2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice(s), this list of conditions and the following disclaimer as
12 * the first lines of this file unmodified other than the possible
13 * addition of one or more copyright notices.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice(s), this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
19 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
22 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
25 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
31 #include "opt_witness.h"
32 #include "opt_hwpmc_hooks.h"
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
41 #include <sys/mutex.h>
43 #include <sys/epoch.h>
44 #include <sys/rangelock.h>
45 #include <sys/resourcevar.h>
48 #include <sys/sched.h>
49 #include <sys/sleepqueue.h>
50 #include <sys/selinfo.h>
51 #include <sys/syscallsubr.h>
52 #include <sys/sysent.h>
53 #include <sys/turnstile.h>
55 #include <sys/rwlock.h>
57 #include <sys/vmmeter.h>
58 #include <sys/cpuset.h>
60 #include <sys/pmckern.h>
63 #include <security/audit/audit.h>
66 #include <vm/vm_extern.h>
68 #include <sys/eventhandler.h>
71 * Asserts below verify the stability of struct thread and struct proc
72 * layout, as exposed by KBI to modules. On head, the KBI is allowed
73 * to drift, change to the structures must be accompanied by the
76 * On the stable branches after KBI freeze, conditions must not be
77 * violated. Typically new fields are moved to the end of the
81 _Static_assert(offsetof(struct thread, td_flags) == 0xfc,
82 "struct thread KBI td_flags");
83 _Static_assert(offsetof(struct thread, td_pflags) == 0x104,
84 "struct thread KBI td_pflags");
85 _Static_assert(offsetof(struct thread, td_frame) == 0x498,
86 "struct thread KBI td_frame");
87 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6b0,
88 "struct thread KBI td_emuldata");
89 _Static_assert(offsetof(struct proc, p_flag) == 0xb0,
90 "struct proc KBI p_flag");
91 _Static_assert(offsetof(struct proc, p_pid) == 0xbc,
92 "struct proc KBI p_pid");
93 _Static_assert(offsetof(struct proc, p_filemon) == 0x3c8,
94 "struct proc KBI p_filemon");
95 _Static_assert(offsetof(struct proc, p_comm) == 0x3e0,
96 "struct proc KBI p_comm");
97 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4c0,
98 "struct proc KBI p_emuldata");
101 _Static_assert(offsetof(struct thread, td_flags) == 0x98,
102 "struct thread KBI td_flags");
103 _Static_assert(offsetof(struct thread, td_pflags) == 0xa0,
104 "struct thread KBI td_pflags");
105 _Static_assert(offsetof(struct thread, td_frame) == 0x2fc,
106 "struct thread KBI td_frame");
107 _Static_assert(offsetof(struct thread, td_emuldata) == 0x344,
108 "struct thread KBI td_emuldata");
109 _Static_assert(offsetof(struct proc, p_flag) == 0x68,
110 "struct proc KBI p_flag");
111 _Static_assert(offsetof(struct proc, p_pid) == 0x74,
112 "struct proc KBI p_pid");
113 _Static_assert(offsetof(struct proc, p_filemon) == 0x278,
114 "struct proc KBI p_filemon");
115 _Static_assert(offsetof(struct proc, p_comm) == 0x28c,
116 "struct proc KBI p_comm");
117 _Static_assert(offsetof(struct proc, p_emuldata) == 0x318,
118 "struct proc KBI p_emuldata");
121 SDT_PROVIDER_DECLARE(proc);
122 SDT_PROBE_DEFINE(proc, , , lwp__exit);
125 * thread related storage.
127 static uma_zone_t thread_zone;
129 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
130 static struct mtx zombie_lock;
131 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
133 static void thread_zombie(struct thread *);
134 static int thread_unsuspend_one(struct thread *td, struct proc *p,
137 #define TID_BUFFER_SIZE 1024
140 static struct unrhdr *tid_unrhdr;
141 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
142 static int tid_head, tid_tail;
143 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
145 struct tidhashhead *tidhashtbl;
147 struct rwlock tidhash_lock;
149 EVENTHANDLER_LIST_DEFINE(thread_ctor);
150 EVENTHANDLER_LIST_DEFINE(thread_dtor);
151 EVENTHANDLER_LIST_DEFINE(thread_init);
152 EVENTHANDLER_LIST_DEFINE(thread_fini);
159 tid = alloc_unr(tid_unrhdr);
163 if (tid_head == tid_tail) {
164 mtx_unlock(&tid_lock);
167 tid = tid_buffer[tid_head];
168 tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
169 mtx_unlock(&tid_lock);
174 tid_free(lwpid_t tid)
176 lwpid_t tmp_tid = -1;
179 if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
180 tmp_tid = tid_buffer[tid_head];
181 tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
183 tid_buffer[tid_tail] = tid;
184 tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE;
185 mtx_unlock(&tid_lock);
187 free_unr(tid_unrhdr, tmp_tid);
191 * Prepare a thread for use.
