2 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
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.
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
29 #include "opt_witness.h"
30 #include "opt_kdtrace.h"
31 #include "opt_hwpmc_hooks.h"
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
40 #include <sys/mutex.h>
42 #include <sys/resourcevar.h>
45 #include <sys/sched.h>
46 #include <sys/sleepqueue.h>
47 #include <sys/selinfo.h>
48 #include <sys/turnstile.h>
51 #include <sys/cpuset.h>
53 #include <sys/pmckern.h>
56 #include <security/audit/audit.h>
59 #include <vm/vm_extern.h>
61 #include <sys/eventhandler.h>
63 SDT_PROVIDER_DECLARE(proc);
64 SDT_PROBE_DEFINE(proc, , , lwp_exit, lwp-exit);
68 * thread related storage.
70 static uma_zone_t thread_zone;
72 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
73 static struct mtx zombie_lock;
74 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
76 static void thread_zombie(struct thread *);
79 static struct unrhdr *tid_unrhdr;
82 * Prepare a thread for use.
85 thread_ctor(void *mem, int size, void *arg, int flags)
89 td = (struct thread *)mem;
90 td->td_state = TDS_INACTIVE;
93 td->td_tid = alloc_unr(tid_unrhdr);
97 * Note that td_critnest begins life as 1 because the thread is not
98 * running and is thereby implicitly waiting to be on the receiving
99 * end of a context switch.
102 EVENTHANDLER_INVOKE(thread_ctor, td);
104 audit_thread_alloc(td);
106 umtx_thread_alloc(td);
111 * Reclaim a thread after use.
114 thread_dtor(void *mem, int size, void *arg)
118 td = (struct thread *)mem;
121 /* Verify that this thread is in a safe state to free. */
122 switch (td->td_state) {
128 * We must never unlink a thread that is in one of
129 * these states, because it is currently active.
131 panic("bad state for thread unlinking");
136 panic("bad thread state");
141 audit_thread_free(td);
143 /* Free all OSD associated to this thread. */
146 EVENTHANDLER_INVOKE(thread_dtor, td);
147 free_unr(tid_unrhdr, td->td_tid);
151 * Initialize type-stable parts of a thread (when newly created).
154 thread_init(void *mem, int size, int flags)
158 td = (struct thread *)mem;
160 td->td_sleepqueue = sleepq_alloc();
161 td->td_turnstile = turnstile_alloc();
162 EVENTHANDLER_INVOKE(thread_init, td);
163 td->td_sched = (struct td_sched *)&td[1];
164 umtx_thread_init(td);
170 * Tear down type-stable parts of a thread (just before being discarded).
173 thread_fini(void *mem, int size)
177 td = (struct thread *)mem;
178 EVENTHANDLER_INVOKE(thread_fini, td);
179 turnstile_free(td->td_turnstile);
180 sleepq_free(td->td_sleepqueue);
181 umtx_thread_fini(td);
186 * For a newly created process,
187 * link up all the structures and its initial threads etc.
189 * {arch}/{arch}/machdep.c ia64_init(), init386() etc.
190 * proc_dtor() (should go away)
194 proc_linkup0(struct proc *p, struct thread *td)
196 TAILQ_INIT(&p->p_threads); /* all threads in proc */
201 proc_linkup(struct proc *p, struct thread *td)
204 sigqueue_init(&p->p_sigqueue, p);
205 p->p_ksi = ksiginfo_alloc(1);
206 if (p->p_ksi != NULL) {
207 /* XXX p_ksi may be null if ksiginfo zone is not ready */
208 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
210 LIST_INIT(&p->p_mqnotifier);
216 * Initialize global thread allocation resources.
222 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
223 /* leave one number for thread0 */
224 tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
226 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
227 thread_ctor, thread_dtor, thread_init, thread_fini,
232 * Place an unused thread on the zombie list.
233 * Use the slpq as that must be unused by now.
