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"
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
38 #include <sys/mutex.h>
40 #include <sys/resourcevar.h>
42 #include <sys/sysctl.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 <security/audit/audit.h>
54 #include <vm/vm_extern.h>
56 #include <sys/eventhandler.h>
59 * thread related storage.
61 static uma_zone_t thread_zone;
63 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
65 int max_threads_per_proc = 1500;
66 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
67 &max_threads_per_proc, 0, "Limit on threads per proc");
70 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
71 &max_threads_hits, 0, "");
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);
77 static void thread_zombie(struct thread *);
80 static struct unrhdr *tid_unrhdr;
83 * Prepare a thread for use.
86 thread_ctor(void *mem, int size, void *arg, int flags)
90 td = (struct thread *)mem;
91 td->td_state = TDS_INACTIVE;
94 td->td_tid = alloc_unr(tid_unrhdr);
98 * Note that td_critnest begins life as 1 because the thread is not
99 * running and is thereby implicitly waiting to be on the receiving
100 * end of a context switch.
103 EVENTHANDLER_INVOKE(thread_ctor, td);
105 audit_thread_alloc(td);
107 umtx_thread_alloc(td);
112 * Reclaim a thread after use.
115 thread_dtor(void *mem, int size, void *arg)
119 td = (struct thread *)mem;
122 /* Verify that this thread is in a safe state to free. */
123 switch (td->td_state) {
129 * We must never unlink a thread that is in one of
130 * these states, because it is currently active.
132 panic("bad state for thread unlinking");
137 panic("bad thread state");
142 audit_thread_free(td);
144 /* Free all OSD associated to this thread. */
147 EVENTHANDLER_INVOKE(thread_dtor, td);
148 free_unr(tid_unrhdr, td->td_tid);
152 * Initialize type-stable parts of a thread (when newly created).
155 thread_init(void *mem, int size, int flags)
159 td = (struct thread *)mem;
161 td->td_sleepqueue = sleepq_alloc();
162 td->td_turnstile = turnstile_alloc();
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 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 */
372 /* Do the same timestamp bookkeeping that mi_switch() would do. */
373 new_switchtime = cpu_ticks();
374 p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
375 PCPU_SET(switchtime, new_switchtime);
376 PCPU_SET(switchticks, ticks);
377 PCPU_INC(cnt.v_swtch);
378 /* Save our resource usage in our process. */
379 td->td_ru.ru_nvcsw++;
380 rucollect(&p->p_ru, &td->td_ru);
382 * The last thread is left attached to the process
383 * So that the whole bundle gets recycled. Skip
384 * all this stuff if we never had threads.
385 * EXIT clears all sign of other threads when
386 * it goes to single threading, so the last thread always
387 * takes the short path.
389 if (p->p_flag & P_HADTHREADS) {
390 if (p->p_numthreads > 1) {
392 td2 = FIRST_THREAD_IN_PROC(p);
393 sched_exit_thread(td2, td);
396 * The test below is NOT true if we are the
397 * sole exiting thread. P_STOPPED_SNGL is unset
398 * in exit1() after it is the only survivor.
400 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
401 if (p->p_numthreads == p->p_suspcount) {
402 thread_lock(p->p_singlethread);
403 wakeup_swapper = thread_unsuspend_one(
405 thread_unlock(p->p_singlethread);
411 atomic_add_int(&td->td_proc->p_exitthreads, 1);
412 PCPU_SET(deadthread, td);
415 * The last thread is exiting.. but not through exit()
417 panic ("thread_exit: Last thread exiting on its own");
422 /* Save our tick information with both the thread and proc locked */
423 ruxagg(&p->p_rux, td);
425 td->td_state = TDS_INACTIVE;
427 witness_thread_exit(td);
429 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
431 panic("I'm a teapot!");
436 * Do any thread specific cleanups that may be needed in wait()
437 * called with Giant, proc and schedlock not held.
440 thread_wait(struct proc *p)
444 mtx_assert(&Giant, MA_NOTOWNED);
445 KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
446 td = FIRST_THREAD_IN_PROC(p);
447 /* Lock the last thread so we spin until it exits cpu_throw(). */
450 /* Wait for any remaining threads to exit cpu_throw(). */
451 while (p->p_exitthreads)
452 sched_relinquish(curthread);
453 lock_profile_thread_exit(td);
454 cpuset_rel(td->td_cpuset);
455 td->td_cpuset = NULL;
456 cpu_thread_clean(td);
457 crfree(td->td_ucred);
458 thread_reap(); /* check for zombie threads etc. */
462 * Link a thread to a process.
463 * set up anything that needs to be initialized for it to
464 * be used by the process.
467 thread_link(struct thread *td, struct proc *p)
471 * XXX This can't be enabled because it's called for proc0 before
472 * its lock has been created.
