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 EVENTHANDLER_INVOKE(thread_dtor, td);
145 free_unr(tid_unrhdr, td->td_tid);
149 * Initialize type-stable parts of a thread (when newly created).
152 thread_init(void *mem, int size, int flags)
156 td = (struct thread *)mem;
158 td->td_sleepqueue = sleepq_alloc();
159 td->td_turnstile = turnstile_alloc();
160 EVENTHANDLER_INVOKE(thread_init, td);
161 td->td_sched = (struct td_sched *)&td[1];
162 umtx_thread_init(td);
168 * Tear down type-stable parts of a thread (just before being discarded).
171 thread_fini(void *mem, int size)
175 td = (struct thread *)mem;
176 EVENTHANDLER_INVOKE(thread_fini, td);
177 turnstile_free(td->td_turnstile);
178 sleepq_free(td->td_sleepqueue);
179 umtx_thread_fini(td);
184 * For a newly created process,
185 * link up all the structures and its initial threads etc.
187 * {arch}/{arch}/machdep.c ia64_init(), init386() etc.
188 * proc_dtor() (should go away)
192 proc_linkup0(struct proc *p, struct thread *td)
194 TAILQ_INIT(&p->p_threads); /* all threads in proc */
199 proc_linkup(struct proc *p, struct thread *td)
202 sigqueue_init(&p->p_sigqueue, p);
203 p->p_ksi = ksiginfo_alloc(1);
204 if (p->p_ksi != NULL) {
205 /* XXX p_ksi may be null if ksiginfo zone is not ready */
206 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
208 LIST_INIT(&p->p_mqnotifier);
214 * Initialize global thread allocation resources.
220 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
221 /* leave one number for thread0 */
222 tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
224 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
225 thread_ctor, thread_dtor, thread_init, thread_fini,
230 * Place an unused thread on the zombie list.
231 * Use the slpq as that must be unused by now.
234 thread_zombie(struct thread *td)
236 mtx_lock_spin(&zombie_lock);
237 TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
238 mtx_unlock_spin(&zombie_lock);
242 * Release a thread that has exited after cpu_throw().
245 thread_stash(struct thread *td)
247 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
252 * Reap zombie resources.
257 struct thread *td_first, *td_next;
260 * Don't even bother to lock if none at this instant,
261 * we really don't care about the next instant..
263 if (!TAILQ_EMPTY(&zombie_threads)) {
264 mtx_lock_spin(&zombie_lock);
265 td_first = TAILQ_FIRST(&zombie_threads);
267 TAILQ_INIT(&zombie_threads);
268 mtx_unlock_spin(&zombie_lock);
270 td_next = TAILQ_NEXT(td_first, td_slpq);
271 if (td_first->td_ucred)
272 crfree(td_first->td_ucred);
273 thread_free(td_first);
287 thread_reap(); /* check if any zombies to get */
289 td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
290 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
291 if (!vm_thread_new(td, 0)) {
292 uma_zfree(thread_zone, td);
295 cpu_thread_alloc(td);
301 * Deallocate a thread.
304 thread_free(struct thread *td)
307 cpuset_rel(td->td_cpuset);
308 td->td_cpuset = NULL;
310 if (td->td_altkstack != 0)
311 vm_thread_dispose_altkstack(td);
312 if (td->td_kstack != 0)
313 vm_thread_dispose(td);
314 uma_zfree(thread_zone, td);
318 * Discard the current thread and exit from its context.
319 * Always called with scheduler locked.
321 * Because we can't free a thread while we're operating under its context,
322 * push the current thread into our CPU's deadthread holder. This means
323 * we needn't worry about someone else grabbing our context before we
329 uint64_t new_switchtime;
338 PROC_SLOCK_ASSERT(p, MA_OWNED);
339 mtx_assert(&Giant, MA_NOTOWNED);
341 PROC_LOCK_ASSERT(p, MA_OWNED);
342 KASSERT(p != NULL, ("thread exiting without a process"));
343 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
344 (long)p->p_pid, td->td_name);
345 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
348 AUDIT_SYSCALL_EXIT(0, td);
350 umtx_thread_exit(td);
352 * drop FPU & debug register state storage, or any other
353 * architecture specific resources that
354 * would not be on a new untouched process.
