struct thread *)mem;
228 	EVENTHANDLER_INVOKE(thread_fini, td);
229 	rlqentry_free(td->td_rlqe);
230 	turnstile_free(td->td_turnstile);
231 	sleepq_free(td->td_sleepqueue);
232 	umtx_thread_fini(td);
233 	seltdfini(td);
234 }
235 
236 /*
237  * For a newly created process,
238  * link up all the structures and its initial threads etc.
239  * called from:
240  * {arch}/{arch}/machdep.c   {arch}_init(), init386() etc.
241  * proc_dtor() (should go away)
242  * proc_init()
243  */
244 void
245 proc_linkup0(struct proc *p, struct thread *td)
246 {
247 	TAILQ_INIT(&p->p_threads);	     /* all threads in proc */
248 	proc_linkup(p, td);
249 }
250 
251 void
252 proc_linkup(struct proc *p, struct thread *td)
253 {
254 
255 	sigqueue_init(&p->p_sigqueue, p);
256 	p->p_ksi = ksiginfo_alloc(1);
257 	if (p->p_ksi != NULL) {
258 		/* XXX p_ksi may be null if ksiginfo zone is not ready */
259 		p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
260 	}
261 	LIST_INIT(&p->p_mqnotifier);
262 	p->p_numthreads = 0;
263 	thread_link(td, p);
264 }
265 
266 /*
267  * Initialize global thread allocation resources.
268  */
269 void
270 threadinit(void)
271 {
272 
273 	mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
274 
275 	/*
276 	 * pid_max cannot be greater than PID_MAX.
277 	 * leave one number for thread0.
278 	 */
279 	tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
280 
281 	thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
282 	    thread_ctor, thread_dtor, thread_init, thread_fini,
283 	    16 - 1, 0);
284 	tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
285 	rw_init(&tidhash_lock, "tidhash");
286 }
287 
288 /*
289  * Place an unused thread on the zombie list.
290  * Use the slpq as that must be unused by now.
291  */
292 void
293 thread_zombie(struct thread *td)
294 {
295 	mtx_lock_spin(&zombie_lock);
296 	TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
297 	mtx_unlock_spin(&zombie_lock);
298 }
299 
300 /*
301  * Release a thread that has exited after cpu_throw().
302  */
303 void
304 thread_stash(struct thread *td)
305 {
306 	atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
307 	thread_zombie(td);
308 }
309 
310 /*
311  * Reap zombie resources.
312  */
313 void
314 thread_reap(void)
315 {
316 	struct thread *td_first, *td_next;
317 
318 	/*
319 	 * Don't even bother to lock if none at this instant,
320 	 * we really don't care about the next instant..
321 	 */
322 	if (!TAILQ_EMPTY(&zombie_threads)) {
323 		mtx_lock_spin(&zombie_lock);
324 		td_first = TAILQ_FIRST(&zombie_threads);
325 		if (td_first)
326 			TAILQ_INIT(&zombie_threads);
327 		mtx_unlock_spin(&zombie_lock);
328 		while (td_first) {
329 			td_next = TAILQ_NEXT(td_first, td_slpq);
330 			if (td_first->td_ucred)
331 				crfree(td_first->td_ucred);
332 			thread_free(td_first);
333 			td_first = td_next;
334 		}
335 	}
336 }
337 
338 /*
339  * Allocate a thread.
340  */
341 struct thread *
342 thread_alloc(int pages)
343 {
344 	struct thread *td;
345 
346 	thread_reap(); /* check if any zombies to get */
347 
348 	td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
349 	KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
350 	if (!vm_thread_new(td, pages)) {
351 		uma_zfree(thread_zone, td);
352 		return (NULL);
353 	}
354 	cpu_thread_alloc(td);
355 	return (td);
356 }
357 
358 int
359 thread_alloc_stack(struct thread *td, int pages)
360 {
361 
362 	KASSERT(td->td_kstack == 0,
363 	    ("thread_alloc_stack called on a thread with kstack"));
364 	if (!vm_thread_new(td, pages))
365 		return (0);
366 	cpu_thread_alloc(td);
367 	return (1);
368 }
369 
370 /*
371  * Deallocate a thread.
