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>
42 #include <sys/mutex.h>
44 #include <sys/bitstring.h>
45 #include <sys/epoch.h>
46 #include <sys/rangelock.h>
47 #include <sys/resourcevar.h>
50 #include <sys/sched.h>
51 #include <sys/sleepqueue.h>
52 #include <sys/selinfo.h>
53 #include <sys/syscallsubr.h>
54 #include <sys/dtrace_bsd.h>
55 #include <sys/sysent.h>
56 #include <sys/turnstile.h>
57 #include <sys/taskqueue.h>
59 #include <sys/rwlock.h>
60 #include <sys/umtxvar.h>
61 #include <sys/vmmeter.h>
62 #include <sys/cpuset.h>
64 #include <sys/pmckern.h>
68 #include <security/audit/audit.h>
72 #include <vm/vm_extern.h>
74 #include <vm/vm_phys.h>
75 #include <sys/eventhandler.h>
78 * Asserts below verify the stability of struct thread and struct proc
79 * layout, as exposed by KBI to modules. On head, the KBI is allowed
80 * to drift, change to the structures must be accompanied by the
83 * On the stable branches after KBI freeze, conditions must not be
84 * violated. Typically new fields are moved to the end of the
88 _Static_assert(offsetof(struct thread, td_flags) == 0x108,
89 "struct thread KBI td_flags");
90 _Static_assert(offsetof(struct thread, td_pflags) == 0x114,
91 "struct thread KBI td_pflags");
92 _Static_assert(offsetof(struct thread, td_frame) == 0x4b0,
93 "struct thread KBI td_frame");
94 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6c0,
95 "struct thread KBI td_emuldata");
96 _Static_assert(offsetof(struct proc, p_flag) == 0xb8,
97 "struct proc KBI p_flag");
98 _Static_assert(offsetof(struct proc, p_pid) == 0xc4,
99 "struct proc KBI p_pid");
100 _Static_assert(offsetof(struct proc, p_filemon) == 0x3c8,
101 "struct proc KBI p_filemon");
102 _Static_assert(offsetof(struct proc, p_comm) == 0x3e0,
103 "struct proc KBI p_comm");
104 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4c8,
105 "struct proc KBI p_emuldata");
108 _Static_assert(offsetof(struct thread, td_flags) == 0x9c,
109 "struct thread KBI td_flags");
110 _Static_assert(offsetof(struct thread, td_pflags) == 0xa8,
111 "struct thread KBI td_pflags");
112 _Static_assert(offsetof(struct thread, td_frame) == 0x30c,
113 "struct thread KBI td_frame");
114 _Static_assert(offsetof(struct thread, td_emuldata) == 0x350,
115 "struct thread KBI td_emuldata");
116 _Static_assert(offsetof(struct proc, p_flag) == 0x6c,
117 "struct proc KBI p_flag");
118 _Static_assert(offsetof(struct proc, p_pid) == 0x78,
119 "struct proc KBI p_pid");
120 _Static_assert(offsetof(struct proc, p_filemon) == 0x270,
121 "struct proc KBI p_filemon");
122 _Static_assert(offsetof(struct proc, p_comm) == 0x284,
123 "struct proc KBI p_comm");
124 _Static_assert(offsetof(struct proc, p_emuldata) == 0x310,
125 "struct proc KBI p_emuldata");
128 SDT_PROVIDER_DECLARE(proc);
129 SDT_PROBE_DEFINE(proc, , , lwp__exit);
132 * thread related storage.
134 static uma_zone_t thread_zone;
136 struct thread_domain_data {
137 struct thread *tdd_zombies;
139 } __aligned(CACHE_LINE_SIZE);
141 static struct thread_domain_data thread_domain_data[MAXMEMDOM];
143 static struct task thread_reap_task;
144 static struct callout thread_reap_callout;
146 static void thread_zombie(struct thread *);
147 static void thread_reap(void);
148 static void thread_reap_all(void);
149 static void thread_reap_task_cb(void *, int);
150 static void thread_reap_callout_cb(void *);
151 static int thread_unsuspend_one(struct thread *td, struct proc *p,
153 static void thread_free_batched(struct thread *td);
155 static __exclusive_cache_line struct mtx tid_lock;
156 static bitstr_t *tid_bitmap;
158 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
160 static int maxthread;
161 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
162 &maxthread, 0, "Maximum number of threads");
164 static __exclusive_cache_line int nthreads;
166 static LIST_HEAD(tidhashhead, thread) *tidhashtbl;
167 static u_long tidhash;
168 static u_long tidhashlock;
169 static struct rwlock *tidhashtbl_lock;
170 #define TIDHASH(tid) (&tidhashtbl[(tid) & tidhash])
171 #define TIDHASHLOCK(tid) (&tidhashtbl_lock[(tid) & tidhashlock])
173 EVENTHANDLER_LIST_DEFINE(thread_ctor);
174 EVENTHANDLER_LIST_DEFINE(thread_dtor);
175 EVENTHANDLER_LIST_DEFINE(thread_init);
176 EVENTHANDLER_LIST_DEFINE(thread_fini);
179 thread_count_inc_try(void)
183 nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1;
184 if (nthreads_new >= maxthread - 100) {
185 if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
186 nthreads_new >= maxthread) {
187 atomic_subtract_int(&nthreads, 1);
195 thread_count_inc(void)
197 static struct timeval lastfail;
201 if (thread_count_inc_try()) {
206 if (thread_count_inc_try()) {
210 if (ppsratecheck(&lastfail, &curfail, 1)) {
211 printf("maxthread limit exceeded by uid %u "
212 "(pid %d); consider increasing kern.maxthread\n",
213 curthread->td_ucred->cr_ruid, curproc->p_pid);
219 thread_count_sub(int n)
222 atomic_subtract_int(&nthreads, n);
226 thread_count_dec(void)
235 static lwpid_t trytid;
240 * It is an invariant that the bitmap is big enough to hold maxthread
241 * IDs. If we got to this point there has to be at least one free.