194 thread_ctor(void *mem, int size, void *arg, int flags)
198 td = (struct thread *)mem;
199 td->td_state = TDS_INACTIVE;
200 td->td_lastcpu = td->td_oncpu = NOCPU;
202 td->td_tid = tid_alloc();
205 * Note that td_critnest begins life as 1 because the thread is not
206 * running and is thereby implicitly waiting to be on the receiving
207 * end of a context switch.
210 td->td_lend_user_pri = PRI_MAX;
211 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
213 audit_thread_alloc(td);
215 umtx_thread_alloc(td);
220 * Reclaim a thread after use.
223 thread_dtor(void *mem, int size, void *arg)
227 td = (struct thread *)mem;
230 /* Verify that this thread is in a safe state to free. */
231 switch (td->td_state) {
237 * We must never unlink a thread that is in one of
238 * these states, because it is currently active.
240 panic("bad state for thread unlinking");
245 panic("bad thread state");
250 audit_thread_free(td);
252 /* Free all OSD associated to this thread. */
254 td_softdep_cleanup(td);
255 MPASS(td->td_su == NULL);
257 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
258 tid_free(td->td_tid);
262 * Initialize type-stable parts of a thread (when newly created).
265 thread_init(void *mem, int size, int flags)
269 td = (struct thread *)mem;
271 td->td_sleepqueue = sleepq_alloc();
272 td->td_turnstile = turnstile_alloc();
274 EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
275 umtx_thread_init(td);
282 * Tear down type-stable parts of a thread (just before being discarded).
285 thread_fini(void *mem, int size)
289 td = (struct thread *)mem;
290 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
291 rlqentry_free(td->td_rlqe);
292 turnstile_free(td->td_turnstile);
293 sleepq_free(td->td_sleepqueue);
294 umtx_thread_fini(td);
299 * For a newly created process,
300 * link up all the structures and its initial threads etc.
302 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc.
303 * proc_dtor() (should go away)
307 proc_linkup0(struct proc *p, struct thread *td)
309 TAILQ_INIT(&p->p_threads); /* all threads in proc */
314 proc_linkup(struct proc *p, struct thread *td)
317 sigqueue_init(&p->p_sigqueue, p);
318 p->p_ksi = ksiginfo_alloc(1);
319 if (p->p_ksi != NULL) {
320 /* XXX p_ksi may be null if ksiginfo zone is not ready */
321 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
323 LIST_INIT(&p->p_mqnotifier);
329 * Initialize global thread allocation resources.
335 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
338 * pid_max cannot be greater than PID_MAX.
339 * leave one number for thread0.
341 tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
343 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
344 thread_ctor, thread_dtor, thread_init, thread_fini,
345 32 - 1, UMA_ZONE_NOFREE);
346 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
347 rw_init(&tidhash_lock, "tidhash");
351 * Place an unused thread on the zombie list.
352 * Use the slpq as that must be unused by now.
355 thread_zombie(struct thread *td)
357 mtx_lock_spin(&zombie_lock);
358 TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
359 mtx_unlock_spin(&zombie_lock);
363 * Release a thread that has exited after cpu_throw().
366 thread_stash(struct thread *td)
368 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
373 * Reap zombie resources.
378 struct thread *td_first, *td_next;
381 * Don't even bother to lock if none at this instant,
382 * we really don't care about the next instant.
384 if (!TAILQ_EMPTY(&zombie_threads)) {
385 mtx_lock_spin(&zombie_lock);
386 td_first = TAILQ_FIRST(&zombie_threads);
388 TAILQ_INIT(&zombie_threads);
389 mtx_unlock_spin(&zombie_lock);
391 td_next = TAILQ_NEXT(td_first, td_slpq);
392 thread_cow_free(td_first);
393 thread_free(td_first);
403 thread_alloc(int pages)
407 thread_reap(); /* check if any zombies to get */
409 td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
410 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
411 if (!vm_thread_new(td, pages)) {
412 uma_zfree(thread_zone, td);
415 cpu_thread_alloc(td);
420 thread_alloc_stack(struct thread *td, int pages)
423 KASSERT(td->td_kstack == 0,
424 ("thread_alloc_stack called on a thread with kstack"));
425 if (!vm_thread_new(td, pages))
427 cpu_thread_alloc(td);
432 * Deallocate a thread.
