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_hwpmc_hooks.h"
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
39 #include <sys/mutex.h>
41 #include <sys/resourcevar.h>
43 #include <sys/sched.h>
44 #include <sys/sleepqueue.h>
45 #include <sys/selinfo.h>
46 #include <sys/turnstile.h>
49 #include <sys/cpuset.h>
51 #include <sys/pmckern.h>
54 #include <security/audit/audit.h>
57 #include <vm/vm_extern.h>
59 #include <sys/eventhandler.h>
62 * thread related storage.
64 static uma_zone_t thread_zone;
66 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
67 static struct mtx zombie_lock;
68 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
70 static void thread_zombie(struct thread *);
73 static struct unrhdr *tid_unrhdr;
76 * Prepare a thread for use.
79 thread_ctor(void *mem, int size, void *arg, int flags)
83 td = (struct thread *)mem;
84 td->td_state = TDS_INACTIVE;
87 td->td_tid = alloc_unr(tid_unrhdr);
91 * Note that td_critnest begins life as 1 because the thread is not
92 * running and is thereby implicitly waiting to be on the receiving
93 * end of a context switch.
96 EVENTHANDLER_INVOKE(thread_ctor, td);
98 audit_thread_alloc(td);
100 umtx_thread_alloc(td);
105 * Reclaim a thread after use.
108 thread_dtor(void *mem, int size, void *arg)
112 td = (struct thread *)mem;
115 /* Verify that this thread is in a safe state to free. */
116 switch (td->td_state) {
122 * We must never unlink a thread that is in one of
123 * these states, because it is currently active.
125 panic("bad state for thread unlinking");
130 panic("bad thread state");
135 audit_thread_free(td);
137 /* Free all OSD associated to this thread. */
140 EVENTHANDLER_INVOKE(thread_dtor, td);
141 free_unr(tid_unrhdr, td->td_tid);
145 * Initialize type-stable parts of a thread (when newly created).
148 thread_init(void *mem, int size, int flags)
152 td = (struct thread *)mem;
154 td->td_sleepqueue = sleepq_alloc();
155 td->td_turnstile = turnstile_alloc();
156 EVENTHANDLER_INVOKE(thread_init, td);
157 td->td_sched = (struct td_sched *)&td[1];
158 umtx_thread_init(td);
164 * Tear down type-stable parts of a thread (just before being discarded).
167 thread_fini(void *mem, int size)
171 td = (struct thread *)mem;
172 EVENTHANDLER_INVOKE(thread_fini, td);
173 turnstile_free(td->td_turnstile);
174 sleepq_free(td->td_sleepqueue);
175 umtx_thread_fini(td);
180 * For a newly created process,
181 * link up all the structures and its initial threads etc.
183 * {arch}/{arch}/machdep.c ia64_init(), init386() etc.
184 * proc_dtor() (should go away)
188 proc_linkup0(struct proc *p, struct thread *td)
190 TAILQ_INIT(&p->p_threads); /* all threads in proc */
195 proc_linkup(struct proc *p, struct thread *td)
198 sigqueue_init(&p->p_sigqueue, p);
199 p->p_ksi = ksiginfo_alloc(1);
200 if (p->p_ksi != NULL) {
201 /* XXX p_ksi may be null if ksiginfo zone is not ready */
202 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
204 LIST_INIT(&p->p_mqnotifier);
210 * Initialize global thread allocation resources.
216 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
217 /* leave one number for thread0 */
218 tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
220 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
221 thread_ctor, thread_dtor, thread_init, thread_fini,
226 * Place an unused thread on the zombie list.
227 * Use the slpq as that must be unused by now.
230 thread_zombie(struct thread *td)
232 mtx_lock_spin(&zombie_lock);
233 TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
234 mtx_unlock_spin(&zombie_lock);
238 * Release a thread that has exited after cpu_throw().
241 thread_stash(struct thread *td)
243 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
248 * Reap zombie resources.