372  */
373 void
374 thread_free(struct thread *td)
375 {
376 
377 	lock_profile_thread_exit(td);
378 	if (td->td_cpuset)
379 		cpuset_rel(td->td_cpuset);
380 	td->td_cpuset = NULL;
381 	cpu_thread_free(td);
382 	if (td->td_kstack != 0)
383 		vm_thread_dispose(td);
384 	uma_zfree(thread_zone, td);
385 }
386 
387 /*
388  * Discard the current thread and exit from its context.
389  * Always called with scheduler locked.
390  *
391  * Because we can't free a thread while we're operating under its context,
392  * push the current thread into our CPU's deadthread holder. This means
393  * we needn't worry about someone else grabbing our context before we
394  * do a cpu_throw().
395  */
396 void
397 thread_exit(void)
398 {
399 	uint64_t runtime, new_switchtime;
400 	struct thread *td;
401 	struct thread *td2;
402 	struct proc *p;
403 	int wakeup_swapper;
404 
405 	td = curthread;
406 	p = td->td_proc;
407 
408 	PROC_SLOCK_ASSERT(p, MA_OWNED);
409 	mtx_assert(&Giant, MA_NOTOWNED);
410 
411 	PROC_LOCK_ASSERT(p, MA_OWNED);
412 	KASSERT(p != NULL, ("thread exiting without a process"));
413 	CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
414 	    (long)p->p_pid, td->td_name);
415 	KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
416 
417 #ifdef AUDIT
418 	AUDIT_SYSCALL_EXIT(0, td);
419 #endif
420 	/*
421 	 * drop FPU & debug register state storage, or any other
422 	 * architecture specific resources that
423 	 * would not be on a new untouched process.
424 	 */
425 	cpu_thread_exit(td);	/* XXXSMP */
426 
427 	/*
428 	 * The last thread is left attached to the process
429 	 * So that the whole bundle gets recycled. Skip
430 	 * all this stuff if we never had threads.
431 	 * EXIT clears all sign of other threads when
432 	 * it goes to single threading, so the last thread always
433 	 * takes the short path.
434 	 */
435 	if (p->p_flag & P_HADTHREADS) {
436 		if (p->p_numthreads > 1) {
437 			atomic_add_int(&td->td_proc->p_exitthreads, 1);
438 			thread_unlink(td);
439 			td2 = FIRST_THREAD_IN_PROC(p);
440 			sched_exit_thread(td2, td);
441 
442 			/*
443 			 * The test below is NOT true if we are the
444 			 * sole exiting thread. P_STOPPED_SINGLE is unset
445 			 * in exit1() after it is the only survivor.
446 			 */
447 			if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
448 				if (p->p_numthreads == p->p_suspcount) {
449 					thread_lock(p->p_singlethread);
450 					wakeup_swapper = thread_unsuspend_one(
451 						p->p_singlethread, p, false);
452 					thread_unlock(p->p_singlethread);
453 					if (wakeup_swapper)
454 						kick_proc0();
455 				}
456 			}
457 
458 			PCPU_SET(deadthread, td);
459 		} else {
460 			/*
461 			 * The last thread is exiting.. but not through exit()
462 			 */
463 			panic ("thread_exit: Last thread exiting on its own");
464 		}
465 	}
466 #ifdef	HWPMC_HOOKS
467 	/*
468 	 * If this thread is part of a process that is being tracked by hwpmc(4),
469 	 * inform the module of the thread's impending exit.