435 thread_free(struct thread *td)
438 lock_profile_thread_exit(td);
440 cpuset_rel(td->td_cpuset);
441 td->td_cpuset = NULL;
443 if (td->td_kstack != 0)
444 vm_thread_dispose(td);
445 callout_drain(&td->td_slpcallout);
446 uma_zfree(thread_zone, td);
450 thread_cow_get_proc(struct thread *newtd, struct proc *p)
453 PROC_LOCK_ASSERT(p, MA_OWNED);
454 newtd->td_ucred = crhold(p->p_ucred);
455 newtd->td_limit = lim_hold(p->p_limit);
456 newtd->td_cowgen = p->p_cowgen;
460 thread_cow_get(struct thread *newtd, struct thread *td)
463 newtd->td_ucred = crhold(td->td_ucred);
464 newtd->td_limit = lim_hold(td->td_limit);
465 newtd->td_cowgen = td->td_cowgen;
469 thread_cow_free(struct thread *td)
472 if (td->td_ucred != NULL)
473 crfree(td->td_ucred);
474 if (td->td_limit != NULL)
475 lim_free(td->td_limit);
479 thread_cow_update(struct thread *td)
482 struct ucred *oldcred;
483 struct plimit *oldlimit;
489 if (td->td_ucred != p->p_ucred) {
490 oldcred = td->td_ucred;
491 td->td_ucred = crhold(p->p_ucred);
493 if (td->td_limit != p->p_limit) {
494 oldlimit = td->td_limit;
495 td->td_limit = lim_hold(p->p_limit);
497 td->td_cowgen = p->p_cowgen;
501 if (oldlimit != NULL)
506 * Discard the current thread and exit from its context.
507 * Always called with scheduler locked.
509 * Because we can't free a thread while we're operating under its context,
510 * push the current thread into our CPU's deadthread holder. This means
511 * we needn't worry about someone else grabbing our context before we
517 uint64_t runtime, new_switchtime;
526 PROC_SLOCK_ASSERT(p, MA_OWNED);
527 mtx_assert(&Giant, MA_NOTOWNED);
529 PROC_LOCK_ASSERT(p, MA_OWNED);
530 KASSERT(p != NULL, ("thread exiting without a process"));
531 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
532 (long)p->p_pid, td->td_name);
533 SDT_PROBE0(proc, , , lwp__exit);
534 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
537 * drop FPU & debug register state storage, or any other
538 * architecture specific resources that
539 * would not be on a new untouched process.
544 * The last thread is left attached to the process
545 * So that the whole bundle gets recycled. Skip
546 * all this stuff if we never had threads.
547 * EXIT clears all sign of other threads when
548 * it goes to single threading, so the last thread always
549 * takes the short path.
551 if (p->p_flag & P_HADTHREADS) {
552 if (p->p_numthreads > 1) {
553 atomic_add_int(&td->td_proc->p_exitthreads, 1);
555 td2 = FIRST_THREAD_IN_PROC(p);
556 sched_exit_thread(td2, td);
559 * The test below is NOT true if we are the
560 * sole exiting thread. P_STOPPED_SINGLE is unset
561 * in exit1() after it is the only survivor.
563 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
564 if (p->p_numthreads == p->p_suspcount) {
565 thread_lock(p->p_singlethread);
566 wakeup_swapper = thread_unsuspend_one(
567 p->p_singlethread, p, false);
573 PCPU_SET(deadthread, td);
576 * The last thread is exiting.. but not through exit()
578 panic ("thread_exit: Last thread exiting on its own");
583 * If this thread is part of a process that is being tracked by hwpmc(4),
584 * inform the module of the thread's impending exit.