243 if (trytid >= maxthread)
245 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
247 KASSERT(trytid != 0, ("unexpectedly ran out of IDs"));
249 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
250 KASSERT(tid != -1, ("unexpectedly ran out of IDs"));
252 bit_set(tid_bitmap, tid);
254 mtx_unlock(&tid_lock);
255 return (tid + NO_PID);
259 tid_free_locked(lwpid_t rtid)
263 mtx_assert(&tid_lock, MA_OWNED);
264 KASSERT(rtid >= NO_PID,
265 ("%s: invalid tid %d\n", __func__, rtid));
267 KASSERT(bit_test(tid_bitmap, tid) != 0,
268 ("thread ID %d not allocated\n", rtid));
269 bit_clear(tid_bitmap, tid);
273 tid_free(lwpid_t rtid)
277 tid_free_locked(rtid);
278 mtx_unlock(&tid_lock);
282 tid_free_batch(lwpid_t *batch, int n)
287 for (i = 0; i < n; i++) {
288 tid_free_locked(batch[i]);
290 mtx_unlock(&tid_lock);
294 * Batching for thread reapping.
302 tidbatch_prep(struct tidbatch *tb)
309 tidbatch_add(struct tidbatch *tb, struct thread *td)
312 KASSERT(tb->n < nitems(tb->tab),
313 ("%s: count too high %d", __func__, tb->n));
314 tb->tab[tb->n] = td->td_tid;
319 tidbatch_process(struct tidbatch *tb)
322 KASSERT(tb->n <= nitems(tb->tab),
323 ("%s: count too high %d", __func__, tb->n));
324 if (tb->n == nitems(tb->tab)) {
325 tid_free_batch(tb->tab, tb->n);
331 tidbatch_final(struct tidbatch *tb)
334 KASSERT(tb->n <= nitems(tb->tab),
335 ("%s: count too high %d", __func__, tb->n));
337 tid_free_batch(tb->tab, tb->n);
342 * Prepare a thread for use.
345 thread_ctor(void *mem, int size, void *arg, int flags)
349 td = (struct thread *)mem;
350 TD_SET_STATE(td, TDS_INACTIVE);
351 td->td_lastcpu = td->td_oncpu = NOCPU;
354 * Note that td_critnest begins life as 1 because the thread is not
355 * running and is thereby implicitly waiting to be on the receiving
356 * end of a context switch.
359 td->td_lend_user_pri = PRI_MAX;
361 audit_thread_alloc(td);
364 kdtrace_thread_ctor(td);
366 umtx_thread_alloc(td);
367 MPASS(td->td_sel == NULL);
372 * Reclaim a thread after use.
375 thread_dtor(void *mem, int size, void *arg)
379 td = (struct thread *)mem;
382 /* Verify that this thread is in a safe state to free. */
383 switch (TD_GET_STATE(td)) {
389 * We must never unlink a thread that is in one of
390 * these states, because it is currently active.
392 panic("bad state for thread unlinking");
397 panic("bad thread state");
402 audit_thread_free(td);
405 kdtrace_thread_dtor(td);
407 /* Free all OSD associated to this thread. */
414 * Initialize type-stable parts of a thread (when newly created).
417 thread_init(void *mem, int size, int flags)
421 td = (struct thread *)mem;
423 td->td_allocdomain = vm_phys_domain(vtophys(td));
424 td->td_sleepqueue = sleepq_alloc();
425 td->td_turnstile = turnstile_alloc();
427 EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
428 umtx_thread_init(td);
435 * Tear down type-stable parts of a thread (just before being discarded).
438 thread_fini(void *mem, int size)
442 td = (struct thread *)mem;
443 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
444 rlqentry_free(td->td_rlqe);
445 turnstile_free(td->td_turnstile);
446 sleepq_free(td->td_sleepqueue);
447 umtx_thread_fini(td);
448 MPASS(td->td_sel == NULL);
452 * For a newly created process,
453 * link up all the structures and its initial threads etc.
455 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc.