356 cpu_thread_exit(td); /* XXXSMP */
358 /* Do the same timestamp bookkeeping that mi_switch() would do. */
359 new_switchtime = cpu_ticks();
360 p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
361 PCPU_SET(switchtime, new_switchtime);
362 PCPU_SET(switchticks, ticks);
363 PCPU_INC(cnt.v_swtch);
364 /* Save our resource usage in our process. */
365 td->td_ru.ru_nvcsw++;
366 rucollect(&p->p_ru, &td->td_ru);
368 * The last thread is left attached to the process
369 * So that the whole bundle gets recycled. Skip
370 * all this stuff if we never had threads.
371 * EXIT clears all sign of other threads when
372 * it goes to single threading, so the last thread always
373 * takes the short path.
375 if (p->p_flag & P_HADTHREADS) {
376 if (p->p_numthreads > 1) {
378 td2 = FIRST_THREAD_IN_PROC(p);
379 sched_exit_thread(td2, td);
382 * The test below is NOT true if we are the
383 * sole exiting thread. P_STOPPED_SNGL is unset
384 * in exit1() after it is the only survivor.
386 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
387 if (p->p_numthreads == p->p_suspcount) {
388 thread_lock(p->p_singlethread);
389 wakeup_swapper = thread_unsuspend_one(
391 thread_unlock(p->p_singlethread);
397 atomic_add_int(&td->td_proc->p_exitthreads, 1);
398 PCPU_SET(deadthread, td);
401 * The last thread is exiting.. but not through exit()
403 panic ("thread_exit: Last thread exiting on its own");
408 /* Save our tick information with both the thread and proc locked */
409 ruxagg(&p->p_rux, td);
411 td->td_state = TDS_INACTIVE;
413 witness_thread_exit(td);
415 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
417 panic("I'm a teapot!");
422 * Do any thread specific cleanups that may be needed in wait()
423 * called with Giant, proc and schedlock not held.
426 thread_wait(struct proc *p)
430 mtx_assert(&Giant, MA_NOTOWNED);
431 KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
432 td = FIRST_THREAD_IN_PROC(p);
433 /* Lock the last thread so we spin until it exits cpu_throw(). */
436 /* Wait for any remaining threads to exit cpu_throw(). */
437 while (p->p_exitthreads)
438 sched_relinquish(curthread);
439 cpuset_rel(td->td_cpuset);
440 td->td_cpuset = NULL;
441 cpu_thread_clean(td);
442 crfree(td->td_ucred);
443 thread_reap(); /* check for zombie threads etc. */
447 * Link a thread to a process.
448 * set up anything that needs to be initialized for it to
449 * be used by the process.
452 thread_link(struct thread *td, struct proc *p)
456 * XXX This can't be enabled because it's called for proc0 before
457 * its lock has been created.
458 * PROC_LOCK_ASSERT(p, MA_OWNED);
460 td->td_state = TDS_INACTIVE;
462 td->td_flags = TDF_INMEM;
464 LIST_INIT(&td->td_contested);
465 LIST_INIT(&td->td_lprof[0]);
466 LIST_INIT(&td->td_lprof[1]);
467 sigqueue_init(&td->td_sigqueue, p);
468 callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
469 TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
474 * Convert a process with one thread to an unthreaded process.
477 thread_unthread(struct thread *td)
479 struct proc *p = td->td_proc;
481 KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
482 p->p_flag &= ~P_HADTHREADS;
490 thread_unlink(struct thread *td)
492 struct proc *p = td->td_proc;
494 PROC_LOCK_ASSERT(p, MA_OWNED);
495 TAILQ_REMOVE(&p->p_threads, td, td_plist);
497 /* could clear a few other things here */
498 /* Must NOT clear links to proc! */
502 * Enforce single-threading.
504 * Returns 1 if the caller must abort (another thread is waiting to
505 * exit the process or similar). Process is locked!
506 * Returns 0 when you are successfully the only thread running.
507 * A process has successfully single threaded in the suspend mode when
508 * There are no threads in user mode. Threads in the kernel must be
509 * allowed to continue until they get to the user boundary. They may even
510 * copy out their return values and data before suspending. They may however be
511 * accelerated in reaching the user boundary as we will wake up
512 * any sleeping threads that are interruptable. (PCATCH).
