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) == 0x110,
91 "struct thread KBI td_pflags");
92 _Static_assert(offsetof(struct thread, td_frame) == 0x4a8,
93 "struct thread KBI td_frame");
94 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6b0,
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) == 0xa4,
111 "struct thread KBI td_pflags");
112 _Static_assert(offsetof(struct thread, td_frame) == 0x308,
113 "struct thread KBI td_frame");
114 _Static_assert(offsetof(struct thread, td_emuldata) == 0x34c,
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. */
409 td_softdep_cleanup(td);
410 MPASS(td->td_su == NULL);
415 * Initialize type-stable parts of a thread (when newly created).
418 thread_init(void *mem, int size, int flags)
422 td = (struct thread *)mem;
424 td->td_allocdomain = vm_phys_domain(vtophys(td));
425 td->td_sleepqueue = sleepq_alloc();
426 td->td_turnstile = turnstile_alloc();
428 EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
429 umtx_thread_init(td);
436 * Tear down type-stable parts of a thread (just before being discarded).
439 thread_fini(void *mem, int size)
443 td = (struct thread *)mem;
444 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
445 rlqentry_free(td->td_rlqe);
446 turnstile_free(td->td_turnstile);
447 sleepq_free(td->td_sleepqueue);
448 umtx_thread_fini(td);
449 MPASS(td->td_sel == NULL);
453 * For a newly created process,
454 * link up all the structures and its initial threads etc.
456 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc.
457 * proc_dtor() (should go away)
461 proc_linkup0(struct proc *p, struct thread *td)
463 TAILQ_INIT(&p->p_threads); /* all threads in proc */
468 proc_linkup(struct proc *p, struct thread *td)
471 sigqueue_init(&p->p_sigqueue, p);
472 p->p_ksi = ksiginfo_alloc(1);
473 if (p->p_ksi != NULL) {
474 /* XXX p_ksi may be null if ksiginfo zone is not ready */
475 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
477 LIST_INIT(&p->p_mqnotifier);
482 extern int max_threads_per_proc;
485 * Initialize global thread allocation resources.
495 * Place an upper limit on threads which can be allocated.
497 * Note that other factors may make the de facto limit much lower.
499 * Platform limits are somewhat arbitrary but deemed "more than good
500 * enough" for the foreseable future.
502 if (maxthread == 0) {
504 maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
506 maxthread = MIN(maxproc * max_threads_per_proc, 100000);
510 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
511 tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
517 if (tid0 != THREAD0_TID)
518 panic("tid0 %d != %d\n", tid0, THREAD0_TID);
520 flags = UMA_ZONE_NOFREE;
523 * Force thread structures to be allocated from the direct map.
524 * Otherwise, superpage promotions and demotions may temporarily
525 * invalidate thread structure mappings. For most dynamically allocated
526 * structures this is not a problem, but translation faults cannot be
527 * handled without accessing curthread.
529 flags |= UMA_ZONE_CONTIG;
531 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
532 thread_ctor, thread_dtor, thread_init, thread_fini,
534 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
535 tidhashlock = (tidhash + 1) / 64;
538 tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
539 M_TIDHASH, M_WAITOK | M_ZERO);
540 for (i = 0; i < tidhashlock + 1; i++)
541 rw_init(&tidhashtbl_lock[i], "tidhash");
543 TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL);
544 callout_init(&thread_reap_callout, 1);
545 callout_reset(&thread_reap_callout, 5 * hz,
546 thread_reap_callout_cb, NULL);
550 * Place an unused thread on the zombie list.
553 thread_zombie(struct thread *td)
555 struct thread_domain_data *tdd;
558 tdd = &thread_domain_data[td->td_allocdomain];
559 ztd = atomic_load_ptr(&tdd->tdd_zombies);
562 if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
563 (uintptr_t *)&ztd, (uintptr_t)td))
570 * Release a thread that has exited after cpu_throw().
573 thread_stash(struct thread *td)
575 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
580 * Reap zombies from passed domain.