456 * proc_dtor() (should go away)
460 proc_linkup0(struct proc *p, struct thread *td)
462 TAILQ_INIT(&p->p_threads); /* all threads in proc */
467 proc_linkup(struct proc *p, struct thread *td)
470 sigqueue_init(&p->p_sigqueue, p);
471 p->p_ksi = ksiginfo_alloc(M_WAITOK);
472 if (p->p_ksi != NULL) {
473 /* XXX p_ksi may be null if ksiginfo zone is not ready */
474 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
476 LIST_INIT(&p->p_mqnotifier);
482 ast_suspend(struct thread *td, int tda __unused)
488 * We need to check to see if we have to exit or wait due to a
489 * single threading requirement or some other STOP condition.
492 thread_suspend_check(0);
496 extern int max_threads_per_proc;
499 * Initialize global thread allocation resources.
509 * Place an upper limit on threads which can be allocated.
511 * Note that other factors may make the de facto limit much lower.
513 * Platform limits are somewhat arbitrary but deemed "more than good
514 * enough" for the foreseable future.
516 if (maxthread == 0) {
518 maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
520 maxthread = MIN(maxproc * max_threads_per_proc, 100000);
524 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
525 tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
531 if (tid0 != THREAD0_TID)
532 panic("tid0 %d != %d\n", tid0, THREAD0_TID);
534 flags = UMA_ZONE_NOFREE;
537 * Force thread structures to be allocated from the direct map.
538 * Otherwise, superpage promotions and demotions may temporarily
539 * invalidate thread structure mappings. For most dynamically allocated
540 * structures this is not a problem, but translation faults cannot be
541 * handled without accessing curthread.
543 flags |= UMA_ZONE_CONTIG;
545 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
546 thread_ctor, thread_dtor, thread_init, thread_fini,
548 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
549 tidhashlock = (tidhash + 1) / 64;
552 tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
553 M_TIDHASH, M_WAITOK | M_ZERO);
554 for (i = 0; i < tidhashlock + 1; i++)
555 rw_init(&tidhashtbl_lock[i], "tidhash");
557 TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL);
558 callout_init(&thread_reap_callout, 1);
559 callout_reset(&thread_reap_callout, 5 * hz,
560 thread_reap_callout_cb, NULL);
561 ast_register(TDA_SUSPEND, ASTR_ASTF_REQUIRED, 0, ast_suspend);
565 * Place an unused thread on the zombie list.
568 thread_zombie(struct thread *td)
570 struct thread_domain_data *tdd;
573 tdd = &thread_domain_data[td->td_allocdomain];
574 ztd = atomic_load_ptr(&tdd->tdd_zombies);
577 if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
578 (uintptr_t *)&ztd, (uintptr_t)td))
585 * Release a thread that has exited after cpu_throw().
588 thread_stash(struct thread *td)
590 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
595 * Reap zombies from passed domain.
598 thread_reap_domain(struct thread_domain_data *tdd)
600 struct thread *itd, *ntd;
601 struct tidbatch tidbatch;
602 struct credbatch credbatch;
608 * Reading upfront is pessimal if followed by concurrent atomic_swap,
609 * but most of the time the list is empty.
611 if (tdd->tdd_zombies == NULL)
614 itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
620 * Multiple CPUs can get here, the race is fine as ticks is only
623 tdd->tdd_reapticks = ticks;
625 tidbatch_prep(&tidbatch);
626 credbatch_prep(&credbatch);
631 while (itd != NULL) {
632 ntd = itd->td_zombie;
633 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd);
634 tidbatch_add(&tidbatch, itd);
635 credbatch_add(&credbatch, itd);
636 MPASS(itd->td_limit != NULL);
637 if (lim != itd->td_limit) {
639 lim_freen(lim, limcount);
645 thread_free_batched(itd);
646 tidbatch_process(&tidbatch);
647 credbatch_process(&credbatch);
650 thread_count_sub(tdcount);
656 tidbatch_final(&tidbatch);
657 credbatch_final(&credbatch);
659 thread_count_sub(tdcount);
661 MPASS(limcount != 0);
662 lim_freen(lim, limcount);
666 * Reap zombies from all domains.
669 thread_reap_all(void)
671 struct thread_domain_data *tdd;
674 domain = PCPU_GET(domain);
675 for (i = 0; i < vm_ndomains; i++) {
676 tdd = &thread_domain_data[(i + domain) % vm_ndomains];
677 thread_reap_domain(tdd);
682 * Reap zombies from local domain.
687 struct thread_domain_data *tdd;
690 domain = PCPU_GET(domain);
691 tdd = &thread_domain_data[domain];
693 thread_reap_domain(tdd);
697 thread_reap_task_cb(void *arg __unused, int pending __unused)
704 thread_reap_callout_cb(void *arg __unused)
706 struct thread_domain_data *tdd;
707 int i, cticks, lticks;
711 cticks = atomic_load_int(&ticks);
712 for (i = 0; i < vm_ndomains; i++) {
713 tdd = &thread_domain_data[i];
714 lticks = tdd->tdd_reapticks;
715 if (tdd->tdd_zombies != NULL &&
716 (u_int)(cticks - lticks) > 5 * hz) {
723 taskqueue_enqueue(taskqueue_thread, &thread_reap_task);
724 callout_reset(&thread_reap_callout, 5 * hz,
725 thread_reap_callout_cb, NULL);
729 * Calling this function guarantees that any thread that exited before
730 * the call is reaped when the function returns. By 'exited' we mean
731 * a thread removed from the process linkage with thread_unlink().