586 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
587 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
588 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
589 } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
590 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
597 /* Do the same timestamp bookkeeping that mi_switch() would do. */
598 new_switchtime = cpu_ticks();
599 runtime = new_switchtime - PCPU_GET(switchtime);
600 td->td_runtime += runtime;
601 td->td_incruntime += runtime;
602 PCPU_SET(switchtime, new_switchtime);
603 PCPU_SET(switchticks, ticks);
606 /* Save our resource usage in our process. */
607 td->td_ru.ru_nvcsw++;
608 ruxagg_locked(p, td);
609 rucollect(&p->p_ru, &td->td_ru);
612 td->td_state = TDS_INACTIVE;
614 witness_thread_exit(td);
616 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
618 panic("I'm a teapot!");
623 * Do any thread specific cleanups that may be needed in wait()
624 * called with Giant, proc and schedlock not held.
627 thread_wait(struct proc *p)
631 mtx_assert(&Giant, MA_NOTOWNED);
632 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
633 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
634 td = FIRST_THREAD_IN_PROC(p);
635 /* Lock the last thread so we spin until it exits cpu_throw(). */
638 lock_profile_thread_exit(td);
639 cpuset_rel(td->td_cpuset);
640 td->td_cpuset = NULL;
641 cpu_thread_clean(td);
643 callout_drain(&td->td_slpcallout);
644 thread_reap(); /* check for zombie threads etc. */
648 * Link a thread to a process.
649 * set up anything that needs to be initialized for it to
650 * be used by the process.
653 thread_link(struct thread *td, struct proc *p)
657 * XXX This can't be enabled because it's called for proc0 before
658 * its lock has been created.
659 * PROC_LOCK_ASSERT(p, MA_OWNED);
661 td->td_state = TDS_INACTIVE;
663 td->td_flags = TDF_INMEM;
665 LIST_INIT(&td->td_contested);
666 LIST_INIT(&td->td_lprof[0]);
667 LIST_INIT(&td->td_lprof[1]);
669 SLIST_INIT(&td->td_epochs);
671 sigqueue_init(&td->td_sigqueue, p);
672 callout_init(&td->td_slpcallout, 1);
673 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
682 thread_unlink(struct thread *td)
684 struct proc *p = td->td_proc;
686 PROC_LOCK_ASSERT(p, MA_OWNED);
688 MPASS(SLIST_EMPTY(&td->td_epochs));
691 TAILQ_REMOVE(&p->p_threads, td, td_plist);
693 /* could clear a few other things here */
694 /* Must NOT clear links to proc! */
698 calc_remaining(struct proc *p, int mode)
702 PROC_LOCK_ASSERT(p, MA_OWNED);
703 PROC_SLOCK_ASSERT(p, MA_OWNED);
704 if (mode == SINGLE_EXIT)
705 remaining = p->p_numthreads;
706 else if (mode == SINGLE_BOUNDARY)
707 remaining = p->p_numthreads - p->p_boundary_count;
708 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
709 remaining = p->p_numthreads - p->p_suspcount;
711 panic("calc_remaining: wrong mode %d", mode);
716 remain_for_mode(int mode)
719 return (mode == SINGLE_ALLPROC ? 0 : 1);
723 weed_inhib(int mode, struct thread *td2, struct proc *p)
727 PROC_LOCK_ASSERT(p, MA_OWNED);
728 PROC_SLOCK_ASSERT(p, MA_OWNED);
729 THREAD_LOCK_ASSERT(td2, MA_OWNED);
734 * Since the thread lock is dropped by the scheduler we have
735 * to retry to check for races.
740 if (TD_IS_SUSPENDED(td2)) {
741 wakeup_swapper |= thread_unsuspend_one(td2, p, true);
745 if (TD_CAN_ABORT(td2)) {
746 wakeup_swapper |= sleepq_abort(td2, EINTR);
747 return (wakeup_swapper);
750 case SINGLE_BOUNDARY:
752 if (TD_IS_SUSPENDED(td2) &&
753 (td2->td_flags & TDF_BOUNDARY) == 0) {
754 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
758 if (TD_CAN_ABORT(td2)) {
759 wakeup_swapper |= sleepq_abort(td2, ERESTART);
760 return (wakeup_swapper);
765 * ALLPROC suspend tries to avoid spurious EINTR for
766 * threads sleeping interruptable, by suspending the
767 * thread directly, similarly to sig_suspend_threads().
768 * Since such sleep is not performed at the user
769 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
770 * is used to avoid immediate un-suspend.
