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 td->td_vp_reserv = 0;
163 EVENTHANDLER_INVOKE(thread_init, td);
164 td->td_sched = (struct td_sched *)&td[1];
165 umtx_thread_init(td);
171 * Tear down type-stable parts of a thread (just before being discarded).
174 thread_fini(void *mem, int size)
178 td = (struct thread *)mem;
179 EVENTHANDLER_INVOKE(thread_fini, td);
180 turnstile_free(td->td_turnstile);
181 sleepq_free(td->td_sleepqueue);
182 umtx_thread_fini(td);
187 * For a newly created process,
188 * link up all the structures and its initial threads etc.
190 * {arch}/{arch}/machdep.c ia64_init(), init386() etc.
191 * proc_dtor() (should go away)
195 proc_linkup0(struct proc *p, struct thread *td)
197 TAILQ_INIT(&p->p_threads); /* all threads in proc */
202 proc_linkup(struct proc *p, struct thread *td)
205 sigqueue_init(&p->p_sigqueue, p);
206 p->p_ksi = ksiginfo_alloc(1);
207 if (p->p_ksi != NULL) {
208 /* XXX p_ksi may be null if ksiginfo zone is not ready */
209 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
211 LIST_INIT(&p->p_mqnotifier);
217 * Initialize global thread allocation resources.
223 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
224 /* leave one number for thread0 */
225 tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
227 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
228 thread_ctor, thread_dtor, thread_init, thread_fini,
233 * Place an unused thread on the zombie list.
234 * Use the slpq as that must be unused by now.
237 thread_zombie(struct thread *td)
239 mtx_lock_spin(&zombie_lock);
240 TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
241 mtx_unlock_spin(&zombie_lock);
245 * Release a thread that has exited after cpu_throw().
248 thread_stash(struct thread *td)
250 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
255 * Reap zombie resources.
260 struct thread *td_first, *td_next;
263 * Don't even bother to lock if none at this instant,
264 * we really don't care about the next instant..
266 if (!TAILQ_EMPTY(&zombie_threads)) {
267 mtx_lock_spin(&zombie_lock);
268 td_first = TAILQ_FIRST(&zombie_threads);
270 TAILQ_INIT(&zombie_threads);
271 mtx_unlock_spin(&zombie_lock);
273 td_next = TAILQ_NEXT(td_first, td_slpq);
274 if (td_first->td_ucred)
275 crfree(td_first->td_ucred);
276 thread_free(td_first);
286 thread_alloc(int pages)
290 thread_reap(); /* check if any zombies to get */
292 td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
293 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
294 if (!vm_thread_new(td, pages)) {
295 uma_zfree(thread_zone, td);
298 cpu_thread_alloc(td);
303 thread_alloc_stack(struct thread *td, int pages)
306 KASSERT(td->td_kstack == 0,
307 ("thread_alloc_stack called on a thread with kstack"));
308 if (!vm_thread_new(td, pages))
310 cpu_thread_alloc(td);
315 * Deallocate a thread.
318 thread_free(struct thread *td)
321 lock_profile_thread_exit(td);
323 cpuset_rel(td->td_cpuset);
324 td->td_cpuset = NULL;
326 if (td->td_kstack != 0)
327 vm_thread_dispose(td);
328 uma_zfree(thread_zone, td);
332 * Discard the current thread and exit from its context.
333 * Always called with scheduler locked.
335 * Because we can't free a thread while we're operating under its context,
336 * push the current thread into our CPU's deadthread holder. This means
337 * we needn't worry about someone else grabbing our context before we
343 uint64_t runtime, new_switchtime;
352 PROC_SLOCK_ASSERT(p, MA_OWNED);
353 mtx_assert(&Giant, MA_NOTOWNED);
355 PROC_LOCK_ASSERT(p, MA_OWNED);
356 KASSERT(p != NULL, ("thread exiting without a process"));
357 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
358 (long)p->p_pid, td->td_name);
359 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
362 AUDIT_SYSCALL_EXIT(0, td);
364 umtx_thread_exit(td);
366 * drop FPU & debug register state storage, or any other
367 * architecture specific resources that
368 * would not be on a new untouched process.
370 cpu_thread_exit(td); /* XXXSMP */
373 * The last thread is left attached to the process
374 * So that the whole bundle gets recycled. Skip
375 * all this stuff if we never had threads.
376 * EXIT clears all sign of other threads when
377 * it goes to single threading, so the last thread always
378 * takes the short path.