732 * Practically this means that caller must lock/unlock corresponding
733 * process lock before the call, to synchronize with thread_exit().
736 thread_reap_barrier(void)
741 * First do context switches to each CPU to ensure that all
742 * PCPU pc_deadthreads are moved to zombie list.
744 quiesce_all_cpus("", PDROP);
747 * Second, fire the task in the same thread as normal
748 * thread_reap() is done, to serialize reaping.
750 t = malloc(sizeof(*t), M_TEMP, M_WAITOK);
751 TASK_INIT(t, 0, thread_reap_task_cb, t);
752 taskqueue_enqueue(taskqueue_thread, t);
753 taskqueue_drain(taskqueue_thread, t);
761 thread_alloc(int pages)
766 if (!thread_count_inc()) {
771 td = uma_zalloc(thread_zone, M_WAITOK);
772 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
773 if (!vm_thread_new(td, pages)) {
774 uma_zfree(thread_zone, td);
780 bzero(&td->td_sa.args, sizeof(td->td_sa.args));
781 kmsan_thread_alloc(td);
782 cpu_thread_alloc(td);
783 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
788 thread_alloc_stack(struct thread *td, int pages)
791 KASSERT(td->td_kstack == 0,
792 ("thread_alloc_stack called on a thread with kstack"));
793 if (!vm_thread_new(td, pages))
795 cpu_thread_alloc(td);
800 * Deallocate a thread.
803 thread_free_batched(struct thread *td)
806 lock_profile_thread_exit(td);
808 cpuset_rel(td->td_cpuset);
809 td->td_cpuset = NULL;
811 if (td->td_kstack != 0)
812 vm_thread_dispose(td);
813 callout_drain(&td->td_slpcallout);
815 * Freeing handled by the caller.
818 kmsan_thread_free(td);
819 uma_zfree(thread_zone, td);
823 thread_free(struct thread *td)
827 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
829 thread_free_batched(td);
835 thread_cow_get_proc(struct thread *newtd, struct proc *p)
838 PROC_LOCK_ASSERT(p, MA_OWNED);
839 newtd->td_realucred = crcowget(p->p_ucred);
840 newtd->td_ucred = newtd->td_realucred;
841 newtd->td_limit = lim_hold(p->p_limit);
842 newtd->td_cowgen = p->p_cowgen;
846 thread_cow_get(struct thread *newtd, struct thread *td)
849 MPASS(td->td_realucred == td->td_ucred);
850 newtd->td_realucred = crcowget(td->td_realucred);
851 newtd->td_ucred = newtd->td_realucred;
852 newtd->td_limit = lim_hold(td->td_limit);
853 newtd->td_cowgen = td->td_cowgen;
857 thread_cow_free(struct thread *td)
860 if (td->td_realucred != NULL)
862 if (td->td_limit != NULL)
863 lim_free(td->td_limit);
867 thread_cow_update(struct thread *td)
870 struct ucred *oldcred;
871 struct plimit *oldlimit;
875 oldcred = crcowsync();
876 oldlimit = lim_cowsync();
877 td->td_cowgen = p->p_cowgen;
881 if (oldlimit != NULL)
886 thread_cow_synced(struct thread *td)
891 PROC_LOCK_ASSERT(p, MA_OWNED);
892 MPASS(td->td_cowgen != p->p_cowgen);
893 MPASS(td->td_ucred == p->p_ucred);
894 MPASS(td->td_limit == p->p_limit);
895 td->td_cowgen = p->p_cowgen;
899 * Discard the current thread and exit from its context.
900 * Always called with scheduler locked.
902 * Because we can't free a thread while we're operating under its context,
903 * push the current thread into our CPU's deadthread holder. This means
904 * we needn't worry about someone else grabbing our context before we
910 uint64_t runtime, new_switchtime;
919 PROC_SLOCK_ASSERT(p, MA_OWNED);
920 mtx_assert(&Giant, MA_NOTOWNED);
922 PROC_LOCK_ASSERT(p, MA_OWNED);
923 KASSERT(p != NULL, ("thread exiting without a process"));
924 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
925 (long)p->p_pid, td->td_name);
926 SDT_PROBE0(proc, , , lwp__exit);
927 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
928 MPASS(td->td_realucred == td->td_ucred);
931 * drop FPU & debug register state storage, or any other
932 * architecture specific resources that
933 * would not be on a new untouched process.
938 * The last thread is left attached to the process
939 * So that the whole bundle gets recycled. Skip
940 * all this stuff if we never had threads.
941 * EXIT clears all sign of other threads when
942 * it goes to single threading, so the last thread always
943 * takes the short path.
945 if (p->p_flag & P_HADTHREADS) {
946 if (p->p_numthreads > 1) {
947 atomic_add_int(&td->td_proc->p_exitthreads, 1);
949 td2 = FIRST_THREAD_IN_PROC(p);
950 sched_exit_thread(td2, td);
953 * The test below is NOT true if we are the
954 * sole exiting thread. P_STOPPED_SINGLE is unset
955 * in exit1() after it is the only survivor.