236 thread_zombie(struct thread *td)
238 mtx_lock_spin(&zombie_lock);
239 TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
240 mtx_unlock_spin(&zombie_lock);
244 * Release a thread that has exited after cpu_throw().
247 thread_stash(struct thread *td)
249 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
254 * Reap zombie resources.
259 struct thread *td_first, *td_next;
262 * Don't even bother to lock if none at this instant,
263 * we really don't care about the next instant..
265 if (!TAILQ_EMPTY(&zombie_threads)) {
266 mtx_lock_spin(&zombie_lock);
267 td_first = TAILQ_FIRST(&zombie_threads);
269 TAILQ_INIT(&zombie_threads);
270 mtx_unlock_spin(&zombie_lock);
272 td_next = TAILQ_NEXT(td_first, td_slpq);
273 if (td_first->td_ucred)
274 crfree(td_first->td_ucred);
275 thread_free(td_first);
285 thread_alloc(int pages)
289 thread_reap(); /* check if any zombies to get */
291 td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
292 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
293 if (!vm_thread_new(td, pages)) {
294 uma_zfree(thread_zone, td);
297 cpu_thread_alloc(td);
302 thread_alloc_stack(struct thread *td, int pages)
305 KASSERT(td->td_kstack == 0,
306 ("thread_alloc_stack called on a thread with kstack"));
307 if (!vm_thread_new(td, pages))
309 cpu_thread_alloc(td);
314 * Deallocate a thread.
317 thread_free(struct thread *td)
320 lock_profile_thread_exit(td);
322 cpuset_rel(td->td_cpuset);
323 td->td_cpuset = NULL;
325 if (td->td_kstack != 0)
326 vm_thread_dispose(td);
327 uma_zfree(thread_zone, td);
331 * Discard the current thread and exit from its context.
332 * Always called with scheduler locked.
334 * Because we can't free a thread while we're operating under its context,
335 * push the current thread into our CPU's deadthread holder. This means
336 * we needn't worry about someone else grabbing our context before we
342 uint64_t runtime, new_switchtime;
351 PROC_SLOCK_ASSERT(p, MA_OWNED);
352 mtx_assert(&Giant, MA_NOTOWNED);
354 PROC_LOCK_ASSERT(p, MA_OWNED);
355 KASSERT(p != NULL, ("thread exiting without a process"));
356 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
357 (long)p->p_pid, td->td_name);
358 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
361 AUDIT_SYSCALL_EXIT(0, td);
363 umtx_thread_exit(td);
365 * drop FPU & debug register state storage, or any other
366 * architecture specific resources that
367 * would not be on a new untouched process.
369 cpu_thread_exit(td); /* XXXSMP */
372 * The last thread is left attached to the process
373 * So that the whole bundle gets recycled. Skip
374 * all this stuff if we never had threads.
375 * EXIT clears all sign of other threads when
376 * it goes to single threading, so the last thread always
377 * takes the short path.
379 if (p->p_flag & P_HADTHREADS) {
380 if (p->p_numthreads > 1) {
382 td2 = FIRST_THREAD_IN_PROC(p);
383 sched_exit_thread(td2, td);
386 * The test below is NOT true if we are the
387 * sole exiting thread. P_STOPPED_SINGLE is unset
388 * in exit1() after it is the only survivor.
390 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
391 if (p->p_numthreads == p->p_suspcount) {
392 thread_lock(p->p_singlethread);
393 wakeup_swapper = thread_unsuspend_one(
395 thread_unlock(p->p_singlethread);
401 atomic_add_int(&td->td_proc->p_exitthreads, 1);
402 PCPU_SET(deadthread, td);
405 * The last thread is exiting.. but not through exit()
407 panic ("thread_exit: Last thread exiting on its own");
412 * If this thread is part of a process that is being tracked by hwpmc(4),
413 * inform the module of the thread's impending exit.
