2 * SPDX-License-Identifier: BSD-2-Clause
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) == 0x4b8,
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) == 0x4d0,
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) == 0x314,
113 "struct thread KBI td_frame");
114 _Static_assert(offsetof(struct thread, td_emuldata) == 0x358,
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) == 0x318,
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 * Batching thread count free, for consistency
344 struct tdcountbatch {
349 tdcountbatch_prep(struct tdcountbatch *tb)
356 tdcountbatch_add(struct tdcountbatch *tb, struct thread *td __unused)
363 tdcountbatch_process(struct tdcountbatch *tb)
367 thread_count_sub(tb->n);
373 tdcountbatch_final(struct tdcountbatch *tb)
377 thread_count_sub(tb->n);
382 * Prepare a thread for use.
385 thread_ctor(void *mem, int size, void *arg, int flags)
389 td = (struct thread *)mem;
390 TD_SET_STATE(td, TDS_INACTIVE);
391 td->td_lastcpu = td->td_oncpu = NOCPU;
394 * Note that td_critnest begins life as 1 because the thread is not
395 * running and is thereby implicitly waiting to be on the receiving
396 * end of a context switch.
399 td->td_lend_user_pri = PRI_MAX;
401 audit_thread_alloc(td);
404 kdtrace_thread_ctor(td);
406 umtx_thread_alloc(td);
407 MPASS(td->td_sel == NULL);
412 * Reclaim a thread after use.
415 thread_dtor(void *mem, int size, void *arg)
419 td = (struct thread *)mem;
422 /* Verify that this thread is in a safe state to free. */
423 switch (TD_GET_STATE(td)) {
429 * We must never unlink a thread that is in one of
430 * these states, because it is currently active.
432 panic("bad state for thread unlinking");
437 panic("bad thread state");
442 audit_thread_free(td);
445 kdtrace_thread_dtor(td);
447 /* Free all OSD associated to this thread. */
454 * Initialize type-stable parts of a thread (when newly created).
457 thread_init(void *mem, int size, int flags)
461 td = (struct thread *)mem;
463 td->td_allocdomain = vm_phys_domain(vtophys(td));
464 td->td_sleepqueue = sleepq_alloc();
465 td->td_turnstile = turnstile_alloc();
467 EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
468 umtx_thread_init(td);
475 * Tear down type-stable parts of a thread (just before being discarded).
478 thread_fini(void *mem, int size)
482 td = (struct thread *)mem;
483 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
484 rlqentry_free(td->td_rlqe);
485 turnstile_free(td->td_turnstile);
486 sleepq_free(td->td_sleepqueue);
487 umtx_thread_fini(td);
488 MPASS(td->td_sel == NULL);
492 * For a newly created process,
493 * link up all the structures and its initial threads etc.
495 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc.
496 * proc_dtor() (should go away)
500 proc_linkup0(struct proc *p, struct thread *td)
502 TAILQ_INIT(&p->p_threads); /* all threads in proc */
507 proc_linkup(struct proc *p, struct thread *td)
510 sigqueue_init(&p->p_sigqueue, p);
511 p->p_ksi = ksiginfo_alloc(M_WAITOK);
512 if (p->p_ksi != NULL) {
513 /* XXX p_ksi may be null if ksiginfo zone is not ready */
514 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
516 LIST_INIT(&p->p_mqnotifier);
522 ast_suspend(struct thread *td, int tda __unused)
528 * We need to check to see if we have to exit or wait due to a
529 * single threading requirement or some other STOP condition.
532 thread_suspend_check(0);
536 extern int max_threads_per_proc;
539 * Initialize global thread allocation resources.
548 * Place an upper limit on threads which can be allocated.
550 * Note that other factors may make the de facto limit much lower.
552 * Platform limits are somewhat arbitrary but deemed "more than good
553 * enough" for the foreseable future.
555 if (maxthread == 0) {
557 maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
559 maxthread = MIN(maxproc * max_threads_per_proc, 100000);
563 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
564 tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
570 if (tid0 != THREAD0_TID)
571 panic("tid0 %d != %d\n", tid0, THREAD0_TID);
573 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
574 thread_ctor, thread_dtor, thread_init, thread_fini,
575 32 - 1, UMA_ZONE_NOFREE);
576 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
577 tidhashlock = (tidhash + 1) / 64;
580 tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
581 M_TIDHASH, M_WAITOK | M_ZERO);
582 for (i = 0; i < tidhashlock + 1; i++)
583 rw_init(&tidhashtbl_lock[i], "tidhash");
585 TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL);
586 callout_init(&thread_reap_callout, 1);
587 callout_reset(&thread_reap_callout, 5 * hz,
588 thread_reap_callout_cb, NULL);
589 ast_register(TDA_SUSPEND, ASTR_ASTF_REQUIRED, 0, ast_suspend);
593 * Place an unused thread on the zombie list.