473 * PROC_LOCK_ASSERT(p, MA_OWNED);
475 td->td_state = TDS_INACTIVE;
477 td->td_flags = TDF_INMEM;
479 LIST_INIT(&td->td_contested);
480 LIST_INIT(&td->td_lprof[0]);
481 LIST_INIT(&td->td_lprof[1]);
482 sigqueue_init(&td->td_sigqueue, p);
483 callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
484 TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
489 * Convert a process with one thread to an unthreaded process.
492 thread_unthread(struct thread *td)
494 struct proc *p = td->td_proc;
496 KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
497 p->p_flag &= ~P_HADTHREADS;
505 thread_unlink(struct thread *td)
507 struct proc *p = td->td_proc;
509 PROC_LOCK_ASSERT(p, MA_OWNED);
510 TAILQ_REMOVE(&p->p_threads, td, td_plist);
512 /* could clear a few other things here */
513 /* Must NOT clear links to proc! */
517 calc_remaining(struct proc *p, int mode)
521 if (mode == SINGLE_EXIT)
522 remaining = p->p_numthreads;
523 else if (mode == SINGLE_BOUNDARY)
524 remaining = p->p_numthreads - p->p_boundary_count;
525 else if (mode == SINGLE_NO_EXIT)
526 remaining = p->p_numthreads - p->p_suspcount;
528 panic("calc_remaining: wrong mode %d", mode);
533 * Enforce single-threading.
535 * Returns 1 if the caller must abort (another thread is waiting to
536 * exit the process or similar). Process is locked!
537 * Returns 0 when you are successfully the only thread running.
538 * A process has successfully single threaded in the suspend mode when
539 * There are no threads in user mode. Threads in the kernel must be
540 * allowed to continue until they get to the user boundary. They may even
541 * copy out their return values and data before suspending. They may however be
542 * accelerated in reaching the user boundary as we will wake up
543 * any sleeping threads that are interruptable. (PCATCH).
546 thread_single(int mode)
551 int remaining, wakeup_swapper;
555 mtx_assert(&Giant, MA_NOTOWNED);
556 PROC_LOCK_ASSERT(p, MA_OWNED);
557 KASSERT((td != NULL), ("curthread is NULL"));
559 if ((p->p_flag & P_HADTHREADS) == 0)
562 /* Is someone already single threading? */
563 if (p->p_singlethread != NULL && p->p_singlethread != td)
566 if (mode == SINGLE_EXIT) {
567 p->p_flag |= P_SINGLE_EXIT;
568 p->p_flag &= ~P_SINGLE_BOUNDARY;
570 p->p_flag &= ~P_SINGLE_EXIT;
571 if (mode == SINGLE_BOUNDARY)
572 p->p_flag |= P_SINGLE_BOUNDARY;
574 p->p_flag &= ~P_SINGLE_BOUNDARY;
576 p->p_flag |= P_STOPPED_SINGLE;
578 p->p_singlethread = td;
579 remaining = calc_remaining(p, mode);
580 while (remaining != 1) {
581 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
584 FOREACH_THREAD_IN_PROC(p, td2) {
588 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
589 if (TD_IS_INHIBITED(td2)) {
592 if (TD_IS_SUSPENDED(td2))
594 thread_unsuspend_one(td2);
595 if (TD_ON_SLEEPQ(td2) &&
596 (td2->td_flags & TDF_SINTR))
598 sleepq_abort(td2, EINTR);
600 case SINGLE_BOUNDARY:
601 if (TD_IS_SUSPENDED(td2) &&
602 !(td2->td_flags & TDF_BOUNDARY))
604 thread_unsuspend_one(td2);
605 if (TD_ON_SLEEPQ(td2) &&
606 (td2->td_flags & TDF_SINTR))
608 sleepq_abort(td2, ERESTART);
611 if (TD_IS_SUSPENDED(td2) &&
612 !(td2->td_flags & TDF_BOUNDARY))
614 thread_unsuspend_one(td2);
615 if (TD_ON_SLEEPQ(td2) &&
616 (td2->td_flags & TDF_SINTR))
618 sleepq_abort(td2, ERESTART);
625 else if (TD_IS_RUNNING(td2) && td != td2) {
633 remaining = calc_remaining(p, mode);
636 * Maybe we suspended some threads.. was it enough?
643 * Wake us up when everyone else has suspended.
644 * In the mean time we suspend as well.
646 thread_suspend_switch(td);
647 remaining = calc_remaining(p, mode);
649 if (mode == SINGLE_EXIT) {
651 * We have gotten rid of all the other threads and we
652 * are about to either exit or exec. In either case,
653 * we try our utmost to revert to being a non-threaded
656 p->p_singlethread = NULL;
657 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
665 * Called in from locations that can safely check to see
666 * whether we have to suspend or at least throttle for a
667 * single-thread event (e.g. fork).
669 * Such locations include userret().
670 * If the "return_instead" argument is non zero, the thread must be able to
671 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
673 * The 'return_instead' argument tells the function if it may do a
674 * thread_exit() or suspend, or whether the caller must abort and back
677 * If the thread that set the single_threading request has set the
678 * P_SINGLE_EXIT bit in the process flags then this call will never return
679 * if 'return_instead' is false, but will exit.