415 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
416 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
420 /* Do the same timestamp bookkeeping that mi_switch() would do. */
421 new_switchtime = cpu_ticks();
422 runtime = new_switchtime - PCPU_GET(switchtime);
423 td->td_runtime += runtime;
424 td->td_incruntime += runtime;
425 PCPU_SET(switchtime, new_switchtime);
426 PCPU_SET(switchticks, ticks);
427 PCPU_INC(cnt.v_swtch);
429 /* Save our resource usage in our process. */
430 td->td_ru.ru_nvcsw++;
432 rucollect(&p->p_ru, &td->td_ru);
436 td->td_state = TDS_INACTIVE;
438 witness_thread_exit(td);
440 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
442 panic("I'm a teapot!");
447 * Do any thread specific cleanups that may be needed in wait()
448 * called with Giant, proc and schedlock not held.
451 thread_wait(struct proc *p)
455 mtx_assert(&Giant, MA_NOTOWNED);
456 KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
457 td = FIRST_THREAD_IN_PROC(p);
458 /* Lock the last thread so we spin until it exits cpu_throw(). */
461 /* Wait for any remaining threads to exit cpu_throw(). */
462 while (p->p_exitthreads)
463 sched_relinquish(curthread);
464 lock_profile_thread_exit(td);
465 cpuset_rel(td->td_cpuset);
466 td->td_cpuset = NULL;
467 cpu_thread_clean(td);
468 crfree(td->td_ucred);
469 thread_reap(); /* check for zombie threads etc. */
473 * Link a thread to a process.
474 * set up anything that needs to be initialized for it to
475 * be used by the process.
478 thread_link(struct thread *td, struct proc *p)
482 * XXX This can't be enabled because it's called for proc0 before
483 * its lock has been created.
484 * PROC_LOCK_ASSERT(p, MA_OWNED);
486 td->td_state = TDS_INACTIVE;
488 td->td_flags = TDF_INMEM;
490 LIST_INIT(&td->td_contested);
491 LIST_INIT(&td->td_lprof[0]);
492 LIST_INIT(&td->td_lprof[1]);
493 sigqueue_init(&td->td_sigqueue, p);
494 callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
495 TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
500 * Convert a process with one thread to an unthreaded process.
503 thread_unthread(struct thread *td)
505 struct proc *p = td->td_proc;
507 KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
508 p->p_flag &= ~P_HADTHREADS;
516 thread_unlink(struct thread *td)
518 struct proc *p = td->td_proc;
520 PROC_LOCK_ASSERT(p, MA_OWNED);
521 TAILQ_REMOVE(&p->p_threads, td, td_plist);
523 /* could clear a few other things here */
524 /* Must NOT clear links to proc! */
528 calc_remaining(struct proc *p, int mode)
532 PROC_LOCK_ASSERT(p, MA_OWNED);
533 PROC_SLOCK_ASSERT(p, MA_OWNED);
534 if (mode == SINGLE_EXIT)
535 remaining = p->p_numthreads;
536 else if (mode == SINGLE_BOUNDARY)
537 remaining = p->p_numthreads - p->p_boundary_count;
538 else if (mode == SINGLE_NO_EXIT)
539 remaining = p->p_numthreads - p->p_suspcount;
541 panic("calc_remaining: wrong mode %d", mode);
546 * Enforce single-threading.
548 * Returns 1 if the caller must abort (another thread is waiting to
549 * exit the process or similar). Process is locked!
550 * Returns 0 when you are successfully the only thread running.
551 * A process has successfully single threaded in the suspend mode when
552 * There are no threads in user mode. Threads in the kernel must be
553 * allowed to continue until they get to the user boundary. They may even
554 * copy out their return values and data before suspending. They may however be
555 * accelerated in reaching the user boundary as we will wake up
556 * any sleeping threads that are interruptable. (PCATCH).