957 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
958 if (p->p_numthreads == p->p_suspcount) {
959 thread_lock(p->p_singlethread);
960 wakeup_swapper = thread_unsuspend_one(
961 p->p_singlethread, p, false);
967 PCPU_SET(deadthread, td);
970 * The last thread is exiting.. but not through exit()
972 panic ("thread_exit: Last thread exiting on its own");
977 * If this thread is part of a process that is being tracked by hwpmc(4),
978 * inform the module of the thread's impending exit.
980 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
981 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
982 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
983 } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
984 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
991 /* Do the same timestamp bookkeeping that mi_switch() would do. */
992 new_switchtime = cpu_ticks();
993 runtime = new_switchtime - PCPU_GET(switchtime);
994 td->td_runtime += runtime;
995 td->td_incruntime += runtime;
996 PCPU_SET(switchtime, new_switchtime);
997 PCPU_SET(switchticks, ticks);
1000 /* Save our resource usage in our process. */
1001 td->td_ru.ru_nvcsw++;
1002 ruxagg_locked(p, td);
1003 rucollect(&p->p_ru, &td->td_ru);
1006 TD_SET_STATE(td, TDS_INACTIVE);
1008 witness_thread_exit(td);
1010 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
1012 panic("I'm a teapot!");
1017 * Do any thread specific cleanups that may be needed in wait()
1018 * called with Giant, proc and schedlock not held.
1021 thread_wait(struct proc *p)
1025 mtx_assert(&Giant, MA_NOTOWNED);
1026 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
1027 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
1028 td = FIRST_THREAD_IN_PROC(p);
1029 /* Lock the last thread so we spin until it exits cpu_throw(). */
1032 lock_profile_thread_exit(td);
1033 cpuset_rel(td->td_cpuset);
1034 td->td_cpuset = NULL;
1035 cpu_thread_clean(td);
1036 thread_cow_free(td);
1037 callout_drain(&td->td_slpcallout);
1038 thread_reap(); /* check for zombie threads etc. */
1042 * Link a thread to a process.
1043 * set up anything that needs to be initialized for it to
1044 * be used by the process.
1047 thread_link(struct thread *td, struct proc *p)
1051 * XXX This can't be enabled because it's called for proc0 before
1052 * its lock has been created.
1053 * PROC_LOCK_ASSERT(p, MA_OWNED);
1055 TD_SET_STATE(td, TDS_INACTIVE);
1057 td->td_flags = TDF_INMEM;
1059 LIST_INIT(&td->td_contested);
1060 LIST_INIT(&td->td_lprof[0]);
1061 LIST_INIT(&td->td_lprof[1]);
1063 SLIST_INIT(&td->td_epochs);
1065 sigqueue_init(&td->td_sigqueue, p);
1066 callout_init(&td->td_slpcallout, 1);
1067 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
1076 thread_unlink(struct thread *td)
1078 struct proc *p = td->td_proc;
1080 PROC_LOCK_ASSERT(p, MA_OWNED);
1082 MPASS(SLIST_EMPTY(&td->td_epochs));
1085 TAILQ_REMOVE(&p->p_threads, td, td_plist);
1087 /* could clear a few other things here */
1088 /* Must NOT clear links to proc! */
1092 calc_remaining(struct proc *p, int mode)
1096 PROC_LOCK_ASSERT(p, MA_OWNED);
1097 PROC_SLOCK_ASSERT(p, MA_OWNED);
1098 if (mode == SINGLE_EXIT)
1099 remaining = p->p_numthreads;
1100 else if (mode == SINGLE_BOUNDARY)
1101 remaining = p->p_numthreads - p->p_boundary_count;
1102 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
1103 remaining = p->p_numthreads - p->p_suspcount;
1105 panic("calc_remaining: wrong mode %d", mode);
1110 remain_for_mode(int mode)
1113 return (mode == SINGLE_ALLPROC ? 0 : 1);
1117 weed_inhib(int mode, struct thread *td2, struct proc *p)
1121 PROC_LOCK_ASSERT(p, MA_OWNED);
1122 PROC_SLOCK_ASSERT(p, MA_OWNED);
1123 THREAD_LOCK_ASSERT(td2, MA_OWNED);
1128 * Since the thread lock is dropped by the scheduler we have
1129 * to retry to check for races.
1134 if (TD_IS_SUSPENDED(td2)) {
1135 wakeup_swapper |= thread_unsuspend_one(td2, p, true);
1139 if (TD_CAN_ABORT(td2)) {
1140 wakeup_swapper |= sleepq_abort(td2, EINTR);
1141 return (wakeup_swapper);
1144 case SINGLE_BOUNDARY:
1145 case SINGLE_NO_EXIT:
1146 if (TD_IS_SUSPENDED(td2) &&
1147 (td2->td_flags & TDF_BOUNDARY) == 0) {
1148 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1152 if (TD_CAN_ABORT(td2)) {
1153 wakeup_swapper |= sleepq_abort(td2, ERESTART);
1154 return (wakeup_swapper);
1157 case SINGLE_ALLPROC:
1159 * ALLPROC suspend tries to avoid spurious EINTR for
1160 * threads sleeping interruptable, by suspending the
1161 * thread directly, similarly to sig_suspend_threads().