470 	 */
471 	if (PMC_PROC_IS_USING_PMCS(td->td_proc))
472 		PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
473 #endif
474 	PROC_UNLOCK(p);
475 	PROC_STATLOCK(p);
476 	thread_lock(td);
477 	PROC_SUNLOCK(p);
478 
479 	/* Do the same timestamp bookkeeping that mi_switch() would do. */
480 	new_switchtime = cpu_ticks();
481 	runtime = new_switchtime - PCPU_GET(switchtime);
482 	td->td_runtime += runtime;
483 	td->td_incruntime += runtime;
484 	PCPU_SET(switchtime, new_switchtime);
485 	PCPU_SET(switchticks, ticks);
486 	PCPU_INC(cnt.v_swtch);
487 
488 	/* Save our resource usage in our process. */
489 	td->td_ru.ru_nvcsw++;
490 	ruxagg(p, td);
491 	rucollect(&p->p_ru, &td->td_ru);
492 	PROC_STATUNLOCK(p);
493 
494 	td->td_state = TDS_INACTIVE;
495 #ifdef WITNESS
496 	witness_thread_exit(td);
497 #endif
498 	CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
499 	sched_throw(td);
500 	panic("I'm a teapot!");
501 	/* NOTREACHED */
502 }
503 
504 /*
505  * Do any thread specific cleanups that may be needed in wait()
506  * called with Giant, proc and schedlock not held.
507  */
508 void
509 thread_wait(struct proc *p)
510 {
511 	struct thread *td;
512 
513 	mtx_assert(&Giant, MA_NOTOWNED);
514 	KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
515 	KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
516 	td = FIRST_THREAD_IN_PROC(p);
517 	/* Lock the last thread so we spin until it exits cpu_throw(). */
518 	thread_lock(td);
519 	thread_unlock(td);
520 	lock_profile_thread_exit(td);
521 	cpuset_rel(td->td_cpuset);
522 	td->td_cpuset = NULL;
523 	cpu_thread_clean(td);
524 	crfree(td->td_ucred);
525 	thread_reap();	/* check for zombie threads etc. */
526 }
527 
528 /*
529  * Link a thread to a process.
530  * set up anything that needs to be initialized for it to
531  * be used by the process.
532  */
533 void
534 thread_link(struct thread *td, struct proc *p)
535 {
536 
537 	/*
538 	 * XXX This can't be enabled because it's called for proc0 before
539 	 * its lock has been created.
540 	 * PROC_LOCK_ASSERT(p, MA_OWNED);
541 	 */
542 	td->td_state    = TDS_INACTIVE;
543 	td->td_proc     = p;
544 	td->td_flags    = TDF_INMEM;
545 
546 	LIST_INIT(&td->td_contested);
547 	LIST_INIT(&td->td_lprof[0]);
548 	LIST_INIT(&td->td_lprof[1]);
549 	sigqueue_init(&td->td_sigqueue, p);
550 	callout_init(&td->td_slpcallout, 1);
551 	TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
552 	p->p_numthreads++;
553 }
554 
555 /*
556  * Called from:
557  *  thread_exit()
558  */
559 void
560 thread_unlink(struct thread *td)
561 {
562 	struct proc *p = td->td_proc;
563 
564 	PROC_LOCK_ASSERT(p, MA_OWNED);
565 	TAILQ_REMOVE(&p->p_threads, td, td_plist);
566 	p->p_numthreads--;
567 	/* could clear a few other things here */
568 	/* Must  NOT clear links to proc! */
569 }
570 
571 static int
572 calc_remaining(struct proc *p, int mode)
573 {
574 	int remaining;
575 
576 	PROC_LOCK_ASSERT(p, MA_OWNED);
577 	PROC_SLOCK_ASSERT(p, MA_OWNED);
578 	if (mode == SINGLE_EXIT)
579 		remaining = p->p_numthreads;
580 	else if (mode == SINGLE_BOUNDARY)
581 		remaining = p->p_numthreads - p->p_boundary_count;
582 	else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
583 		remaining = p->p_numthreads - p->p_suspcount;
584 	else
585 		panic("calc_remaining: wrong mode %d", mode);
586 	return (remaining);
587 }
588 