253 struct thread *td_first, *td_next;
256 * Don't even bother to lock if none at this instant,
257 * we really don't care about the next instant..
259 if (!TAILQ_EMPTY(&zombie_threads)) {
260 mtx_lock_spin(&zombie_lock);
261 td_first = TAILQ_FIRST(&zombie_threads);
263 TAILQ_INIT(&zombie_threads);
264 mtx_unlock_spin(&zombie_lock);
266 td_next = TAILQ_NEXT(td_first, td_slpq);
267 if (td_first->td_ucred)
268 crfree(td_first->td_ucred);
269 thread_free(td_first);
279 thread_alloc(int pages)
283 thread_reap(); /* check if any zombies to get */
285 td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
286 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
287 if (!vm_thread_new(td, pages)) {
288 uma_zfree(thread_zone, td);
291 cpu_thread_alloc(td);
296 thread_alloc_stack(struct thread *td, int pages)
299 KASSERT(td->td_kstack == 0,
300 ("thread_alloc_stack called on a thread with kstack"));
301 if (!vm_thread_new(td, pages))
303 cpu_thread_alloc(td);
308 * Deallocate a thread.
311 thread_free(struct thread *td)
314 lock_profile_thread_exit(td);
316 cpuset_rel(td->td_cpuset);
317 td->td_cpuset = NULL;
319 if (td->td_kstack != 0)
320 vm_thread_dispose(td);
321 uma_zfree(thread_zone, td);
325 * Discard the current thread and exit from its context.
326 * Always called with scheduler locked.
328 * Because we can't free a thread while we're operating under its context,
329 * push the current thread into our CPU's deadthread holder. This means
330 * we needn't worry about someone else grabbing our context before we
336 uint64_t runtime, new_switchtime;
345 PROC_SLOCK_ASSERT(p, MA_OWNED);
346 mtx_assert(&Giant, MA_NOTOWNED);
348 PROC_LOCK_ASSERT(p, MA_OWNED);
349 KASSERT(p != NULL, ("thread exiting without a process"));
350 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
351 (long)p->p_pid, td->td_name);
352 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
355 AUDIT_SYSCALL_EXIT(0, td);
357 umtx_thread_exit(td);
359 * drop FPU & debug register state storage, or any other
360 * architecture specific resources that
361 * would not be on a new untouched process.
363 cpu_thread_exit(td); /* XXXSMP */
366 * The last thread is left attached to the process
367 * So that the whole bundle gets recycled. Skip
368 * all this stuff if we never had threads.
369 * EXIT clears all sign of other threads when
370 * it goes to single threading, so the last thread always
371 * takes the short path.
373 if (p->p_flag & P_HADTHREADS) {
374 if (p->p_numthreads > 1) {
376 td2 = FIRST_THREAD_IN_PROC(p);
377 sched_exit_thread(td2, td);
380 * The test below is NOT true if we are the
381 * sole exiting thread. P_STOPPED_SINGLE is unset
382 * in exit1() after it is the only survivor.
384 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
385 if (p->p_numthreads == p->p_suspcount) {
386 thread_lock(p->p_singlethread);
387 wakeup_swapper = thread_unsuspend_one(
389 thread_unlock(p->p_singlethread);
395 atomic_add_int(&td->td_proc->p_exitthreads, 1);
396 PCPU_SET(deadthread, td);
399 * The last thread is exiting.. but not through exit()
401 panic ("thread_exit: Last thread exiting on its own");
406 * If this thread is part of a process that is being tracked by hwpmc(4),
407 * inform the module of the thread's impending exit.