1162 * Since such sleep is not neccessary performed at the user
1163 * boundary, TDF_ALLPROCSUSP is used to avoid immediate
1166 if (TD_IS_SUSPENDED(td2) && (td2->td_flags &
1167 TDF_ALLPROCSUSP) == 0) {
1168 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1172 if (TD_CAN_ABORT(td2)) {
1173 td2->td_flags |= TDF_ALLPROCSUSP;
1174 wakeup_swapper |= sleepq_abort(td2, ERESTART);
1175 return (wakeup_swapper);
1182 return (wakeup_swapper);
1186 * Enforce single-threading.
1188 * Returns 1 if the caller must abort (another thread is waiting to
1189 * exit the process or similar). Process is locked!
1190 * Returns 0 when you are successfully the only thread running.
1191 * A process has successfully single threaded in the suspend mode when
1192 * There are no threads in user mode. Threads in the kernel must be
1193 * allowed to continue until they get to the user boundary. They may even
1194 * copy out their return values and data before suspending. They may however be
1195 * accelerated in reaching the user boundary as we will wake up
1196 * any sleeping threads that are interruptable. (PCATCH).
1199 thread_single(struct proc *p, int mode)
1203 int remaining, wakeup_swapper;
1206 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1207 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1208 ("invalid mode %d", mode));
1210 * If allowing non-ALLPROC singlethreading for non-curproc
1211 * callers, calc_remaining() and remain_for_mode() should be
1212 * adjusted to also account for td->td_proc != p. For now
1213 * this is not implemented because it is not used.
1215 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
1216 (mode != SINGLE_ALLPROC && td->td_proc == p),
1217 ("mode %d proc %p curproc %p", mode, p, td->td_proc));
1218 mtx_assert(&Giant, MA_NOTOWNED);
1219 PROC_LOCK_ASSERT(p, MA_OWNED);
1222 * Is someone already single threading?
1223 * Or may be singlethreading is not needed at all.
1225 if (mode == SINGLE_ALLPROC) {
1226 while ((p->p_flag & P_STOPPED_SINGLE) != 0) {
1227 if ((p->p_flag2 & P2_WEXIT) != 0)
1229 msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0);
1231 } else if ((p->p_flag & P_HADTHREADS) == 0)
1233 if (p->p_singlethread != NULL && p->p_singlethread != td)
1236 if (mode == SINGLE_EXIT) {
1237 p->p_flag |= P_SINGLE_EXIT;
1238 p->p_flag &= ~P_SINGLE_BOUNDARY;
1240 p->p_flag &= ~P_SINGLE_EXIT;
1241 if (mode == SINGLE_BOUNDARY)
1242 p->p_flag |= P_SINGLE_BOUNDARY;
1244 p->p_flag &= ~P_SINGLE_BOUNDARY;
1246 if (mode == SINGLE_ALLPROC)
1247 p->p_flag |= P_TOTAL_STOP;
1248 p->p_flag |= P_STOPPED_SINGLE;
1250 p->p_singlethread = td;
1251 remaining = calc_remaining(p, mode);
1252 while (remaining != remain_for_mode(mode)) {
1253 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
1256 FOREACH_THREAD_IN_PROC(p, td2) {
1260 ast_sched_locked(td2, TDA_SUSPEND);
1261 if (TD_IS_INHIBITED(td2)) {
1262 wakeup_swapper |= weed_inhib(mode, td2, p);
1264 } else if (TD_IS_RUNNING(td2)) {
1265 forward_signal(td2);
1273 remaining = calc_remaining(p, mode);
1276 * Maybe we suspended some threads.. was it enough?
1278 if (remaining == remain_for_mode(mode))
1283 * Wake us up when everyone else has suspended.
1284 * In the mean time we suspend as well.
1286 thread_suspend_switch(td, p);
1287 remaining = calc_remaining(p, mode);
1289 if (mode == SINGLE_EXIT) {
1291 * Convert the process to an unthreaded process. The
1292 * SINGLE_EXIT is called by exit1() or execve(), in
1293 * both cases other threads must be retired.
1295 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1296 p->p_singlethread = NULL;
1297 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1300 * Wait for any remaining threads to exit cpu_throw().
1302 while (p->p_exitthreads != 0) {
1305 sched_relinquish(td);
1309 } else if (mode == SINGLE_BOUNDARY) {
1311 * Wait until all suspended threads are removed from
1312 * the processors. The thread_suspend_check()
1313 * increments p_boundary_count while it is still
1314 * running, which makes it possible for the execve()
1315 * to destroy vmspace while our other threads are
1316 * still using the address space.
1318 * We lock the thread, which is only allowed to
1319 * succeed after context switch code finished using
1320 * the address space.