515 thread_single(int mode)
520 int remaining, wakeup_swapper;
524 mtx_assert(&Giant, MA_NOTOWNED);
525 PROC_LOCK_ASSERT(p, MA_OWNED);
526 KASSERT((td != NULL), ("curthread is NULL"));
528 if ((p->p_flag & P_HADTHREADS) == 0)
531 /* Is someone already single threading? */
532 if (p->p_singlethread != NULL && p->p_singlethread != td)
535 if (mode == SINGLE_EXIT) {
536 p->p_flag |= P_SINGLE_EXIT;
537 p->p_flag &= ~P_SINGLE_BOUNDARY;
539 p->p_flag &= ~P_SINGLE_EXIT;
540 if (mode == SINGLE_BOUNDARY)
541 p->p_flag |= P_SINGLE_BOUNDARY;
543 p->p_flag &= ~P_SINGLE_BOUNDARY;
545 p->p_flag |= P_STOPPED_SINGLE;
547 p->p_singlethread = td;
548 if (mode == SINGLE_EXIT)
549 remaining = p->p_numthreads;
550 else if (mode == SINGLE_BOUNDARY)
551 remaining = p->p_numthreads - p->p_boundary_count;
553 remaining = p->p_numthreads - p->p_suspcount;
554 while (remaining != 1) {
555 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
558 FOREACH_THREAD_IN_PROC(p, td2) {
562 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
563 if (TD_IS_INHIBITED(td2)) {
566 if (TD_IS_SUSPENDED(td2))
568 thread_unsuspend_one(td2);
569 if (TD_ON_SLEEPQ(td2) &&
570 (td2->td_flags & TDF_SINTR))
572 sleepq_abort(td2, EINTR);
574 case SINGLE_BOUNDARY:
575 if (TD_IS_SUSPENDED(td2) &&
576 !(td2->td_flags & TDF_BOUNDARY))
578 thread_unsuspend_one(td2);
579 if (TD_ON_SLEEPQ(td2) &&
580 (td2->td_flags & TDF_SINTR))
582 sleepq_abort(td2, ERESTART);
585 if (TD_IS_SUSPENDED(td2)) {
590 * maybe other inhibited states too?
592 if ((td2->td_flags & TDF_SINTR) &&
593 (td2->td_inhibitors &
594 (TDI_SLEEPING | TDI_SWAPPED)))
595 thread_suspend_one(td2);
600 else if (TD_IS_RUNNING(td2) && td != td2) {
608 if (mode == SINGLE_EXIT)
609 remaining = p->p_numthreads;
610 else if (mode == SINGLE_BOUNDARY)
611 remaining = p->p_numthreads - p->p_boundary_count;
613 remaining = p->p_numthreads - p->p_suspcount;
616 * Maybe we suspended some threads.. was it enough?
623 * Wake us up when everyone else has suspended.
624 * In the mean time we suspend as well.
626 thread_suspend_switch(td);
627 if (mode == SINGLE_EXIT)
628 remaining = p->p_numthreads;
629 else if (mode == SINGLE_BOUNDARY)
630 remaining = p->p_numthreads - p->p_boundary_count;
632 remaining = p->p_numthreads - p->p_suspcount;
634 if (mode == SINGLE_EXIT) {
636 * We have gotten rid of all the other threads and we
637 * are about to either exit or exec. In either case,
638 * we try our utmost to revert to being a non-threaded
641 p->p_singlethread = NULL;
642 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
650 * Called in from locations that can safely check to see
651 * whether we have to suspend or at least throttle for a
652 * single-thread event (e.g. fork).
654 * Such locations include userret().
655 * If the "return_instead" argument is non zero, the thread must be able to
656 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
658 * The 'return_instead' argument tells the function if it may do a
659 * thread_exit() or suspend, or whether the caller must abort and back
662 * If the thread that set the single_threading request has set the
663 * P_SINGLE_EXIT bit in the process flags then this call will never return
664 * if 'return_instead' is false, but will exit.
666 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
667 *---------------+--------------------+---------------------
668 * 0 | returns 0 | returns 0 or 1
669 * | when ST ends | immediatly
670 *---------------+--------------------+---------------------
671 * 1 | thread exits | returns 1
673 * 0 = thread_exit() or suspension ok,
674 * other = return error instead of stopping the thread.
676 * While a full suspension is under effect, even a single threading
677 * thread would be suspended if it made this call (but it shouldn't).
678 * This call should only be made from places where
679 * thread_exit() would be safe as that may be the outcome unless
680 * return_instead is set.