409 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
410 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
414 /* Do the same timestamp bookkeeping that mi_switch() would do. */
415 new_switchtime = cpu_ticks();
416 runtime = new_switchtime - PCPU_GET(switchtime);
417 td->td_runtime += runtime;
418 td->td_incruntime += runtime;
419 PCPU_SET(switchtime, new_switchtime);
420 PCPU_SET(switchticks, ticks);
421 PCPU_INC(cnt.v_swtch);
423 /* Save our resource usage in our process. */
424 td->td_ru.ru_nvcsw++;
426 rucollect(&p->p_ru, &td->td_ru);
430 td->td_state = TDS_INACTIVE;
432 witness_thread_exit(td);
434 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
436 panic("I'm a teapot!");
441 * Do any thread specific cleanups that may be needed in wait()
442 * called with Giant, proc and schedlock not held.
445 thread_wait(struct proc *p)
449 mtx_assert(&Giant, MA_NOTOWNED);
450 KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
451 td = FIRST_THREAD_IN_PROC(p);
452 /* Lock the last thread so we spin until it exits cpu_throw(). */
455 /* Wait for any remaining threads to exit cpu_throw(). */
456 while (p->p_exitthreads)
457 sched_relinquish(curthread);
458 lock_profile_thread_exit(td);
459 cpuset_rel(td->td_cpuset);
460 td->td_cpuset = NULL;
461 cpu_thread_clean(td);
462 crfree(td->td_ucred);
463 thread_reap(); /* check for zombie threads etc. */
467 * Link a thread to a process.
468 * set up anything that needs to be initialized for it to
469 * be used by the process.
472 thread_link(struct thread *td, struct proc *p)
476 * XXX This can't be enabled because it's called for proc0 before
477 * its lock has been created.
478 * PROC_LOCK_ASSERT(p, MA_OWNED);
480 td->td_state = TDS_INACTIVE;
482 td->td_flags = TDF_INMEM;
484 LIST_INIT(&td->td_contested);
485 LIST_INIT(&td->td_lprof[0]);
486 LIST_INIT(&td->td_lprof[1]);
487 sigqueue_init(&td->td_sigqueue, p);
488 callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
489 TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
494 * Convert a process with one thread to an unthreaded process.
497 thread_unthread(struct thread *td)
499 struct proc *p = td->td_proc;
501 KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
502 p->p_flag &= ~P_HADTHREADS;
510 thread_unlink(struct thread *td)
512 struct proc *p = td->td_proc;
514 PROC_LOCK_ASSERT(p, MA_OWNED);
515 TAILQ_REMOVE(&p->p_threads, td, td_plist);
517 /* could clear a few other things here */
518 /* Must NOT clear links to proc! */
522 calc_remaining(struct proc *p, int mode)
526 PROC_LOCK_ASSERT(p, MA_OWNED);
527 PROC_SLOCK_ASSERT(p, MA_OWNED);
528 if (mode == SINGLE_EXIT)
529 remaining = p->p_numthreads;
530 else if (mode == SINGLE_BOUNDARY)
531 remaining = p->p_numthreads - p->p_boundary_count;
532 else if (mode == SINGLE_NO_EXIT)
533 remaining = p->p_numthreads - p->p_suspcount;
535 panic("calc_remaining: wrong mode %d", mode);
540 * Enforce single-threading.
542 * Returns 1 if the caller must abort (another thread is waiting to
543 * exit the process or similar). Process is locked!
544 * Returns 0 when you are successfully the only thread running.
545 * A process has successfully single threaded in the suspend mode when
546 * There are no threads in user mode. Threads in the kernel must be
547 * allowed to continue until they get to the user boundary. They may even
548 * copy out their return values and data before suspending. They may however be
549 * accelerated in reaching the user boundary as we will wake up
550 * any sleeping threads that are interruptable. (PCATCH).