583 thread_reap_domain(struct thread_domain_data *tdd)
585 struct thread *itd, *ntd;
586 struct tidbatch tidbatch;
587 struct credbatch credbatch;
593 * Reading upfront is pessimal if followed by concurrent atomic_swap,
594 * but most of the time the list is empty.
596 if (tdd->tdd_zombies == NULL)
599 itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
605 * Multiple CPUs can get here, the race is fine as ticks is only
608 tdd->tdd_reapticks = ticks;
610 tidbatch_prep(&tidbatch);
611 credbatch_prep(&credbatch);
616 while (itd != NULL) {
617 ntd = itd->td_zombie;
618 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd);
619 tidbatch_add(&tidbatch, itd);
620 credbatch_add(&credbatch, itd);
621 MPASS(itd->td_limit != NULL);
622 if (lim != itd->td_limit) {
624 lim_freen(lim, limcount);
630 thread_free_batched(itd);
631 tidbatch_process(&tidbatch);
632 credbatch_process(&credbatch);
635 thread_count_sub(tdcount);
641 tidbatch_final(&tidbatch);
642 credbatch_final(&credbatch);
644 thread_count_sub(tdcount);
646 MPASS(limcount != 0);
647 lim_freen(lim, limcount);
651 * Reap zombies from all domains.
654 thread_reap_all(void)
656 struct thread_domain_data *tdd;
659 domain = PCPU_GET(domain);
660 for (i = 0; i < vm_ndomains; i++) {
661 tdd = &thread_domain_data[(i + domain) % vm_ndomains];
662 thread_reap_domain(tdd);
667 * Reap zombies from local domain.
672 struct thread_domain_data *tdd;
675 domain = PCPU_GET(domain);
676 tdd = &thread_domain_data[domain];
678 thread_reap_domain(tdd);
682 thread_reap_task_cb(void *arg __unused, int pending __unused)
689 thread_reap_callout_cb(void *arg __unused)
691 struct thread_domain_data *tdd;
692 int i, cticks, lticks;
696 cticks = atomic_load_int(&ticks);
697 for (i = 0; i < vm_ndomains; i++) {
698 tdd = &thread_domain_data[i];
699 lticks = tdd->tdd_reapticks;
700 if (tdd->tdd_zombies != NULL &&
701 (u_int)(cticks - lticks) > 5 * hz) {
708 taskqueue_enqueue(taskqueue_thread, &thread_reap_task);
709 callout_reset(&thread_reap_callout, 5 * hz,
710 thread_reap_callout_cb, NULL);
714 * Calling this function guarantees that any thread that exited before
715 * the call is reaped when the function returns. By 'exited' we mean
716 * a thread removed from the process linkage with thread_unlink().
717 * Practically this means that caller must lock/unlock corresponding
718 * process lock before the call, to synchronize with thread_exit().
721 thread_reap_barrier(void)
726 * First do context switches to each CPU to ensure that all
727 * PCPU pc_deadthreads are moved to zombie list.
729 quiesce_all_cpus("", PDROP);
732 * Second, fire the task in the same thread as normal
733 * thread_reap() is done, to serialize reaping.
735 t = malloc(sizeof(*t), M_TEMP, M_WAITOK);
736 TASK_INIT(t, 0, thread_reap_task_cb, t);
737 taskqueue_enqueue(taskqueue_thread, t);
738 taskqueue_drain(taskqueue_thread, t);
746 thread_alloc(int pages)
751 if (!thread_count_inc()) {
756 td = uma_zalloc(thread_zone, M_WAITOK);
757 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
758 if (!vm_thread_new(td, pages)) {
759 uma_zfree(thread_zone, td);
765 bzero(&td->td_sa.args, sizeof(td->td_sa.args));
766 kmsan_thread_alloc(td);
767 cpu_thread_alloc(td);
768 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
773 thread_alloc_stack(struct thread *td, int pages)
776 KASSERT(td->td_kstack == 0,
777 ("thread_alloc_stack called on a thread with kstack"));
778 if (!vm_thread_new(td, pages))
780 cpu_thread_alloc(td);
785 * Deallocate a thread.