589 static int
590 remain_for_mode(int mode)
591 {
592 
593 	return (mode == SINGLE_ALLPROC ? 0 : 1);
594 }
595 
596 static int
597 weed_inhib(int mode, struct thread *td2, struct proc *p)
598 {
599 	int wakeup_swapper;
600 
601 	PROC_LOCK_ASSERT(p, MA_OWNED);
602 	PROC_SLOCK_ASSERT(p, MA_OWNED);
603 	THREAD_LOCK_ASSERT(td2, MA_OWNED);
604 
605 	wakeup_swapper = 0;
606 	switch (mode) {
607 	case SINGLE_EXIT:
608 		if (TD_IS_SUSPENDED(td2))
609 			wakeup_swapper |= thread_unsuspend_one(td2, p, true);
610 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0)
611 			wakeup_swapper |= sleepq_abort(td2, EINTR);
612 		break;
613 	case SINGLE_BOUNDARY:
614 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & TDF_BOUNDARY) == 0)
615 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
616 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0)
617 			wakeup_swapper |= sleepq_abort(td2, ERESTART);
618 		break;
619 	case SINGLE_NO_EXIT:
620 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & TDF_BOUNDARY) == 0)
621 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
622 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0)
623 			wakeup_swapper |= sleepq_abort(td2, ERESTART);
624 		break;
625 	case SINGLE_ALLPROC:
626 		/*
627 		 * ALLPROC suspend tries to avoid spurious EINTR for
628 		 * threads sleeping interruptable, by suspending the
629 		 * thread directly, similarly to sig_suspend_threads().
630 		 * Since such sleep is not performed at the user
631 		 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
632 		 * is used to avoid immediate un-suspend.
633 		 */
634 		if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
635 		    TDF_ALLPROCSUSP)) == 0)
636 			wakeup_swapper |= thread_unsuspend_one(td2, p, false);
637 		if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR) != 0) {
638 			if ((td2->td_flags & TDF_SBDRY) == 0) {
639 				thread_suspend_one(td2);
640 				td2->td_flags |= TDF_ALLPROCSUSP;
641 			} else {
642 				wakeup_swapper |= sleepq_abort(td2, ERESTART);
643 			}
644 		}
645 		break;
646 	}
647 	return (wakeup_swapper);
648 }
649 
650 /*
651  * Enforce single-threading.
652  *
653  * Returns 1 if the caller must abort (another thread is waiting to
654  * exit the process or similar). Process is locked!
655  * Returns 0 when you are successfully the only thread running.
656  * A process has successfully single threaded in the suspend mode when
657  * There are no threads in user mode. Threads in the kernel must be
658  * allowed to continue until they get to the user boundary. They may even
659  * copy out their return values and data before suspending. They may however be
660  * accelerated in reaching the user boundary as we will wake up
661  * any sleeping threads that are interruptable. (PCATCH).
662  */
663 int
664 thread_single(struct proc *p, int mode)
665 {
666 	struct thread *td;
667 	struct thread *td2;
668 	int remaining, wakeup_swapper;
669 
670 	td = curthread;
671 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
672 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
673 	    ("invalid mode %d", mode));
674 	/*
675 	 * If allowing non-ALLPROC singlethreading for non-curproc
676 	 * callers, calc_remaining() and remain_for_mode() should be
677 	 * adjusted to also account for td->td_proc != p.  For now
678 	 * this is not implemented because it is not used.