553 thread_single(int mode)
558 int remaining, wakeup_swapper;
562 mtx_assert(&Giant, MA_NOTOWNED);
563 PROC_LOCK_ASSERT(p, MA_OWNED);
564 KASSERT((td != NULL), ("curthread is NULL"));
566 if ((p->p_flag & P_HADTHREADS) == 0)
569 /* Is someone already single threading? */
570 if (p->p_singlethread != NULL && p->p_singlethread != td)
573 if (mode == SINGLE_EXIT) {
574 p->p_flag |= P_SINGLE_EXIT;
575 p->p_flag &= ~P_SINGLE_BOUNDARY;
577 p->p_flag &= ~P_SINGLE_EXIT;
578 if (mode == SINGLE_BOUNDARY)
579 p->p_flag |= P_SINGLE_BOUNDARY;
581 p->p_flag &= ~P_SINGLE_BOUNDARY;
583 p->p_flag |= P_STOPPED_SINGLE;
585 p->p_singlethread = td;
586 remaining = calc_remaining(p, mode);
587 while (remaining != 1) {
588 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
591 FOREACH_THREAD_IN_PROC(p, td2) {
595 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
596 if (TD_IS_INHIBITED(td2)) {
599 if (TD_IS_SUSPENDED(td2))
601 thread_unsuspend_one(td2);
602 if (TD_ON_SLEEPQ(td2) &&
603 (td2->td_flags & TDF_SINTR))
605 sleepq_abort(td2, EINTR);
607 case SINGLE_BOUNDARY:
608 if (TD_IS_SUSPENDED(td2) &&
609 !(td2->td_flags & TDF_BOUNDARY))
611 thread_unsuspend_one(td2);
612 if (TD_ON_SLEEPQ(td2) &&
613 (td2->td_flags & TDF_SINTR))
615 sleepq_abort(td2, ERESTART);
618 if (TD_IS_SUSPENDED(td2) &&
619 !(td2->td_flags & TDF_BOUNDARY))
621 thread_unsuspend_one(td2);
622 if (TD_ON_SLEEPQ(td2) &&
623 (td2->td_flags & TDF_SINTR))
625 sleepq_abort(td2, ERESTART);
632 else if (TD_IS_RUNNING(td2) && td != td2) {
640 remaining = calc_remaining(p, mode);
643 * Maybe we suspended some threads.. was it enough?
650 * Wake us up when everyone else has suspended.
651 * In the mean time we suspend as well.
653 thread_suspend_switch(td);
654 remaining = calc_remaining(p, mode);
656 if (mode == SINGLE_EXIT) {
658 * We have gotten rid of all the other threads and we
659 * are about to either exit or exec. In either case,
660 * we try our utmost to revert to being a non-threaded
663 p->p_singlethread = NULL;
664 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
672 * Called in from locations that can safely check to see
673 * whether we have to suspend or at least throttle for a
674 * single-thread event (e.g. fork).
676 * Such locations include userret().
677 * If the "return_instead" argument is non zero, the thread must be able to
678 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
680 * The 'return_instead' argument tells the function if it may do a
681 * thread_exit() or suspend, or whether the caller must abort and back
684 * If the thread that set the single_threading request has set the
685 * P_SINGLE_EXIT bit in the process flags then this call will never return
686 * if 'return_instead' is false, but will exit.
688 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
689 *---------------+--------------------+---------------------
690 * 0 | returns 0 | returns 0 or 1
691 * | when ST ends | immediatly
692 *---------------+--------------------+---------------------
693 * 1 | thread exits | returns 1
695 * 0 = thread_exit() or suspension ok,
696 * other = return error instead of stopping the thread.
698 * While a full suspension is under effect, even a single threading
699 * thread would be suspended if it made this call (but it shouldn't).
700 * This call should only be made from places where
701 * thread_exit() would be safe as that may be the outcome unless
702 * return_instead is set.
705 thread_suspend_check(int return_instead)
713 mtx_assert(&Giant, MA_NOTOWNED);
714 PROC_LOCK_ASSERT(p, MA_OWNED);
715 while (P_SHOULDSTOP(p) ||
716 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
717 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
718 KASSERT(p->p_singlethread != NULL,
719 ("singlethread not set"));
721 * The only suspension in action is a
722 * single-threading. Single threader need not stop.