788 thread_free_batched(struct thread *td)
791 lock_profile_thread_exit(td);
793 cpuset_rel(td->td_cpuset);
794 td->td_cpuset = NULL;
796 if (td->td_kstack != 0)
797 vm_thread_dispose(td);
798 callout_drain(&td->td_slpcallout);
800 * Freeing handled by the caller.
803 kmsan_thread_free(td);
804 uma_zfree(thread_zone, td);
808 thread_free(struct thread *td)
812 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
814 thread_free_batched(td);
820 thread_cow_get_proc(struct thread *newtd, struct proc *p)
823 PROC_LOCK_ASSERT(p, MA_OWNED);
824 newtd->td_realucred = crcowget(p->p_ucred);
825 newtd->td_ucred = newtd->td_realucred;
826 newtd->td_limit = lim_hold(p->p_limit);
827 newtd->td_cowgen = p->p_cowgen;
831 thread_cow_get(struct thread *newtd, struct thread *td)
834 MPASS(td->td_realucred == td->td_ucred);
835 newtd->td_realucred = crcowget(td->td_realucred);
836 newtd->td_ucred = newtd->td_realucred;
837 newtd->td_limit = lim_hold(td->td_limit);
838 newtd->td_cowgen = td->td_cowgen;
842 thread_cow_free(struct thread *td)
845 if (td->td_realucred != NULL)
847 if (td->td_limit != NULL)
848 lim_free(td->td_limit);
852 thread_cow_update(struct thread *td)
855 struct ucred *oldcred;
856 struct plimit *oldlimit;
860 oldcred = crcowsync();
861 oldlimit = lim_cowsync();
862 td->td_cowgen = p->p_cowgen;
866 if (oldlimit != NULL)
871 thread_cow_synced(struct thread *td)
876 PROC_LOCK_ASSERT(p, MA_OWNED);
877 MPASS(td->td_cowgen != p->p_cowgen);
878 MPASS(td->td_ucred == p->p_ucred);
879 MPASS(td->td_limit == p->p_limit);
880 td->td_cowgen = p->p_cowgen;
884 * Discard the current thread and exit from its context.
885 * Always called with scheduler locked.
887 * Because we can't free a thread while we're operating under its context,
888 * push the current thread into our CPU's deadthread holder. This means
889 * we needn't worry about someone else grabbing our context before we
895 uint64_t runtime, new_switchtime;
904 PROC_SLOCK_ASSERT(p, MA_OWNED);
905 mtx_assert(&Giant, MA_NOTOWNED);
907 PROC_LOCK_ASSERT(p, MA_OWNED);
908 KASSERT(p != NULL, ("thread exiting without a process"));
909 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
910 (long)p->p_pid, td->td_name);
911 SDT_PROBE0(proc, , , lwp__exit);
912 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
913 MPASS(td->td_realucred == td->td_ucred);
916 * drop FPU & debug register state storage, or any other
917 * architecture specific resources that
918 * would not be on a new untouched process.
923 * The last thread is left attached to the process
924 * So that the whole bundle gets recycled. Skip
925 * all this stuff if we never had threads.
926 * EXIT clears all sign of other threads when
927 * it goes to single threading, so the last thread always
928 * takes the short path.
930 if (p->p_flag & P_HADTHREADS) {
931 if (p->p_numthreads > 1) {
932 atomic_add_int(&td->td_proc->p_exitthreads, 1);
934 td2 = FIRST_THREAD_IN_PROC(p);
935 sched_exit_thread(td2, td);
938 * The test below is NOT true if we are the
939 * sole exiting thread. P_STOPPED_SINGLE is unset
940 * in exit1() after it is the only survivor.