1322 FOREACH_THREAD_IN_PROC(p, td2) {
1326 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1327 ("td %p not on boundary", td2));
1328 KASSERT(TD_IS_SUSPENDED(td2),
1329 ("td %p is not suspended", td2));
1338 thread_suspend_check_needed(void)
1345 PROC_LOCK_ASSERT(p, MA_OWNED);
1346 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1347 (td->td_dbgflags & TDB_SUSPEND) != 0));
1351 * Called in from locations that can safely check to see
1352 * whether we have to suspend or at least throttle for a
1353 * single-thread event (e.g. fork).
1355 * Such locations include userret().
1356 * If the "return_instead" argument is non zero, the thread must be able to
1357 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1359 * The 'return_instead' argument tells the function if it may do a
1360 * thread_exit() or suspend, or whether the caller must abort and back
1363 * If the thread that set the single_threading request has set the
1364 * P_SINGLE_EXIT bit in the process flags then this call will never return
1365 * if 'return_instead' is false, but will exit.
1367 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1368 *---------------+--------------------+---------------------
1369 * 0 | returns 0 | returns 0 or 1
1370 * | when ST ends | immediately
1371 *---------------+--------------------+---------------------
1372 * 1 | thread exits | returns 1
1374 * 0 = thread_exit() or suspension ok,
1375 * other = return error instead of stopping the thread.
1377 * While a full suspension is under effect, even a single threading
1378 * thread would be suspended if it made this call (but it shouldn't).
1379 * This call should only be made from places where
1380 * thread_exit() would be safe as that may be the outcome unless
1381 * return_instead is set.
1384 thread_suspend_check(int return_instead)
1392 mtx_assert(&Giant, MA_NOTOWNED);
1393 PROC_LOCK_ASSERT(p, MA_OWNED);
1394 while (thread_suspend_check_needed()) {
1395 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1396 KASSERT(p->p_singlethread != NULL,
1397 ("singlethread not set"));
1399 * The only suspension in action is a
1400 * single-threading. Single threader need not stop.
1401 * It is safe to access p->p_singlethread unlocked
1402 * because it can only be set to our address by us.
1404 if (p->p_singlethread == td)
1405 return (0); /* Exempt from stopping. */
1407 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1410 /* Should we goto user boundary if we didn't come from there? */
1411 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1412 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1416 * Ignore suspend requests if they are deferred.
1418 if ((td->td_flags & TDF_SBDRY) != 0) {
1419 KASSERT(return_instead,
1420 ("TDF_SBDRY set for unsafe thread_suspend_check"));
1421 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1422 (TDF_SEINTR | TDF_SERESTART),
1423 ("both TDF_SEINTR and TDF_SERESTART"));
1424 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1428 * If the process is waiting for us to exit,
1429 * this thread should just suicide.
1430 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1432 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1436 * Allow Linux emulation layer to do some work
1437 * before thread suicide.
1439 if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1440 (p->p_sysent->sv_thread_detach)(td);
1441 umtx_thread_exit(td);
1443 panic("stopped thread did not exit");
1448 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1449 if (p->p_numthreads == p->p_suspcount + 1) {
1450 thread_lock(p->p_singlethread);
1451 wakeup_swapper = thread_unsuspend_one(
1452 p->p_singlethread, p, false);
1460 * When a thread suspends, it just
1461 * gets taken off all queues.
1463 thread_suspend_one(td);
1464 if (return_instead == 0) {
1465 p->p_boundary_count++;
1466 td->td_flags |= TDF_BOUNDARY;
1469 mi_switch(SW_INVOL | SWT_SUSPEND);
1476 * Check for possible stops and suspensions while executing a
1477 * casueword or similar transiently failing operation.
1479 * The sleep argument controls whether the function can handle a stop
1480 * request itself or it should return ERESTART and the request is
1481 * proceed at the kernel/user boundary in ast.
1483 * Typically, when retrying due to casueword(9) failure (rv == 1), we
1484 * should handle the stop requests there, with exception of cases when
1485 * the thread owns a kernel resource, for instance busied the umtx
1486 * key, or when functions return immediately if thread_check_susp()
1487 * returned non-zero. On the other hand, retrying the whole lock
1488 * operation, we better not stop there but delegate the handling to
1491 * If the request is for thread termination P_SINGLE_EXIT, we cannot
1492 * handle it at all, and simply return EINTR.
1495 thread_check_susp(struct thread *td, bool sleep)
1501 * The check for TDA_SUSPEND is racy, but it is enough to
1502 * eventually break the lockstep loop.
1504 if (!td_ast_pending(td, TDA_SUSPEND))
1509 if (p->p_flag & P_SINGLE_EXIT)
1511 else if (P_SHOULDSTOP(p) ||
1512 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1513 error = sleep ? thread_suspend_check(0) : ERESTART;
1519 thread_suspend_switch(struct thread *td, struct proc *p)
1522 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1523 PROC_LOCK_ASSERT(p, MA_OWNED);
1524 PROC_SLOCK_ASSERT(p, MA_OWNED);
1526 * We implement thread_suspend_one in stages here to avoid
1527 * dropping the proc lock while the thread lock is owned.