679 	 */
680 	KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
681 	    (mode != SINGLE_ALLPROC && td->td_proc == p),
682 	    ("mode %d proc %p curproc %p", mode, p, td->td_proc));
683 	mtx_assert(&Giant, MA_NOTOWNED);
684 	PROC_LOCK_ASSERT(p, MA_OWNED);
685 
686 	if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
687 		return (0);
688 
689 	/* Is someone already single threading? */
690 	if (p->p_singlethread != NULL && p->p_singlethread != td)
691 		return (1);
692 
693 	if (mode == SINGLE_EXIT) {
694 		p->p_flag |= P_SINGLE_EXIT;
695 		p->p_flag &= ~P_SINGLE_BOUNDARY;
696 	} else {
697 		p->p_flag &= ~P_SINGLE_EXIT;
698 		if (mode == SINGLE_BOUNDARY)
699 			p->p_flag |= P_SINGLE_BOUNDARY;
700 		else
701 			p->p_flag &= ~P_SINGLE_BOUNDARY;
702 	}
703 	if (mode == SINGLE_ALLPROC)
704 		p->p_flag |= P_TOTAL_STOP;
705 	p->p_flag |= P_STOPPED_SINGLE;
706 	PROC_SLOCK(p);
707 	p->p_singlethread = td;
708 	remaining = calc_remaining(p, mode);
709 	while (remaining != remain_for_mode(mode)) {
710 		if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
711 			goto stopme;
712 		wakeup_swapper = 0;
713 		FOREACH_THREAD_IN_PROC(p, td2) {
714 			if (td2 == td)
715 				continue;
716 			thread_lock(td2);
717 			td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
718 			if (TD_IS_INHIBITED(td2)) {
719 				wakeup_swapper |= weed_inhib(mode, td2, p);
720 #ifdef SMP
721 			} else if (TD_IS_RUNNING(td2) && td != td2) {
722 				forward_signal(td2);
723 #endif
724 			}
725 			thread_unlock(td2);
726 		}
727 		if (wakeup_swapper)
728 			kick_proc0();
729 		remaining = calc_remaining(p, mode);
730 
731 		/*
732 		 * Maybe we suspended some threads.. was it enough?
733 		 */
734 		if (remaining == remain_for_mode(mode))
735 			break;
736 
737 stopme:
738 		/*
739 		 * Wake us up when everyone else has suspended.
740 		 * In the mean time we suspend as well.
741 		 */
742 		thread_suspend_switch(td, p);
743 		remaining = calc_remaining(p, mode);
744 	}
745 	if (mode == SINGLE_EXIT) {
746 		/*
747 		 * Convert the process to an unthreaded process.  The
748 		 * SINGLE_EXIT is called by exit1() or execve(), in
749 		 * both cases other threads must be retired.
750 		 */
751 		KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
752 		p->p_singlethread = NULL;
753 		p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
754 
755 		/*
756 		 * Wait for any remaining threads to exit cpu_throw().
757 		 */
758 		while (p->p_exitthreads != 0) {
759 			PROC_SUNLOCK(p);
760 			PROC_UNLOCK(p);
761 			sched_relinquish(td);
762 			PROC_LOCK(p);
763 			PROC_SLOCK(p);
764 		}
765 	} else if (mode == SINGLE_BOUNDARY) {
766 		/*
767 		 * Wait until all suspended threads are removed from
768 		 * the processors.  The thread_suspend_check()
769 		 * increments p_boundary_count while it is still
770 		 * running, which makes it possible for the execve()
771 		 * to destroy vmspace while our other threads are
772 		 * still using the address space.
773 		 *
774 		 * We lock the thread, which is only allowed to
775 		 * succeed after context switch code finished using
776 		 * the address space.
777 		 */
778 		FOREACH_THREAD_IN_PROC(p, td2) {
779 			if (td2 == td)
780 				continue;
781 			thread_lock(td2);
782 			KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
783 			    ("td %p not on boundary", td2));
784 			KASSERT(TD_IS_SUSPENDED(td2),
785 			    ("td %p is not suspended", td2));
786 			thread_unlock(td2);
787 		}
788 	}
789 	PROC_SUNLOCK(p);
790 	return (0);
791 }
792 
793 bool
794 thread_suspend_check_needed(void)
795 {
796 	struct proc *p;
797 	struct thread *td;
798 
799 	td = curthread;
800 	p = td->td_proc;
801 	PROC_LOCK_ASSERT(p, MA_OWNED);
802 	return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
803 	    (td->td_dbgflags & TDB_SUSPEND) != 0));
804 }
805 
806 /*
807  * Called in from locations that can safely check to see
808  * whether we have to suspend or at least throttle for a
809  * single-thread event (e.g. fork).