942 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
943 if (p->p_numthreads == p->p_suspcount) {
944 thread_lock(p->p_singlethread);
945 wakeup_swapper = thread_unsuspend_one(
946 p->p_singlethread, p, false);
952 PCPU_SET(deadthread, td);
955 * The last thread is exiting.. but not through exit()
957 panic ("thread_exit: Last thread exiting on its own");
962 * If this thread is part of a process that is being tracked by hwpmc(4),
963 * inform the module of the thread's impending exit.
965 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
966 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
967 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
968 } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
969 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
976 /* Do the same timestamp bookkeeping that mi_switch() would do. */
977 new_switchtime = cpu_ticks();
978 runtime = new_switchtime - PCPU_GET(switchtime);
979 td->td_runtime += runtime;
980 td->td_incruntime += runtime;
981 PCPU_SET(switchtime, new_switchtime);
982 PCPU_SET(switchticks, ticks);
985 /* Save our resource usage in our process. */
986 td->td_ru.ru_nvcsw++;
987 ruxagg_locked(p, td);
988 rucollect(&p->p_ru, &td->td_ru);
991 TD_SET_STATE(td, TDS_INACTIVE);
993 witness_thread_exit(td);
995 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
997 panic("I'm a teapot!");
1002 * Do any thread specific cleanups that may be needed in wait()
1003 * called with Giant, proc and schedlock not held.
1006 thread_wait(struct proc *p)
1010 mtx_assert(&Giant, MA_NOTOWNED);
1011 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
1012 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
1013 td = FIRST_THREAD_IN_PROC(p);
1014 /* Lock the last thread so we spin until it exits cpu_throw(). */
1017 lock_profile_thread_exit(td);
1018 cpuset_rel(td->td_cpuset);
1019 td->td_cpuset = NULL;
1020 cpu_thread_clean(td);
1021 thread_cow_free(td);
1022 callout_drain(&td->td_slpcallout);
1023 thread_reap(); /* check for zombie threads etc. */
1027 * Link a thread to a process.
1028 * set up anything that needs to be initialized for it to
1029 * be used by the process.
1032 thread_link(struct thread *td, struct proc *p)
1036 * XXX This can't be enabled because it's called for proc0 before
1037 * its lock has been created.
1038 * PROC_LOCK_ASSERT(p, MA_OWNED);
1040 TD_SET_STATE(td, TDS_INACTIVE);
1042 td->td_flags = TDF_INMEM;
1044 LIST_INIT(&td->td_contested);
1045 LIST_INIT(&td->td_lprof[0]);
1046 LIST_INIT(&td->td_lprof[1]);
1048 SLIST_INIT(&td->td_epochs);
1050 sigqueue_init(&td->td_sigqueue, p);
1051 callout_init(&td->td_slpcallout, 1);
1052 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
1061 thread_unlink(struct thread *td)
1063 struct proc *p = td->td_proc;
1065 PROC_LOCK_ASSERT(p, MA_OWNED);
1067 MPASS(SLIST_EMPTY(&td->td_epochs));
1070 TAILQ_REMOVE(&p->p_threads, td, td_plist);
1072 /* could clear a few other things here */
1073 /* Must NOT clear links to proc! */
1077 calc_remaining(struct proc *p, int mode)
1081 PROC_LOCK_ASSERT(p, MA_OWNED);
1082 PROC_SLOCK_ASSERT(p, MA_OWNED);
1083 if (mode == SINGLE_EXIT)
1084 remaining = p->p_numthreads;
1085 else if (mode == SINGLE_BOUNDARY)
1086 remaining = p->p_numthreads - p->p_boundary_count;
1087 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
1088 remaining = p->p_numthreads - p->p_suspcount;
1090 panic("calc_remaining: wrong mode %d", mode);
1095 remain_for_mode(int mode)
1098 return (mode == SINGLE_ALLPROC ? 0 : 1);
1102 weed_inhib(int mode, struct thread *td2, struct proc *p)
1106 PROC_LOCK_ASSERT(p, MA_OWNED);
1107 PROC_SLOCK_ASSERT(p, MA_OWNED);
1108 THREAD_LOCK_ASSERT(td2, MA_OWNED);
1113 * Since the thread lock is dropped by the scheduler we have
1114 * to retry to check for races.