380 if (p->p_flag & P_HADTHREADS) {
381 if (p->p_numthreads > 1) {
383 td2 = FIRST_THREAD_IN_PROC(p);
384 sched_exit_thread(td2, td);
387 * The test below is NOT true if we are the
388 * sole exiting thread. P_STOPPED_SINGLE is unset
389 * in exit1() after it is the only survivor.
391 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
392 if (p->p_numthreads == p->p_suspcount) {
393 thread_lock(p->p_singlethread);
394 wakeup_swapper = thread_unsuspend_one(
396 thread_unlock(p->p_singlethread);
402 atomic_add_int(&td->td_proc->p_exitthreads, 1);
403 PCPU_SET(deadthread, td);
406 * The last thread is exiting.. but not through exit()
408 panic ("thread_exit: Last thread exiting on its own");
413 * If this thread is part of a process that is being tracked by hwpmc(4),
414 * inform the module of the thread's impending exit.
416 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
417 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
421 /* Do the same timestamp bookkeeping that mi_switch() would do. */
422 new_switchtime = cpu_ticks();
423 runtime = new_switchtime - PCPU_GET(switchtime);
424 td->td_runtime += runtime;
425 td->td_incruntime += runtime;
426 PCPU_SET(switchtime, new_switchtime);
427 PCPU_SET(switchticks, ticks);
428 PCPU_INC(cnt.v_swtch);
430 /* Save our resource usage in our process. */
431 td->td_ru.ru_nvcsw++;
433 rucollect(&p->p_ru, &td->td_ru);
437 td->td_state = TDS_INACTIVE;
439 witness_thread_exit(td);
441 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
443 panic("I'm a teapot!");
448 * Do any thread specific cleanups that may be needed in wait()
449 * called with Giant, proc and schedlock not held.
452 thread_wait(struct proc *p)
456 mtx_assert(&Giant, MA_NOTOWNED);
457 KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
458 td = FIRST_THREAD_IN_PROC(p);
459 /* Lock the last thread so we spin until it exits cpu_throw(). */
462 /* Wait for any remaining threads to exit cpu_throw(). */
463 while (p->p_exitthreads)
464 sched_relinquish(curthread);
465 lock_profile_thread_exit(td);
466 cpuset_rel(td->td_cpuset);
467 td->td_cpuset = NULL;
468 cpu_thread_clean(td);
469 crfree(td->td_ucred);
470 thread_reap(); /* check for zombie threads etc. */
474 * Link a thread to a process.
475 * set up anything that needs to be initialized for it to
476 * be used by the process.
479 thread_link(struct thread *td, struct proc *p)
483 * XXX This can't be enabled because it's called for proc0 before
484 * its lock has been created.
485 * PROC_LOCK_ASSERT(p, MA_OWNED);
487 td->td_state = TDS_INACTIVE;
489 td->td_flags = TDF_INMEM;
491 LIST_INIT(&td->td_contested);
492 LIST_INIT(&td->td_lprof[0]);
493 LIST_INIT(&td->td_lprof[1]);
494 sigqueue_init(&td->td_sigqueue, p);
495 callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
496 TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
501 * Convert a process with one thread to an unthreaded process.
504 thread_unthread(struct thread *td)
506 struct proc *p = td->td_proc;
508 KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
509 p->p_flag &= ~P_HADTHREADS;
517 thread_unlink(struct thread *td)
519 struct proc *p = td->td_proc;
521 PROC_LOCK_ASSERT(p, MA_OWNED);
522 TAILQ_REMOVE(&p->p_threads, td, td_plist);
524 /* could clear a few other things here */
525 /* Must NOT clear links to proc! */
529 calc_remaining(struct proc *p, int mode)
533 PROC_LOCK_ASSERT(p, MA_OWNED);
534 PROC_SLOCK_ASSERT(p, MA_OWNED);
535 if (mode == SINGLE_EXIT)
536 remaining = p->p_numthreads;
537 else if (mode == SINGLE_BOUNDARY)
538 remaining = p->p_numthreads - p->p_boundary_count;
539 else if (mode == SINGLE_NO_EXIT)
540 remaining = p->p_numthreads - p->p_suspcount;
542 panic("calc_remaining: wrong mode %d", mode);
547 * Enforce single-threading.
549 * Returns 1 if the caller must abort (another thread is waiting to
550 * exit the process or similar). Process is locked!
551 * Returns 0 when you are successfully the only thread running.