683 thread_suspend_check(int return_instead)
691 mtx_assert(&Giant, MA_NOTOWNED);
692 PROC_LOCK_ASSERT(p, MA_OWNED);
693 while (P_SHOULDSTOP(p) ||
694 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
695 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
696 KASSERT(p->p_singlethread != NULL,
697 ("singlethread not set"));
699 * The only suspension in action is a
700 * single-threading. Single threader need not stop.
701 * XXX Should be safe to access unlocked
702 * as it can only be set to be true by us.
704 if (p->p_singlethread == td)
705 return (0); /* Exempt from stopping. */
707 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
710 /* Should we goto user boundary if we didn't come from there? */
711 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
712 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
715 /* If thread will exit, flush its pending signals */
716 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
717 sigqueue_flush(&td->td_sigqueue);
722 * If the process is waiting for us to exit,
723 * this thread should just suicide.
724 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
726 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
728 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
729 if (p->p_numthreads == p->p_suspcount + 1) {
730 thread_lock(p->p_singlethread);
732 thread_unsuspend_one(p->p_singlethread);
733 thread_unlock(p->p_singlethread);
741 * When a thread suspends, it just
742 * gets taken off all queues.
744 thread_suspend_one(td);
745 if (return_instead == 0) {
746 p->p_boundary_count++;
747 td->td_flags |= TDF_BOUNDARY;
750 mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
751 if (return_instead == 0)
752 td->td_flags &= ~TDF_BOUNDARY;
755 if (return_instead == 0)
756 p->p_boundary_count--;
762 thread_suspend_switch(struct thread *td)
767 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
768 PROC_LOCK_ASSERT(p, MA_OWNED);
769 PROC_SLOCK_ASSERT(p, MA_OWNED);
771 * We implement thread_suspend_one in stages here to avoid
772 * dropping the proc lock while the thread lock is owned.
778 td->td_flags &= ~TDF_NEEDSUSPCHK;
779 TD_SET_SUSPENDED(td);
783 mi_switch(SW_VOL | SWT_SUSPEND, NULL);
791 thread_suspend_one(struct thread *td)
793 struct proc *p = td->td_proc;
795 PROC_SLOCK_ASSERT(p, MA_OWNED);
796 THREAD_LOCK_ASSERT(td, MA_OWNED);
797 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
799 td->td_flags &= ~TDF_NEEDSUSPCHK;
800 TD_SET_SUSPENDED(td);
805 thread_unsuspend_one(struct thread *td)
807 struct proc *p = td->td_proc;
809 PROC_SLOCK_ASSERT(p, MA_OWNED);
810 THREAD_LOCK_ASSERT(td, MA_OWNED);
811 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
812 TD_CLR_SUSPENDED(td);
814 return (setrunnable(td));
818 * Allow all threads blocked by single threading to continue running.
821 thread_unsuspend(struct proc *p)
826 PROC_LOCK_ASSERT(p, MA_OWNED);
827 PROC_SLOCK_ASSERT(p, MA_OWNED);
829 if (!P_SHOULDSTOP(p)) {
830 FOREACH_THREAD_IN_PROC(p, td) {
832 if (TD_IS_SUSPENDED(td)) {
833 wakeup_swapper |= thread_unsuspend_one(td);
837 } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
838 (p->p_numthreads == p->p_suspcount)) {
840 * Stopping everything also did the job for the single
841 * threading request. Now we've downgraded to single-threaded,
844 thread_lock(p->p_singlethread);
845 wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
846 thread_unlock(p->p_singlethread);
853 * End the single threading mode..
856 thread_single_end(void)
864 PROC_LOCK_ASSERT(p, MA_OWNED);
865 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
867 p->p_singlethread = NULL;
870 * If there are other threads they may now run,
871 * unless of course there is a blanket 'stop order'
872 * on the process. The single threader must be allowed
873 * to continue however as this is a bad place to stop.
875 if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
876 FOREACH_THREAD_IN_PROC(p, td) {
878 if (TD_IS_SUSPENDED(td)) {
879 wakeup_swapper |= thread_unsuspend_one(td);
890 thread_find(struct proc *p, lwpid_t tid)
894 PROC_LOCK_ASSERT(p, MA_OWNED);
895 FOREACH_THREAD_IN_PROC(p, td) {
896 if (td->td_tid == tid)