1119 if (TD_IS_SUSPENDED(td2)) {
1120 wakeup_swapper |= thread_unsuspend_one(td2, p, true);
1124 if (TD_CAN_ABORT(td2)) {
1125 wakeup_swapper |= sleepq_abort(td2, EINTR);
1126 return (wakeup_swapper);
1129 case SINGLE_BOUNDARY:
1130 case SINGLE_NO_EXIT:
1131 if (TD_IS_SUSPENDED(td2) &&
1132 (td2->td_flags & TDF_BOUNDARY) == 0) {
1133 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1137 if (TD_CAN_ABORT(td2)) {
1138 wakeup_swapper |= sleepq_abort(td2, ERESTART);
1139 return (wakeup_swapper);
1142 case SINGLE_ALLPROC:
1144 * ALLPROC suspend tries to avoid spurious EINTR for
1145 * threads sleeping interruptable, by suspending the
1146 * thread directly, similarly to sig_suspend_threads().
1147 * Since such sleep is not performed at the user
1148 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
1149 * is used to avoid immediate un-suspend.
1151 if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
1152 TDF_ALLPROCSUSP)) == 0) {
1153 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1157 if (TD_CAN_ABORT(td2)) {
1158 if ((td2->td_flags & TDF_SBDRY) == 0) {
1159 thread_suspend_one(td2);
1160 td2->td_flags |= TDF_ALLPROCSUSP;
1162 wakeup_swapper |= sleepq_abort(td2, ERESTART);
1163 return (wakeup_swapper);
1171 return (wakeup_swapper);
1175 * Enforce single-threading.
1177 * Returns 1 if the caller must abort (another thread is waiting to
1178 * exit the process or similar). Process is locked!
1179 * Returns 0 when you are successfully the only thread running.
1180 * A process has successfully single threaded in the suspend mode when
1181 * There are no threads in user mode. Threads in the kernel must be
1182 * allowed to continue until they get to the user boundary. They may even
1183 * copy out their return values and data before suspending. They may however be
1184 * accelerated in reaching the user boundary as we will wake up
1185 * any sleeping threads that are interruptable. (PCATCH).
1188 thread_single(struct proc *p, int mode)
1192 int remaining, wakeup_swapper;
1195 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1196 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1197 ("invalid mode %d", mode));
1199 * If allowing non-ALLPROC singlethreading for non-curproc
1200 * callers, calc_remaining() and remain_for_mode() should be
1201 * adjusted to also account for td->td_proc != p. For now
1202 * this is not implemented because it is not used.
1204 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
1205 (mode != SINGLE_ALLPROC && td->td_proc == p),
1206 ("mode %d proc %p curproc %p", mode, p, td->td_proc));
1207 mtx_assert(&Giant, MA_NOTOWNED);
1208 PROC_LOCK_ASSERT(p, MA_OWNED);
1210 if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
1213 /* Is someone already single threading? */
1214 if (p->p_singlethread != NULL && p->p_singlethread != td)
1217 if (mode == SINGLE_EXIT) {
1218 p->p_flag |= P_SINGLE_EXIT;
1219 p->p_flag &= ~P_SINGLE_BOUNDARY;
1221 p->p_flag &= ~P_SINGLE_EXIT;
1222 if (mode == SINGLE_BOUNDARY)
1223 p->p_flag |= P_SINGLE_BOUNDARY;
1225 p->p_flag &= ~P_SINGLE_BOUNDARY;
1227 if (mode == SINGLE_ALLPROC)
1228 p->p_flag |= P_TOTAL_STOP;
1229 p->p_flag |= P_STOPPED_SINGLE;
1231 p->p_singlethread = td;
1232 remaining = calc_remaining(p, mode);
1233 while (remaining != remain_for_mode(mode)) {
1234 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
1237 FOREACH_THREAD_IN_PROC(p, td2) {
1241 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
1242 if (TD_IS_INHIBITED(td2)) {
1243 wakeup_swapper |= weed_inhib(mode, td2, p);
1245 } else if (TD_IS_RUNNING(td2) && td != td2) {
1246 forward_signal(td2);
1254 remaining = calc_remaining(p, mode);
1257 * Maybe we suspended some threads.. was it enough?