772 if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
773 TDF_ALLPROCSUSP)) == 0) {
774 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
778 if (TD_CAN_ABORT(td2)) {
779 if ((td2->td_flags & TDF_SBDRY) == 0) {
780 thread_suspend_one(td2);
781 td2->td_flags |= TDF_ALLPROCSUSP;
783 wakeup_swapper |= sleepq_abort(td2, ERESTART);
784 return (wakeup_swapper);
792 return (wakeup_swapper);
796 * Enforce single-threading.
798 * Returns 1 if the caller must abort (another thread is waiting to
799 * exit the process or similar). Process is locked!
800 * Returns 0 when you are successfully the only thread running.
801 * A process has successfully single threaded in the suspend mode when
802 * There are no threads in user mode. Threads in the kernel must be
803 * allowed to continue until they get to the user boundary. They may even
804 * copy out their return values and data before suspending. They may however be
805 * accelerated in reaching the user boundary as we will wake up
806 * any sleeping threads that are interruptable. (PCATCH).
809 thread_single(struct proc *p, int mode)
813 int remaining, wakeup_swapper;
816 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
817 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
818 ("invalid mode %d", mode));
820 * If allowing non-ALLPROC singlethreading for non-curproc
821 * callers, calc_remaining() and remain_for_mode() should be
822 * adjusted to also account for td->td_proc != p. For now
823 * this is not implemented because it is not used.
825 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
826 (mode != SINGLE_ALLPROC && td->td_proc == p),
827 ("mode %d proc %p curproc %p", mode, p, td->td_proc));
828 mtx_assert(&Giant, MA_NOTOWNED);
829 PROC_LOCK_ASSERT(p, MA_OWNED);
831 if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
834 /* Is someone already single threading? */
835 if (p->p_singlethread != NULL && p->p_singlethread != td)
838 if (mode == SINGLE_EXIT) {
839 p->p_flag |= P_SINGLE_EXIT;
840 p->p_flag &= ~P_SINGLE_BOUNDARY;
842 p->p_flag &= ~P_SINGLE_EXIT;
843 if (mode == SINGLE_BOUNDARY)
844 p->p_flag |= P_SINGLE_BOUNDARY;
846 p->p_flag &= ~P_SINGLE_BOUNDARY;
848 if (mode == SINGLE_ALLPROC)
849 p->p_flag |= P_TOTAL_STOP;
850 p->p_flag |= P_STOPPED_SINGLE;
852 p->p_singlethread = td;
853 remaining = calc_remaining(p, mode);
854 while (remaining != remain_for_mode(mode)) {
855 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
858 FOREACH_THREAD_IN_PROC(p, td2) {
862 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
863 if (TD_IS_INHIBITED(td2)) {
864 wakeup_swapper |= weed_inhib(mode, td2, p);
866 } else if (TD_IS_RUNNING(td2) && td != td2) {
875 remaining = calc_remaining(p, mode);
878 * Maybe we suspended some threads.. was it enough?
880 if (remaining == remain_for_mode(mode))
885 * Wake us up when everyone else has suspended.
886 * In the mean time we suspend as well.
888 thread_suspend_switch(td, p);
889 remaining = calc_remaining(p, mode);
891 if (mode == SINGLE_EXIT) {
893 * Convert the process to an unthreaded process. The
894 * SINGLE_EXIT is called by exit1() or execve(), in
895 * both cases other threads must be retired.
897 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
898 p->p_singlethread = NULL;
899 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
902 * Wait for any remaining threads to exit cpu_throw().
904 while (p->p_exitthreads != 0) {
907 sched_relinquish(td);
911 } else if (mode == SINGLE_BOUNDARY) {
913 * Wait until all suspended threads are removed from
914 * the processors. The thread_suspend_check()
915 * increments p_boundary_count while it is still
916 * running, which makes it possible for the execve()
917 * to destroy vmspace while our other threads are
918 * still using the address space.
920 * We lock the thread, which is only allowed to
921 * succeed after context switch code finished using
924 FOREACH_THREAD_IN_PROC(p, td2) {
928 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
929 ("td %p not on boundary", td2));
930 KASSERT(TD_IS_SUSPENDED(td2),
931 ("td %p is not suspended", td2));
940 thread_suspend_check_needed(void)
947 PROC_LOCK_ASSERT(p, MA_OWNED);
948 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
949 (td->td_dbgflags & TDB_SUSPEND) != 0));
953 * Called in from locations that can safely check to see
954 * whether we have to suspend or at least throttle for a
955 * single-thread event (e.g. fork).
957 * Such locations include userret().