810  *
811  * Such locations include userret().
812  * If the "return_instead" argument is non zero, the thread must be able to
813  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
814  *
815  * The 'return_instead' argument tells the function if it may do a
816  * thread_exit() or suspend, or whether the caller must abort and back
817  * out instead.
818  *
819  * If the thread that set the single_threading request has set the
820  * P_SINGLE_EXIT bit in the process flags then this call will never return
821  * if 'return_instead' is false, but will exit.
822  *
823  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
824  *---------------+--------------------+---------------------
825  *       0       | returns 0          |   returns 0 or 1
826  *               | when ST ends       |   immediately
827  *---------------+--------------------+---------------------
828  *       1       | thread exits       |   returns 1
829  *               |                    |  immediately
830  * 0 = thread_exit() or suspension ok,
831  * other = return error instead of stopping the thread.
832  *
833  * While a full suspension is under effect, even a single threading
834  * thread would be suspended if it made this call (but it shouldn't).
835  * This call should only be made from places where
836  * thread_exit() would be safe as that may be the outcome unless
837  * return_instead is set.
838  */
839 int
840 thread_suspend_check(int return_instead)
841 {
842 	struct thread *td;
843 	struct proc *p;
844 	int wakeup_swapper;
845 
846 	td = curthread;
847 	p = td->td_proc;
848 	mtx_assert(&Giant, MA_NOTOWNED);
849 	PROC_LOCK_ASSERT(p, MA_OWNED);
850 	while (thread_suspend_check_needed()) {
851 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
852 			KASSERT(p->p_singlethread != NULL,
853 			    ("singlethread not set"));
854 			/*
855 			 * The only suspension in action is a
856 			 * single-threading. Single threader need not stop.
857 			 * XXX Should be safe to access unlocked
858 			 * as it can only be set to be true by us.
859 			 */
860 			if (p->p_singlethread == td)
861 				return (0);	/* Exempt from stopping. */
862 		}
863 		if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
864 			return (EINTR);
865 
866 		/* Should we goto user boundary if we didn't come from there? */
867 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
868 		    (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
869 			return (ERESTART);
870 
871 		/*
872 		 * Ignore suspend requests if they are deferred.
873 		 */
874 		if ((td->td_flags & TDF_SBDRY) != 0) {
875 			KASSERT(return_instead,
876 			    ("TDF_SBDRY set for unsafe thread_suspend_check"));
877 			return (0);
878 		}
879 
880 		/*
881 		 * If the process is waiting for us to exit,
882 		 * this thread should just suicide.
883 		 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
884 		 */
885 		if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
886 			PROC_UNLOCK(p);
887 			tidhash_remove(td);
888 
889 			/*
890 			 * Allow Linux emulation layer to do some work
891 			 * before thread suicide.
892 			 */
893 			if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
894 				(p->p_sysent->sv_thread_detach)(td);
895 
896 			PROC_LOCK(p);
897 			tdsigcleanup(td);
898 			umtx_thread_exit(td);
899 			PROC_SLOCK(p);
900 			thread_stopped(p);
901 			thread_exit();
902 		}
903 
904 		PROC_SLOCK(p);
905 		thread_stopped(p);
906 		if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
907 			if (p->p_numthreads == p->p_suspcount + 1) {
908 				thread_lock(p->p_singlethread);
909 				wakeup_swapper = thread_unsuspend_one(
910 				    p->p_singlethread, p, false);
911 				thread_unlock(p->p_singlethread);
912 				if (wakeup_swapper)
913 					kick_proc0();
914 			}
915 		}
916 		PROC_UNLOCK(p);
917 		thread_lock(td);
918 		/*
919 		 * When a thread suspends, it just
920 		 * gets taken off all queues.