723 * XXX Should be safe to access unlocked
724 * as it can only be set to be true by us.
726 if (p->p_singlethread == td)
727 return (0); /* Exempt from stopping. */
729 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
732 /* Should we goto user boundary if we didn't come from there? */
733 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
734 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
737 /* If thread will exit, flush its pending signals */
738 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
739 sigqueue_flush(&td->td_sigqueue);
744 * If the process is waiting for us to exit,
745 * this thread should just suicide.
746 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
748 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
750 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
751 if (p->p_numthreads == p->p_suspcount + 1) {
752 thread_lock(p->p_singlethread);
754 thread_unsuspend_one(p->p_singlethread);
755 thread_unlock(p->p_singlethread);
763 * When a thread suspends, it just
764 * gets taken off all queues.
766 thread_suspend_one(td);
767 if (return_instead == 0) {
768 p->p_boundary_count++;
769 td->td_flags |= TDF_BOUNDARY;
772 mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
773 if (return_instead == 0)
774 td->td_flags &= ~TDF_BOUNDARY;
777 if (return_instead == 0) {
779 p->p_boundary_count--;
787 thread_suspend_switch(struct thread *td)
792 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
793 PROC_LOCK_ASSERT(p, MA_OWNED);
794 PROC_SLOCK_ASSERT(p, MA_OWNED);
796 * We implement thread_suspend_one in stages here to avoid
797 * dropping the proc lock while the thread lock is owned.
803 td->td_flags &= ~TDF_NEEDSUSPCHK;
804 TD_SET_SUSPENDED(td);
808 mi_switch(SW_VOL | SWT_SUSPEND, NULL);
816 thread_suspend_one(struct thread *td)
818 struct proc *p = td->td_proc;
820 PROC_SLOCK_ASSERT(p, MA_OWNED);
821 THREAD_LOCK_ASSERT(td, MA_OWNED);
822 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
824 td->td_flags &= ~TDF_NEEDSUSPCHK;
825 TD_SET_SUSPENDED(td);
830 thread_unsuspend_one(struct thread *td)
832 struct proc *p = td->td_proc;
834 PROC_SLOCK_ASSERT(p, MA_OWNED);
835 THREAD_LOCK_ASSERT(td, MA_OWNED);
836 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
837 TD_CLR_SUSPENDED(td);
839 return (setrunnable(td));
843 * Allow all threads blocked by single threading to continue running.
846 thread_unsuspend(struct proc *p)
851 PROC_LOCK_ASSERT(p, MA_OWNED);
852 PROC_SLOCK_ASSERT(p, MA_OWNED);
854 if (!P_SHOULDSTOP(p)) {
855 FOREACH_THREAD_IN_PROC(p, td) {
857 if (TD_IS_SUSPENDED(td)) {
858 wakeup_swapper |= thread_unsuspend_one(td);
862 } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
863 (p->p_numthreads == p->p_suspcount)) {
865 * Stopping everything also did the job for the single
866 * threading request. Now we've downgraded to single-threaded,
869 thread_lock(p->p_singlethread);
870 wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
871 thread_unlock(p->p_singlethread);
878 * End the single threading mode..
881 thread_single_end(void)
889 PROC_LOCK_ASSERT(p, MA_OWNED);
890 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
892 p->p_singlethread = NULL;
895 * If there are other threads they may now run,
896 * unless of course there is a blanket 'stop order'
897 * on the process. The single threader must be allowed
898 * to continue however as this is a bad place to stop.
900 if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
901 FOREACH_THREAD_IN_PROC(p, td) {
903 if (TD_IS_SUSPENDED(td)) {
904 wakeup_swapper |= thread_unsuspend_one(td);
915 thread_find(struct proc *p, lwpid_t tid)
919 PROC_LOCK_ASSERT(p, MA_OWNED);
920 FOREACH_THREAD_IN_PROC(p, td) {
921 if (td->td_tid == tid)