958 * If the "return_instead" argument is non zero, the thread must be able to
959 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
961 * The 'return_instead' argument tells the function if it may do a
962 * thread_exit() or suspend, or whether the caller must abort and back
965 * If the thread that set the single_threading request has set the
966 * P_SINGLE_EXIT bit in the process flags then this call will never return
967 * if 'return_instead' is false, but will exit.
969 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
970 *---------------+--------------------+---------------------
971 * 0 | returns 0 | returns 0 or 1
972 * | when ST ends | immediately
973 *---------------+--------------------+---------------------
974 * 1 | thread exits | returns 1
976 * 0 = thread_exit() or suspension ok,
977 * other = return error instead of stopping the thread.
979 * While a full suspension is under effect, even a single threading
980 * thread would be suspended if it made this call (but it shouldn't).
981 * This call should only be made from places where
982 * thread_exit() would be safe as that may be the outcome unless
983 * return_instead is set.
986 thread_suspend_check(int return_instead)
994 mtx_assert(&Giant, MA_NOTOWNED);
995 PROC_LOCK_ASSERT(p, MA_OWNED);
996 while (thread_suspend_check_needed()) {
997 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
998 KASSERT(p->p_singlethread != NULL,
999 ("singlethread not set"));
1001 * The only suspension in action is a
1002 * single-threading. Single threader need not stop.
1003 * It is safe to access p->p_singlethread unlocked
1004 * because it can only be set to our address by us.
1006 if (p->p_singlethread == td)
1007 return (0); /* Exempt from stopping. */
1009 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1012 /* Should we goto user boundary if we didn't come from there? */
1013 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1014 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1018 * Ignore suspend requests if they are deferred.
1020 if ((td->td_flags & TDF_SBDRY) != 0) {
1021 KASSERT(return_instead,
1022 ("TDF_SBDRY set for unsafe thread_suspend_check"));
1023 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1024 (TDF_SEINTR | TDF_SERESTART),
1025 ("both TDF_SEINTR and TDF_SERESTART"));
1026 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1030 * If the process is waiting for us to exit,
1031 * this thread should just suicide.
1032 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1034 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1038 * Allow Linux emulation layer to do some work
1039 * before thread suicide.
1041 if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1042 (p->p_sysent->sv_thread_detach)(td);
1043 umtx_thread_exit(td);
1045 panic("stopped thread did not exit");
1050 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1051 if (p->p_numthreads == p->p_suspcount + 1) {
1052 thread_lock(p->p_singlethread);
1053 wakeup_swapper = thread_unsuspend_one(
1054 p->p_singlethread, p, false);
1062 * When a thread suspends, it just
1063 * gets taken off all queues.
1065 thread_suspend_one(td);
1066 if (return_instead == 0) {
1067 p->p_boundary_count++;
1068 td->td_flags |= TDF_BOUNDARY;
1071 mi_switch(SW_INVOL | SWT_SUSPEND);
1078 * Check for possible stops and suspensions while executing a
1079 * casueword or similar transiently failing operation.
1081 * The sleep argument controls whether the function can handle a stop
1082 * request itself or it should return ERESTART and the request is
1083 * proceed at the kernel/user boundary in ast.
1085 * Typically, when retrying due to casueword(9) failure (rv == 1), we
1086 * should handle the stop requests there, with exception of cases when
1087 * the thread owns a kernel resource, for instance busied the umtx
1088 * key, or when functions return immediately if thread_check_susp()
1089 * returned non-zero. On the other hand, retrying the whole lock
1090 * operation, we better not stop there but delegate the handling to
1093 * If the request is for thread termination P_SINGLE_EXIT, we cannot
1094 * handle it at all, and simply return EINTR.
1097 thread_check_susp(struct thread *td, bool sleep)
1103 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
1104 * eventually break the lockstep loop.
1106 if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
1111 if (p->p_flag & P_SINGLE_EXIT)
1113 else if (P_SHOULDSTOP(p) ||
1114 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1115 error = sleep ? thread_suspend_check(0) : ERESTART;
1121 thread_suspend_switch(struct thread *td, struct proc *p)
1124 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1125 PROC_LOCK_ASSERT(p, MA_OWNED);
1126 PROC_SLOCK_ASSERT(p, MA_OWNED);
1128 * We implement thread_suspend_one in stages here to avoid
1129 * dropping the proc lock while the thread lock is owned.