681 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
682 *---------------+--------------------+---------------------
683 * 0 | returns 0 | returns 0 or 1
684 * | when ST ends | immediatly
685 *---------------+--------------------+---------------------
686 * 1 | thread exits | returns 1
688 * 0 = thread_exit() or suspension ok,
689 * other = return error instead of stopping the thread.
691 * While a full suspension is under effect, even a single threading
692 * thread would be suspended if it made this call (but it shouldn't).
693 * This call should only be made from places where
694 * thread_exit() would be safe as that may be the outcome unless
695 * return_instead is set.
698 thread_suspend_check(int return_instead)
706 mtx_assert(&Giant, MA_NOTOWNED);
707 PROC_LOCK_ASSERT(p, MA_OWNED);
708 while (P_SHOULDSTOP(p) ||
709 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
710 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
711 KASSERT(p->p_singlethread != NULL,
712 ("singlethread not set"));
714 * The only suspension in action is a
715 * single-threading. Single threader need not stop.
716 * XXX Should be safe to access unlocked
717 * as it can only be set to be true by us.
719 if (p->p_singlethread == td)
720 return (0); /* Exempt from stopping. */
722 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
725 /* Should we goto user boundary if we didn't come from there? */
726 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
727 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
730 /* If thread will exit, flush its pending signals */
731 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
732 sigqueue_flush(&td->td_sigqueue);
737 * If the process is waiting for us to exit,
738 * this thread should just suicide.
739 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
741 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
743 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
744 if (p->p_numthreads == p->p_suspcount + 1) {
745 thread_lock(p->p_singlethread);
747 thread_unsuspend_one(p->p_singlethread);
748 thread_unlock(p->p_singlethread);
756 * When a thread suspends, it just
757 * gets taken off all queues.
759 thread_suspend_one(td);
760 if (return_instead == 0) {
761 p->p_boundary_count++;
762 td->td_flags |= TDF_BOUNDARY;
765 mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
766 if (return_instead == 0)
767 td->td_flags &= ~TDF_BOUNDARY;
770 if (return_instead == 0)
771 p->p_boundary_count--;
777 thread_suspend_switch(struct thread *td)
782 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
783 PROC_LOCK_ASSERT(p, MA_OWNED);
784 PROC_SLOCK_ASSERT(p, MA_OWNED);
786 * We implement thread_suspend_one in stages here to avoid
787 * dropping the proc lock while the thread lock is owned.
793 td->td_flags &= ~TDF_NEEDSUSPCHK;
794 TD_SET_SUSPENDED(td);
798 mi_switch(SW_VOL | SWT_SUSPEND, NULL);
806 thread_suspend_one(struct thread *td)
808 struct proc *p = td->td_proc;
810 PROC_SLOCK_ASSERT(p, MA_OWNED);
811 THREAD_LOCK_ASSERT(td, MA_OWNED);
812 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
814 td->td_flags &= ~TDF_NEEDSUSPCHK;
815 TD_SET_SUSPENDED(td);
820 thread_unsuspend_one(struct thread *td)
822 struct proc *p = td->td_proc;
824 PROC_SLOCK_ASSERT(p, MA_OWNED);
825 THREAD_LOCK_ASSERT(td, MA_OWNED);
826 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
827 TD_CLR_SUSPENDED(td);
829 return (setrunnable(td));
833 * Allow all threads blocked by single threading to continue running.
836 thread_unsuspend(struct proc *p)
841 PROC_LOCK_ASSERT(p, MA_OWNED);
842 PROC_SLOCK_ASSERT(p, MA_OWNED);
844 if (!P_SHOULDSTOP(p)) {
845 FOREACH_THREAD_IN_PROC(p, td) {
847 if (TD_IS_SUSPENDED(td)) {
848 wakeup_swapper |= thread_unsuspend_one(td);
852 } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
853 (p->p_numthreads == p->p_suspcount)) {
855 * Stopping everything also did the job for the single
856 * threading request. Now we've downgraded to single-threaded,
859 thread_lock(p->p_singlethread);
860 wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
861 thread_unlock(p->p_singlethread);
868 * End the single threading mode..
871 thread_single_end(void)
879 PROC_LOCK_ASSERT(p, MA_OWNED);
880 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
882 p->p_singlethread = NULL;
885 * If there are other threads they may now run,
886 * unless of course there is a blanket 'stop order'
887 * on the process. The single threader must be allowed
888 * to continue however as this is a bad place to stop.
890 if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
891 FOREACH_THREAD_IN_PROC(p, td) {
893 if (TD_IS_SUSPENDED(td)) {
894 wakeup_swapper |= thread_unsuspend_one(td);
905 thread_find(struct proc *p, lwpid_t tid)
909 PROC_LOCK_ASSERT(p, MA_OWNED);
910 FOREACH_THREAD_IN_PROC(p, td) {
911 if (td->td_tid == tid)