552 * A process has successfully single threaded in the suspend mode when
553 * There are no threads in user mode. Threads in the kernel must be
554 * allowed to continue until they get to the user boundary. They may even
555 * copy out their return values and data before suspending. They may however be
556 * accelerated in reaching the user boundary as we will wake up
557 * any sleeping threads that are interruptable. (PCATCH).
560 thread_single(int mode)
565 int remaining, wakeup_swapper;
569 mtx_assert(&Giant, MA_NOTOWNED);
570 PROC_LOCK_ASSERT(p, MA_OWNED);
571 KASSERT((td != NULL), ("curthread is NULL"));
573 if ((p->p_flag & P_HADTHREADS) == 0)
576 /* Is someone already single threading? */
577 if (p->p_singlethread != NULL && p->p_singlethread != td)
580 if (mode == SINGLE_EXIT) {
581 p->p_flag |= P_SINGLE_EXIT;
582 p->p_flag &= ~P_SINGLE_BOUNDARY;
584 p->p_flag &= ~P_SINGLE_EXIT;
585 if (mode == SINGLE_BOUNDARY)
586 p->p_flag |= P_SINGLE_BOUNDARY;
588 p->p_flag &= ~P_SINGLE_BOUNDARY;
590 p->p_flag |= P_STOPPED_SINGLE;
592 p->p_singlethread = td;
593 remaining = calc_remaining(p, mode);
594 while (remaining != 1) {
595 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
598 FOREACH_THREAD_IN_PROC(p, td2) {
602 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
603 if (TD_IS_INHIBITED(td2)) {
606 if (TD_IS_SUSPENDED(td2))
608 thread_unsuspend_one(td2);
609 if (TD_ON_SLEEPQ(td2) &&
610 (td2->td_flags & TDF_SINTR))
612 sleepq_abort(td2, EINTR);
614 case SINGLE_BOUNDARY:
615 if (TD_IS_SUSPENDED(td2) &&
616 !(td2->td_flags & TDF_BOUNDARY))
618 thread_unsuspend_one(td2);
619 if (TD_ON_SLEEPQ(td2) &&
620 (td2->td_flags & TDF_SINTR))
622 sleepq_abort(td2, ERESTART);
625 if (TD_IS_SUSPENDED(td2) &&
626 !(td2->td_flags & TDF_BOUNDARY))
628 thread_unsuspend_one(td2);
629 if (TD_ON_SLEEPQ(td2) &&
630 (td2->td_flags & TDF_SINTR))
632 sleepq_abort(td2, ERESTART);
639 else if (TD_IS_RUNNING(td2) && td != td2) {
647 remaining = calc_remaining(p, mode);
650 * Maybe we suspended some threads.. was it enough?
657 * Wake us up when everyone else has suspended.
658 * In the mean time we suspend as well.
660 thread_suspend_switch(td);
661 remaining = calc_remaining(p, mode);
663 if (mode == SINGLE_EXIT) {
665 * We have gotten rid of all the other threads and we
666 * are about to either exit or exec. In either case,
667 * we try our utmost to revert to being a non-threaded
670 p->p_singlethread = NULL;
671 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
679 * Called in from locations that can safely check to see
680 * whether we have to suspend or at least throttle for a
681 * single-thread event (e.g. fork).
683 * Such locations include userret().
684 * If the "return_instead" argument is non zero, the thread must be able to
685 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
687 * The 'return_instead' argument tells the function if it may do a
688 * thread_exit() or suspend, or whether the caller must abort and back
691 * If the thread that set the single_threading request has set the
692 * P_SINGLE_EXIT bit in the process flags then this call will never return
693 * if 'return_instead' is false, but will exit.
695 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
696 *---------------+--------------------+---------------------
697 * 0 | returns 0 | returns 0 or 1
698 * | when ST ends | immediatly
699 *---------------+--------------------+---------------------
700 * 1 | thread exits | returns 1
702 * 0 = thread_exit() or suspension ok,
703 * other = return error instead of stopping the thread.
705 * While a full suspension is under effect, even a single threading
706 * thread would be suspended if it made this call (but it shouldn't).
707 * This call should only be made from places where
708 * thread_exit() would be safe as that may be the outcome unless
709 * return_instead is set.