921 		 */
922 		thread_suspend_one(td);
923 		if (return_instead == 0) {
924 			p->p_boundary_count++;
925 			td->td_flags |= TDF_BOUNDARY;
926 		}
927 		PROC_SUNLOCK(p);
928 		mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
929 		thread_unlock(td);
930 		PROC_LOCK(p);
931 	}
932 	return (0);
933 }
934 
935 void
936 thread_suspend_switch(struct thread *td, struct proc *p)
937 {
938 
939 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
940 	PROC_LOCK_ASSERT(p, MA_OWNED);
941 	PROC_SLOCK_ASSERT(p, MA_OWNED);
942 	/*
943 	 * We implement thread_suspend_one in stages here to avoid
944 	 * dropping the proc lock while the thread lock is owned.
945 	 */
946 	if (p == td->td_proc) {
947 		thread_stopped(p);
948 		p->p_suspcount++;
949 	}
950 	PROC_UNLOCK(p);
951 	thread_lock(td);
952 	td->td_flags &= ~TDF_NEEDSUSPCHK;
953 	TD_SET_SUSPENDED(td);
954 	sched_sleep(td, 0);
955 	PROC_SUNLOCK(p);
956 	DROP_GIANT();
957 	mi_switch(SW_VOL | SWT_SUSPEND, NULL);
958 	thread_unlock(td);
959 	PICKUP_GIANT();
960 	PROC_LOCK(p);
961 	PROC_SLOCK(p);
962 }
963 
964 void
965 thread_suspend_one(struct thread *td)
966 {
967 	struct proc *p;
968 
969 	p = td->td_proc;
970 	PROC_SLOCK_ASSERT(p, MA_OWNED);
971 	THREAD_LOCK_ASSERT(td, MA_OWNED);
972 	KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
973 	p->p_suspcount++;
974 	td->td_flags &= ~TDF_NEEDSUSPCHK;
975 	TD_SET_SUSPENDED(td);
976 	sched_sleep(td, 0);
977 }
978 
979 static int
980 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
981 {
982 
983 	THREAD_LOCK_ASSERT(td, MA_OWNED);
984 	KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
985 	TD_CLR_SUSPENDED(td);
986 	td->td_flags &= ~TDF_ALLPROCSUSP;
987 	if (td->td_proc == p) {
988 		PROC_SLOCK_ASSERT(p, MA_OWNED);
989 		p->p_suspcount--;
990 		if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
991 			td->td_flags &= ~TDF_BOUNDARY;
992 			p->p_boundary_count--;
993 		}
994 	}
995 	return (setrunnable(td));
996 }
997 
998 /*
999  * Allow all threads blocked by single threading to continue running.
1000  */
1001 void
1002 thread_unsuspend(struct proc *p)
1003 {
1004 	struct thread *td;
1005 	int wakeup_swapper;
1006 
1007 	PROC_LOCK_ASSERT(p, MA_OWNED);
1008 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1009 	wakeup_swapper = 0;
1010 	if (!P_SHOULDSTOP(p)) {
1011                 FOREACH_THREAD_IN_PROC(p, td) {
1012 			thread_lock(td);
1013 			if (TD_IS_SUSPENDED(td)) {
1014 				wakeup_swapper |= thread_unsuspend_one(td, p,
1015 				    true);
1016 			}
1017 			thread_unlock(td);
1018 		}
1019 	} else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1020 	    p->p_numthreads == p->p_suspcount) {
1021 		/*
1022 		 * Stopping everything also did the job for the single
1023 		 * threading request. Now we've downgraded to single-threaded,
1024 		 * let it continue.
1025 		 */
1026 		if (p->p_singlethread->td_proc == p) {
1027 			thread_lock(p->p_singlethread);
1028 			wakeup_swapper = thread_unsuspend_one(
1029 			    p->p_singlethread, p, false);
1030 			thread_unlock(p->p_singlethread);
1031 		}
1032 	}
1033 	if (wakeup_swapper)
1034 		kick_proc0();
1035 }
1036 
1037 /*
1038  * End the single threading mode..