1529 if (p == td->td_proc) {
1535 ast_unsched_locked(td, TDA_SUSPEND);
1536 TD_SET_SUSPENDED(td);
1540 mi_switch(SW_VOL | SWT_SUSPEND);
1547 thread_suspend_one(struct thread *td)
1552 PROC_SLOCK_ASSERT(p, MA_OWNED);
1553 THREAD_LOCK_ASSERT(td, MA_OWNED);
1554 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1556 ast_unsched_locked(td, TDA_SUSPEND);
1557 TD_SET_SUSPENDED(td);
1562 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1565 THREAD_LOCK_ASSERT(td, MA_OWNED);
1566 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1567 TD_CLR_SUSPENDED(td);
1568 td->td_flags &= ~TDF_ALLPROCSUSP;
1569 if (td->td_proc == p) {
1570 PROC_SLOCK_ASSERT(p, MA_OWNED);
1572 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1573 td->td_flags &= ~TDF_BOUNDARY;
1574 p->p_boundary_count--;
1577 return (setrunnable(td, 0));
1581 thread_run_flash(struct thread *td)
1586 PROC_LOCK_ASSERT(p, MA_OWNED);
1588 if (TD_ON_SLEEPQ(td))
1589 sleepq_remove_nested(td);
1593 THREAD_LOCK_ASSERT(td, MA_OWNED);
1594 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1596 TD_CLR_SUSPENDED(td);
1598 MPASS(p->p_suspcount > 0);
1601 if (setrunnable(td, 0))
1606 * Allow all threads blocked by single threading to continue running.
1609 thread_unsuspend(struct proc *p)
1614 PROC_LOCK_ASSERT(p, MA_OWNED);
1615 PROC_SLOCK_ASSERT(p, MA_OWNED);
1617 if (!P_SHOULDSTOP(p)) {
1618 FOREACH_THREAD_IN_PROC(p, td) {
1620 if (TD_IS_SUSPENDED(td))
1621 wakeup_swapper |= thread_unsuspend_one(td, p,
1626 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1627 p->p_numthreads == p->p_suspcount) {
1629 * Stopping everything also did the job for the single
1630 * threading request. Now we've downgraded to single-threaded,
1633 if (p->p_singlethread->td_proc == p) {
1634 thread_lock(p->p_singlethread);
1635 wakeup_swapper = thread_unsuspend_one(
1636 p->p_singlethread, p, false);
1644 * End the single threading mode..
1647 thread_single_end(struct proc *p, int mode)
1652 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1653 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1654 ("invalid mode %d", mode));
1655 PROC_LOCK_ASSERT(p, MA_OWNED);
1656 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1657 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1658 ("mode %d does not match P_TOTAL_STOP", mode));
1659 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1660 ("thread_single_end from other thread %p %p",
1661 curthread, p->p_singlethread));
1662 KASSERT(mode != SINGLE_BOUNDARY ||
1663 (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1664 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1665 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1668 p->p_singlethread = NULL;
1671 * If there are other threads they may now run,
1672 * unless of course there is a blanket 'stop order'
1673 * on the process. The single threader must be allowed
1674 * to continue however as this is a bad place to stop.
1676 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1677 FOREACH_THREAD_IN_PROC(p, td) {
1679 if (TD_IS_SUSPENDED(td)) {
1680 wakeup_swapper |= thread_unsuspend_one(td, p,
1686 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1687 ("inconsistent boundary count %d", p->p_boundary_count));
1695 * Locate a thread by number and return with proc lock held.
1697 * thread exit establishes proc -> tidhash lock ordering, but lookup
1698 * takes tidhash first and needs to return locked proc.
1700 * The problem is worked around by relying on type-safety of both
1701 * structures and doing the work in 2 steps:
1702 * - tidhash-locked lookup which saves both thread and proc pointers
1703 * - proc-locked verification that the found thread still matches
1706 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1708 #define RUN_THRESH 16
1715 rw_rlock(TIDHASHLOCK(tid));
1717 LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1718 if (td->td_tid != tid) {
1723 if (pid != -1 && p->p_pid != pid) {
1727 if (run > RUN_THRESH) {
1728 if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1729 LIST_REMOVE(td, td_hash);
1730 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1732 rw_wunlock(TIDHASHLOCK(tid));
1740 rw_runlock(TIDHASHLOCK(tid));
1749 tdfind(lwpid_t tid, pid_t pid)
1755 if (td->td_tid == tid) {
1756 if (pid != -1 && td->td_proc->p_pid != pid)
1758 PROC_LOCK(td->td_proc);
1763 if (!tdfind_hash(tid, pid, &p, &td))
1766 if (td->td_tid != tid) {
1770 if (td->td_proc != p) {
1774 if (p->p_state == PRS_NEW) {
1783 tidhash_add(struct thread *td)
1785 rw_wlock(TIDHASHLOCK(td->td_tid));
1786 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1787 rw_wunlock(TIDHASHLOCK(td->td_tid));
1791 tidhash_remove(struct thread *td)
1794 rw_wlock(TIDHASHLOCK(td->td_tid));
1795 LIST_REMOVE(td, td_hash);
1796 rw_wunlock(TIDHASHLOCK(td->td_tid));