1259 if (remaining == remain_for_mode(mode))
1264 * Wake us up when everyone else has suspended.
1265 * In the mean time we suspend as well.
1267 thread_suspend_switch(td, p);
1268 remaining = calc_remaining(p, mode);
1270 if (mode == SINGLE_EXIT) {
1272 * Convert the process to an unthreaded process. The
1273 * SINGLE_EXIT is called by exit1() or execve(), in
1274 * both cases other threads must be retired.
1276 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1277 p->p_singlethread = NULL;
1278 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1281 * Wait for any remaining threads to exit cpu_throw().
1283 while (p->p_exitthreads != 0) {
1286 sched_relinquish(td);
1290 } else if (mode == SINGLE_BOUNDARY) {
1292 * Wait until all suspended threads are removed from
1293 * the processors. The thread_suspend_check()
1294 * increments p_boundary_count while it is still
1295 * running, which makes it possible for the execve()
1296 * to destroy vmspace while our other threads are
1297 * still using the address space.
1299 * We lock the thread, which is only allowed to
1300 * succeed after context switch code finished using
1301 * the address space.
1303 FOREACH_THREAD_IN_PROC(p, td2) {
1307 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1308 ("td %p not on boundary", td2));
1309 KASSERT(TD_IS_SUSPENDED(td2),
1310 ("td %p is not suspended", td2));
1319 thread_suspend_check_needed(void)
1326 PROC_LOCK_ASSERT(p, MA_OWNED);
1327 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1328 (td->td_dbgflags & TDB_SUSPEND) != 0));
1332 * Called in from locations that can safely check to see
1333 * whether we have to suspend or at least throttle for a
1334 * single-thread event (e.g. fork).
1336 * Such locations include userret().
1337 * If the "return_instead" argument is non zero, the thread must be able to
1338 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1340 * The 'return_instead' argument tells the function if it may do a
1341 * thread_exit() or suspend, or whether the caller must abort and back
1344 * If the thread that set the single_threading request has set the
1345 * P_SINGLE_EXIT bit in the process flags then this call will never return
1346 * if 'return_instead' is false, but will exit.
1348 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1349 *---------------+--------------------+---------------------
1350 * 0 | returns 0 | returns 0 or 1
1351 * | when ST ends | immediately
1352 *---------------+--------------------+---------------------
1353 * 1 | thread exits | returns 1
1355 * 0 = thread_exit() or suspension ok,
1356 * other = return error instead of stopping the thread.
1358 * While a full suspension is under effect, even a single threading
1359 * thread would be suspended if it made this call (but it shouldn't).
1360 * This call should only be made from places where
1361 * thread_exit() would be safe as that may be the outcome unless
1362 * return_instead is set.
1365 thread_suspend_check(int return_instead)
1373 mtx_assert(&Giant, MA_NOTOWNED);
1374 PROC_LOCK_ASSERT(p, MA_OWNED);
1375 while (thread_suspend_check_needed()) {
1376 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1377 KASSERT(p->p_singlethread != NULL,
1378 ("singlethread not set"));
1380 * The only suspension in action is a
1381 * single-threading. Single threader need not stop.
1382 * It is safe to access p->p_singlethread unlocked
1383 * because it can only be set to our address by us.
1385 if (p->p_singlethread == td)
1386 return (0); /* Exempt from stopping. */
1388 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1391 /* Should we goto user boundary if we didn't come from there? */
1392 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1393 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1397 * Ignore suspend requests if they are deferred.