1131 if (p == td->td_proc) {
1137 td->td_flags &= ~TDF_NEEDSUSPCHK;
1138 TD_SET_SUSPENDED(td);
1142 mi_switch(SW_VOL | SWT_SUSPEND);
1149 thread_suspend_one(struct thread *td)
1154 PROC_SLOCK_ASSERT(p, MA_OWNED);
1155 THREAD_LOCK_ASSERT(td, MA_OWNED);
1156 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1158 td->td_flags &= ~TDF_NEEDSUSPCHK;
1159 TD_SET_SUSPENDED(td);
1164 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1167 THREAD_LOCK_ASSERT(td, MA_OWNED);
1168 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1169 TD_CLR_SUSPENDED(td);
1170 td->td_flags &= ~TDF_ALLPROCSUSP;
1171 if (td->td_proc == p) {
1172 PROC_SLOCK_ASSERT(p, MA_OWNED);
1174 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1175 td->td_flags &= ~TDF_BOUNDARY;
1176 p->p_boundary_count--;
1179 return (setrunnable(td, 0));
1183 * Allow all threads blocked by single threading to continue running.
1186 thread_unsuspend(struct proc *p)
1191 PROC_LOCK_ASSERT(p, MA_OWNED);
1192 PROC_SLOCK_ASSERT(p, MA_OWNED);
1194 if (!P_SHOULDSTOP(p)) {
1195 FOREACH_THREAD_IN_PROC(p, td) {
1197 if (TD_IS_SUSPENDED(td)) {
1198 wakeup_swapper |= thread_unsuspend_one(td, p,
1203 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1204 p->p_numthreads == p->p_suspcount) {
1206 * Stopping everything also did the job for the single
1207 * threading request. Now we've downgraded to single-threaded,
1210 if (p->p_singlethread->td_proc == p) {
1211 thread_lock(p->p_singlethread);
1212 wakeup_swapper = thread_unsuspend_one(
1213 p->p_singlethread, p, false);
1221 * End the single threading mode..
1224 thread_single_end(struct proc *p, int mode)
1229 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1230 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1231 ("invalid mode %d", mode));
1232 PROC_LOCK_ASSERT(p, MA_OWNED);
1233 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1234 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1235 ("mode %d does not match P_TOTAL_STOP", mode));
1236 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1237 ("thread_single_end from other thread %p %p",
1238 curthread, p->p_singlethread));
1239 KASSERT(mode != SINGLE_BOUNDARY ||
1240 (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1241 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1242 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1245 p->p_singlethread = NULL;
1248 * If there are other threads they may now run,
1249 * unless of course there is a blanket 'stop order'
1250 * on the process. The single threader must be allowed
1251 * to continue however as this is a bad place to stop.
1253 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1254 FOREACH_THREAD_IN_PROC(p, td) {
1256 if (TD_IS_SUSPENDED(td)) {
1257 wakeup_swapper |= thread_unsuspend_one(td, p,
1258 mode == SINGLE_BOUNDARY);
1263 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1264 ("inconsistent boundary count %d", p->p_boundary_count));
1271 thread_find(struct proc *p, lwpid_t tid)
1275 PROC_LOCK_ASSERT(p, MA_OWNED);
1276 FOREACH_THREAD_IN_PROC(p, td) {
1277 if (td->td_tid == tid)
1283 /* Locate a thread by number; return with proc lock held. */
1285 tdfind(lwpid_t tid, pid_t pid)
1287 #define RUN_THRESH 16
1291 rw_rlock(&tidhash_lock);
1292 LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1293 if (td->td_tid == tid) {
1294 if (pid != -1 && td->td_proc->p_pid != pid) {
1298 PROC_LOCK(td->td_proc);
1299 if (td->td_proc->p_state == PRS_NEW) {
1300 PROC_UNLOCK(td->td_proc);
1304 if (run > RUN_THRESH) {
1305 if (rw_try_upgrade(&tidhash_lock)) {
1306 LIST_REMOVE(td, td_hash);
1307 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1309 rw_wunlock(&tidhash_lock);
1317 rw_runlock(&tidhash_lock);
1322 tidhash_add(struct thread *td)
1324 rw_wlock(&tidhash_lock);
1325 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1326 rw_wunlock(&tidhash_lock);
1330 tidhash_remove(struct thread *td)
1332 rw_wlock(&tidhash_lock);
1333 LIST_REMOVE(td, td_hash);
1334 rw_wunlock(&tidhash_lock);