559 thread_single(int mode)
564 int remaining, wakeup_swapper;
568 mtx_assert(&Giant, MA_NOTOWNED);
569 PROC_LOCK_ASSERT(p, MA_OWNED);
570 KASSERT((td != NULL), ("curthread is NULL"));
572 if ((p->p_flag & P_HADTHREADS) == 0)
575 /* Is someone already single threading? */
576 if (p->p_singlethread != NULL && p->p_singlethread != td)
579 if (mode == SINGLE_EXIT) {
580 p->p_flag |= P_SINGLE_EXIT;
581 p->p_flag &= ~P_SINGLE_BOUNDARY;
583 p->p_flag &= ~P_SINGLE_EXIT;
584 if (mode == SINGLE_BOUNDARY)
585 p->p_flag |= P_SINGLE_BOUNDARY;
587 p->p_flag &= ~P_SINGLE_BOUNDARY;
589 p->p_flag |= P_STOPPED_SINGLE;
591 p->p_singlethread = td;
592 remaining = calc_remaining(p, mode);
593 while (remaining != 1) {
594 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
597 FOREACH_THREAD_IN_PROC(p, td2) {
601 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
602 if (TD_IS_INHIBITED(td2)) {
605 if (TD_IS_SUSPENDED(td2))
607 thread_unsuspend_one(td2);
608 if (TD_ON_SLEEPQ(td2) &&
609 (td2->td_flags & TDF_SINTR))
611 sleepq_abort(td2, EINTR);
613 case SINGLE_BOUNDARY:
614 if (TD_IS_SUSPENDED(td2) &&
615 !(td2->td_flags & TDF_BOUNDARY))
617 thread_unsuspend_one(td2);
618 if (TD_ON_SLEEPQ(td2) &&
619 (td2->td_flags & TDF_SINTR))
621 sleepq_abort(td2, ERESTART);
624 if (TD_IS_SUSPENDED(td2) &&
625 !(td2->td_flags & TDF_BOUNDARY))
627 thread_unsuspend_one(td2);
628 if (TD_ON_SLEEPQ(td2) &&
629 (td2->td_flags & TDF_SINTR))
631 sleepq_abort(td2, ERESTART);
638 else if (TD_IS_RUNNING(td2) && td != td2) {
646 remaining = calc_remaining(p, mode);
649 * Maybe we suspended some threads.. was it enough?
656 * Wake us up when everyone else has suspended.
657 * In the mean time we suspend as well.
659 thread_suspend_switch(td);
660 remaining = calc_remaining(p, mode);
662 if (mode == SINGLE_EXIT) {
664 * We have gotten rid of all the other threads and we
665 * are about to either exit or exec. In either case,
666 * we try our utmost to revert to being a non-threaded
669 p->p_singlethread = NULL;
670 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
678 * Called in from locations that can safely check to see
679 * whether we have to suspend or at least throttle for a
680 * single-thread event (e.g. fork).
682 * Such locations include userret().
683 * If the "return_instead" argument is non zero, the thread must be able to
684 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
686 * The 'return_instead' argument tells the function if it may do a
687 * thread_exit() or suspend, or whether the caller must abort and back
690 * If the thread that set the single_threading request has set the
691 * P_SINGLE_EXIT bit in the process flags then this call will never return
692 * if 'return_instead' is false, but will exit.
694 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
695 *---------------+--------------------+---------------------
696 * 0 | returns 0 | returns 0 or 1
697 * | when ST ends | immediatly
698 *---------------+--------------------+---------------------
699 * 1 | thread exits | returns 1
701 * 0 = thread_exit() or suspension ok,
702 * other = return error instead of stopping the thread.
704 * While a full suspension is under effect, even a single threading
705 * thread would be suspended if it made this call (but it shouldn't).
706 * This call should only be made from places where
707 * thread_exit() would be safe as that may be the outcome unless
708 * return_instead is set.
711 thread_suspend_check(int return_instead)
719 mtx_assert(&Giant, MA_NOTOWNED);
720 PROC_LOCK_ASSERT(p, MA_OWNED);
721 while (P_SHOULDSTOP(p) ||
722 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
723 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
724 KASSERT(p->p_singlethread != NULL,
725 ("singlethread not set"));
727 * The only suspension in action is a
728 * single-threading. Single threader need not stop.