1039  */
1040 void
1041 thread_single_end(struct proc *p, int mode)
1042 {
1043 	struct thread *td;
1044 	int wakeup_swapper;
1045 
1046 	KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1047 	    mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1048 	    ("invalid mode %d", mode));
1049 	PROC_LOCK_ASSERT(p, MA_OWNED);
1050 	KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1051 	    (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1052 	    ("mode %d does not match P_TOTAL_STOP", mode));
1053 	KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1054 	    ("thread_single_end from other thread %p %p",
1055 	    curthread, p->p_singlethread));
1056 	KASSERT(mode != SINGLE_BOUNDARY ||
1057 	    (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1058 	    ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1059 	p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1060 	    P_TOTAL_STOP);
1061 	PROC_SLOCK(p);
1062 	p->p_singlethread = NULL;
1063 	wakeup_swapper = 0;
1064 	/*
1065 	 * If there are other threads they may now run,
1066 	 * unless of course there is a blanket 'stop order'
1067 	 * on the process. The single threader must be allowed
1068 	 * to continue however as this is a bad place to stop.
1069 	 */
1070 	if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1071                 FOREACH_THREAD_IN_PROC(p, td) {
1072 			thread_lock(td);
1073 			if (TD_IS_SUSPENDED(td)) {
1074 				wakeup_swapper |= thread_unsuspend_one(td, p,
1075 				    mode == SINGLE_BOUNDARY);
1076 			}
1077 			thread_unlock(td);
1078 		}
1079 	}
1080 	KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1081 	    ("inconsistent boundary count %d", p->p_boundary_count));
1082 	PROC_SUNLOCK(p);
1083 	if (wakeup_swapper)
1084 		kick_proc0();
1085 }
1086 
1087 struct thread *
1088 thread_find(struct proc *p, lwpid_t tid)
1089 {
1090 	struct thread *td;
1091 
1092 	PROC_LOCK_ASSERT(p, MA_OWNED);
1093 	FOREACH_THREAD_IN_PROC(p, td) {
1094 		if (td->td_tid == tid)
1095 			break;
1096 	}
1097 	return (td);
1098 }
1099 
1100 /* Locate a thread by number; return with proc lock held. */
1101 struct thread *
1102 tdfind(lwpid_t tid, pid_t pid)
1103 {
1104 #define RUN_THRESH	16
1105 	struct thread *td;
1106 	int run = 0;
1107 
1108 	rw_rlock(&tidhash_lock);
1109 	LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1110 		if (td->td_tid == tid) {
1111 			if (pid != -1 && td->td_proc->p_pid != pid) {
1112 				td = NULL;
1113 				break;
1114 			}
1115 			PROC_LOCK(td->td_proc);
1116 			if (td->td_proc->p_state == PRS_NEW) {
1117 				PROC_UNLOCK(td->td_proc);
1118 				td = NULL;
1119 				break;
1120 			}
1121 			if (run > RUN_THRESH) {
1122 				if (rw_try_upgrade(&tidhash_lock)) {
1123 					LIST_REMOVE(td, td_hash);
1124 					LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1125 						td, td_hash);
1126 					rw_wunlock(&tidhash_lock);
1127 					return (td);
1128 				}
1129 			}
1130 			break;
1131 		}
1132 		run++;
1133 	}
1134 	rw_runlock(&tidhash_lock);
1135 	return (td);
1136 }
1137 
1138 void
1139 tidhash_add(struct thread *td)
1140 {
1141 	rw_wlock(&tidhash_lock);
1142 	LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1143 	rw_wunlock(&tidhash_lock);
1144 }
1145 
1146 void
1147 tidhash_remove(struct thread *td)
1148 {
1149 	rw_wlock(&tidhash_lock);
1150 	LIST_REMOVE(td, td_hash);
1151 	rw_wunlock(&tidhash_lock);
1152 }
1153