712 thread_suspend_check(int return_instead)
720 mtx_assert(&Giant, MA_NOTOWNED);
721 PROC_LOCK_ASSERT(p, MA_OWNED);
722 while (P_SHOULDSTOP(p) ||
723 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
724 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
725 KASSERT(p->p_singlethread != NULL,
726 ("singlethread not set"));
728 * The only suspension in action is a
729 * single-threading. Single threader need not stop.
730 * XXX Should be safe to access unlocked
731 * as it can only be set to be true by us.
733 if (p->p_singlethread == td)
734 return (0); /* Exempt from stopping. */
736 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
739 /* Should we goto user boundary if we didn't come from there? */
740 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
741 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
745 * Ignore suspend requests for stop signals if they
748 if (P_SHOULDSTOP(p) == P_STOPPED_SIG &&
749 td->td_flags & TDF_SBDRY) {
750 KASSERT(return_instead,
751 ("TDF_SBDRY set for unsafe thread_suspend_check"));
755 /* If thread will exit, flush its pending signals */
756 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
757 sigqueue_flush(&td->td_sigqueue);
762 * If the process is waiting for us to exit,
763 * this thread should just suicide.
764 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
766 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
768 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
769 if (p->p_numthreads == p->p_suspcount + 1) {
770 thread_lock(p->p_singlethread);
772 thread_unsuspend_one(p->p_singlethread);
773 thread_unlock(p->p_singlethread);
781 * When a thread suspends, it just
782 * gets taken off all queues.
784 thread_suspend_one(td);
785 if (return_instead == 0) {
786 p->p_boundary_count++;
787 td->td_flags |= TDF_BOUNDARY;
790 mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
791 if (return_instead == 0)
792 td->td_flags &= ~TDF_BOUNDARY;
795 if (return_instead == 0) {
797 p->p_boundary_count--;
805 thread_suspend_switch(struct thread *td)
810 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
811 PROC_LOCK_ASSERT(p, MA_OWNED);
812 PROC_SLOCK_ASSERT(p, MA_OWNED);
814 * We implement thread_suspend_one in stages here to avoid
815 * dropping the proc lock while the thread lock is owned.
821 td->td_flags &= ~TDF_NEEDSUSPCHK;
822 TD_SET_SUSPENDED(td);
826 mi_switch(SW_VOL | SWT_SUSPEND, NULL);
834 thread_suspend_one(struct thread *td)
836 struct proc *p = td->td_proc;
838 PROC_SLOCK_ASSERT(p, MA_OWNED);
839 THREAD_LOCK_ASSERT(td, MA_OWNED);
840 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
842 td->td_flags &= ~TDF_NEEDSUSPCHK;
843 TD_SET_SUSPENDED(td);
848 thread_unsuspend_one(struct thread *td)
850 struct proc *p = td->td_proc;
852 PROC_SLOCK_ASSERT(p, MA_OWNED);
853 THREAD_LOCK_ASSERT(td, MA_OWNED);
854 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
855 TD_CLR_SUSPENDED(td);
857 return (setrunnable(td));
861 * Allow all threads blocked by single threading to continue running.
864 thread_unsuspend(struct proc *p)
869 PROC_LOCK_ASSERT(p, MA_OWNED);
870 PROC_SLOCK_ASSERT(p, MA_OWNED);
872 if (!P_SHOULDSTOP(p)) {
873 FOREACH_THREAD_IN_PROC(p, td) {
875 if (TD_IS_SUSPENDED(td)) {
876 wakeup_swapper |= thread_unsuspend_one(td);
880 } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
881 (p->p_numthreads == p->p_suspcount)) {
883 * Stopping everything also did the job for the single
884 * threading request. Now we've downgraded to single-threaded,
887 thread_lock(p->p_singlethread);
888 wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
889 thread_unlock(p->p_singlethread);
896 * End the single threading mode..
899 thread_single_end(void)
907 PROC_LOCK_ASSERT(p, MA_OWNED);
908 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
910 p->p_singlethread = NULL;
913 * If there are other threads they may now run,
914 * unless of course there is a blanket 'stop order'
915 * on the process. The single threader must be allowed
916 * to continue however as this is a bad place to stop.
918 if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
919 FOREACH_THREAD_IN_PROC(p, td) {
921 if (TD_IS_SUSPENDED(td)) {
922 wakeup_swapper |= thread_unsuspend_one(td);
933 thread_find(struct proc *p, lwpid_t tid)
937 PROC_LOCK_ASSERT(p, MA_OWNED);
938 FOREACH_THREAD_IN_PROC(p, td) {
939 if (td->td_tid == tid)