729 * XXX Should be safe to access unlocked
730 * as it can only be set to be true by us.
732 if (p->p_singlethread == td)
733 return (0); /* Exempt from stopping. */
735 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
738 /* Should we goto user boundary if we didn't come from there? */
739 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
740 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
743 /* If thread will exit, flush its pending signals */
744 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
745 sigqueue_flush(&td->td_sigqueue);
750 * If the process is waiting for us to exit,
751 * this thread should just suicide.
752 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
754 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
756 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
757 if (p->p_numthreads == p->p_suspcount + 1) {
758 thread_lock(p->p_singlethread);
760 thread_unsuspend_one(p->p_singlethread);
761 thread_unlock(p->p_singlethread);
769 * When a thread suspends, it just
770 * gets taken off all queues.
772 thread_suspend_one(td);
773 if (return_instead == 0) {
774 p->p_boundary_count++;
775 td->td_flags |= TDF_BOUNDARY;
778 mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
779 if (return_instead == 0)
780 td->td_flags &= ~TDF_BOUNDARY;
783 if (return_instead == 0) {
785 p->p_boundary_count--;
793 thread_suspend_switch(struct thread *td)
798 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
799 PROC_LOCK_ASSERT(p, MA_OWNED);
800 PROC_SLOCK_ASSERT(p, MA_OWNED);
802 * We implement thread_suspend_one in stages here to avoid
803 * dropping the proc lock while the thread lock is owned.
809 td->td_flags &= ~TDF_NEEDSUSPCHK;
810 TD_SET_SUSPENDED(td);
814 mi_switch(SW_VOL | SWT_SUSPEND, NULL);
822 thread_suspend_one(struct thread *td)
824 struct proc *p = td->td_proc;
826 PROC_SLOCK_ASSERT(p, MA_OWNED);
827 THREAD_LOCK_ASSERT(td, MA_OWNED);
828 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
830 td->td_flags &= ~TDF_NEEDSUSPCHK;
831 TD_SET_SUSPENDED(td);
836 thread_unsuspend_one(struct thread *td)
838 struct proc *p = td->td_proc;
840 PROC_SLOCK_ASSERT(p, MA_OWNED);
841 THREAD_LOCK_ASSERT(td, MA_OWNED);
842 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
843 TD_CLR_SUSPENDED(td);
845 return (setrunnable(td));
849 * Allow all threads blocked by single threading to continue running.
852 thread_unsuspend(struct proc *p)
857 PROC_LOCK_ASSERT(p, MA_OWNED);
858 PROC_SLOCK_ASSERT(p, MA_OWNED);
860 if (!P_SHOULDSTOP(p)) {
861 FOREACH_THREAD_IN_PROC(p, td) {
863 if (TD_IS_SUSPENDED(td)) {
864 wakeup_swapper |= thread_unsuspend_one(td);
868 } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
869 (p->p_numthreads == p->p_suspcount)) {
871 * Stopping everything also did the job for the single
872 * threading request. Now we've downgraded to single-threaded,
875 thread_lock(p->p_singlethread);
876 wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
877 thread_unlock(p->p_singlethread);
884 * End the single threading mode..
887 thread_single_end(void)
895 PROC_LOCK_ASSERT(p, MA_OWNED);
896 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
898 p->p_singlethread = NULL;
901 * If there are other threads they may now run,
902 * unless of course there is a blanket 'stop order'
903 * on the process. The single threader must be allowed
904 * to continue however as this is a bad place to stop.
906 if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
907 FOREACH_THREAD_IN_PROC(p, td) {
909 if (TD_IS_SUSPENDED(td)) {
910 wakeup_swapper |= thread_unsuspend_one(td);
921 thread_find(struct proc *p, lwpid_t tid)
925 PROC_LOCK_ASSERT(p, MA_OWNED);
926 FOREACH_THREAD_IN_PROC(p, td) {
927 if (td->td_tid == tid)