1399 if ((td->td_flags & TDF_SBDRY) != 0) {
1400 KASSERT(return_instead,
1401 ("TDF_SBDRY set for unsafe thread_suspend_check"));
1402 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1403 (TDF_SEINTR | TDF_SERESTART),
1404 ("both TDF_SEINTR and TDF_SERESTART"));
1405 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1409 * If the process is waiting for us to exit,
1410 * this thread should just suicide.
1411 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1413 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1417 * Allow Linux emulation layer to do some work
1418 * before thread suicide.
1420 if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1421 (p->p_sysent->sv_thread_detach)(td);
1422 umtx_thread_exit(td);
1424 panic("stopped thread did not exit");
1429 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1430 if (p->p_numthreads == p->p_suspcount + 1) {
1431 thread_lock(p->p_singlethread);
1432 wakeup_swapper = thread_unsuspend_one(
1433 p->p_singlethread, p, false);
1441 * When a thread suspends, it just
1442 * gets taken off all queues.
1444 thread_suspend_one(td);
1445 if (return_instead == 0) {
1446 p->p_boundary_count++;
1447 td->td_flags |= TDF_BOUNDARY;
1450 mi_switch(SW_INVOL | SWT_SUSPEND);
1457 * Check for possible stops and suspensions while executing a
1458 * casueword or similar transiently failing operation.
1460 * The sleep argument controls whether the function can handle a stop
1461 * request itself or it should return ERESTART and the request is
1462 * proceed at the kernel/user boundary in ast.
1464 * Typically, when retrying due to casueword(9) failure (rv == 1), we
1465 * should handle the stop requests there, with exception of cases when
1466 * the thread owns a kernel resource, for instance busied the umtx
1467 * key, or when functions return immediately if thread_check_susp()
1468 * returned non-zero. On the other hand, retrying the whole lock
1469 * operation, we better not stop there but delegate the handling to
1472 * If the request is for thread termination P_SINGLE_EXIT, we cannot
1473 * handle it at all, and simply return EINTR.
1476 thread_check_susp(struct thread *td, bool sleep)
1482 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
1483 * eventually break the lockstep loop.
1485 if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
1490 if (p->p_flag & P_SINGLE_EXIT)
1492 else if (P_SHOULDSTOP(p) ||
1493 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1494 error = sleep ? thread_suspend_check(0) : ERESTART;
1500 thread_suspend_switch(struct thread *td, struct proc *p)
1503 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1504 PROC_LOCK_ASSERT(p, MA_OWNED);
1505 PROC_SLOCK_ASSERT(p, MA_OWNED);
1507 * We implement thread_suspend_one in stages here to avoid
1508 * dropping the proc lock while the thread lock is owned.
1510 if (p == td->td_proc) {
1516 td->td_flags &= ~TDF_NEEDSUSPCHK;
1517 TD_SET_SUSPENDED(td);
1521 mi_switch(SW_VOL | SWT_SUSPEND);
1528 thread_suspend_one(struct thread *td)
1533 PROC_SLOCK_ASSERT(p, MA_OWNED);
1534 THREAD_LOCK_ASSERT(td, MA_OWNED);
1535 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1537 td->td_flags &= ~TDF_NEEDSUSPCHK;
1538 TD_SET_SUSPENDED(td);
1543 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1546 THREAD_LOCK_ASSERT(td, MA_OWNED);
1547 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1548 TD_CLR_SUSPENDED(td);
1549 td->td_flags &= ~TDF_ALLPROCSUSP;
1550 if (td->td_proc == p) {
1551 PROC_SLOCK_ASSERT(p, MA_OWNED);
1553 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1554 td->td_flags &= ~TDF_BOUNDARY;
1555 p->p_boundary_count--;
1558 return (setrunnable(td, 0));
1562 thread_run_flash(struct thread *td)
1567 PROC_LOCK_ASSERT(p, MA_OWNED);
1569 if (TD_ON_SLEEPQ(td))
1570 sleepq_remove_nested(td);
1574 THREAD_LOCK_ASSERT(td, MA_OWNED);
1575 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1577 TD_CLR_SUSPENDED(td);
1579 MPASS(p->p_suspcount > 0);
1582 if (setrunnable(td, 0))
1587 * Allow all threads blocked by single threading to continue running.
1590 thread_unsuspend(struct proc *p)
1595 PROC_LOCK_ASSERT(p, MA_OWNED);
1596 PROC_SLOCK_ASSERT(p, MA_OWNED);
1598 if (!P_SHOULDSTOP(p)) {
1599 FOREACH_THREAD_IN_PROC(p, td) {
1601 if (TD_IS_SUSPENDED(td)) {
1602 wakeup_swapper |= thread_unsuspend_one(td, p,
1607 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1608 p->p_numthreads == p->p_suspcount) {
1610 * Stopping everything also did the job for the single
1611 * threading request. Now we've downgraded to single-threaded,
1614 if (p->p_singlethread->td_proc == p) {
1615 thread_lock(p->p_singlethread);
1616 wakeup_swapper = thread_unsuspend_one(
1617 p->p_singlethread, p, false);
1625 * End the single threading mode..
1628 thread_single_end(struct proc *p, int mode)
1633 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1634 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1635 ("invalid mode %d", mode));
1636 PROC_LOCK_ASSERT(p, MA_OWNED);
1637 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1638 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1639 ("mode %d does not match P_TOTAL_STOP", mode));
1640 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1641 ("thread_single_end from other thread %p %p",
1642 curthread, p->p_singlethread));
1643 KASSERT(mode != SINGLE_BOUNDARY ||
1644 (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1645 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1646 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1649 p->p_singlethread = NULL;
1652 * If there are other threads they may now run,
1653 * unless of course there is a blanket 'stop order'
1654 * on the process. The single threader must be allowed
1655 * to continue however as this is a bad place to stop.
1657 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1658 FOREACH_THREAD_IN_PROC(p, td) {
1660 if (TD_IS_SUSPENDED(td)) {
1661 wakeup_swapper |= thread_unsuspend_one(td, p,
1662 mode == SINGLE_BOUNDARY);
1667 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1668 ("inconsistent boundary count %d", p->p_boundary_count));
1675 * Locate a thread by number and return with proc lock held.
1677 * thread exit establishes proc -> tidhash lock ordering, but lookup
1678 * takes tidhash first and needs to return locked proc.
1680 * The problem is worked around by relying on type-safety of both
1681 * structures and doing the work in 2 steps:
1682 * - tidhash-locked lookup which saves both thread and proc pointers
1683 * - proc-locked verification that the found thread still matches
1686 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1688 #define RUN_THRESH 16
1695 rw_rlock(TIDHASHLOCK(tid));
1697 LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1698 if (td->td_tid != tid) {
1703 if (pid != -1 && p->p_pid != pid) {
1707 if (run > RUN_THRESH) {
1708 if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1709 LIST_REMOVE(td, td_hash);
1710 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1712 rw_wunlock(TIDHASHLOCK(tid));
1720 rw_runlock(TIDHASHLOCK(tid));
1729 tdfind(lwpid_t tid, pid_t pid)
1735 if (td->td_tid == tid) {
1736 if (pid != -1 && td->td_proc->p_pid != pid)
1738 PROC_LOCK(td->td_proc);
1743 if (!tdfind_hash(tid, pid, &p, &td))
1746 if (td->td_tid != tid) {
1750 if (td->td_proc != p) {
1754 if (p->p_state == PRS_NEW) {
1763 tidhash_add(struct thread *td)
1765 rw_wlock(TIDHASHLOCK(td->td_tid));
1766 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1767 rw_wunlock(TIDHASHLOCK(td->td_tid));
1771 tidhash_remove(struct thread *td)
1774 rw_wlock(TIDHASHLOCK(td->td_tid));
1775 LIST_REMOVE(td, td_hash);
1776 rw_wunlock(TIDHASHLOCK(td->td_tid));