596 thread_zombie(struct thread *td)
598 struct thread_domain_data *tdd;
601 tdd = &thread_domain_data[td->td_allocdomain];
602 ztd = atomic_load_ptr(&tdd->tdd_zombies);
605 if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
606 (uintptr_t *)&ztd, (uintptr_t)td))
613 * Release a thread that has exited after cpu_throw().
616 thread_stash(struct thread *td)
618 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
623 * Reap zombies from passed domain.
626 thread_reap_domain(struct thread_domain_data *tdd)
628 struct thread *itd, *ntd;
629 struct tidbatch tidbatch;
630 struct credbatch credbatch;
631 struct limbatch limbatch;
632 struct tdcountbatch tdcountbatch;
635 * Reading upfront is pessimal if followed by concurrent atomic_swap,
636 * but most of the time the list is empty.
638 if (tdd->tdd_zombies == NULL)
641 itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
647 * Multiple CPUs can get here, the race is fine as ticks is only
650 tdd->tdd_reapticks = ticks;
652 tidbatch_prep(&tidbatch);
653 credbatch_prep(&credbatch);
654 limbatch_prep(&limbatch);
655 tdcountbatch_prep(&tdcountbatch);
657 while (itd != NULL) {
658 ntd = itd->td_zombie;
659 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd);
661 tidbatch_add(&tidbatch, itd);
662 credbatch_add(&credbatch, itd);
663 limbatch_add(&limbatch, itd);
664 tdcountbatch_add(&tdcountbatch, itd);
666 thread_free_batched(itd);
668 tidbatch_process(&tidbatch);
669 credbatch_process(&credbatch);
670 limbatch_process(&limbatch);
671 tdcountbatch_process(&tdcountbatch);
676 tidbatch_final(&tidbatch);
677 credbatch_final(&credbatch);
678 limbatch_final(&limbatch);
679 tdcountbatch_final(&tdcountbatch);
683 * Reap zombies from all domains.
686 thread_reap_all(void)
688 struct thread_domain_data *tdd;
691 domain = PCPU_GET(domain);
692 for (i = 0; i < vm_ndomains; i++) {
693 tdd = &thread_domain_data[(i + domain) % vm_ndomains];
694 thread_reap_domain(tdd);
699 * Reap zombies from local domain.
704 struct thread_domain_data *tdd;
707 domain = PCPU_GET(domain);
708 tdd = &thread_domain_data[domain];
710 thread_reap_domain(tdd);
714 thread_reap_task_cb(void *arg __unused, int pending __unused)
721 thread_reap_callout_cb(void *arg __unused)
723 struct thread_domain_data *tdd;
724 int i, cticks, lticks;
728 cticks = atomic_load_int(&ticks);
729 for (i = 0; i < vm_ndomains; i++) {
730 tdd = &thread_domain_data[i];
731 lticks = tdd->tdd_reapticks;
732 if (tdd->tdd_zombies != NULL &&
733 (u_int)(cticks - lticks) > 5 * hz) {
740 taskqueue_enqueue(taskqueue_thread, &thread_reap_task);
741 callout_reset(&thread_reap_callout, 5 * hz,
742 thread_reap_callout_cb, NULL);
746 * Calling this function guarantees that any thread that exited before
747 * the call is reaped when the function returns. By 'exited' we mean
748 * a thread removed from the process linkage with thread_unlink().
749 * Practically this means that caller must lock/unlock corresponding
750 * process lock before the call, to synchronize with thread_exit().
753 thread_reap_barrier(void)
758 * First do context switches to each CPU to ensure that all
759 * PCPU pc_deadthreads are moved to zombie list.
761 quiesce_all_cpus("", PDROP);
764 * Second, fire the task in the same thread as normal
765 * thread_reap() is done, to serialize reaping.
767 t = malloc(sizeof(*t), M_TEMP, M_WAITOK);
768 TASK_INIT(t, 0, thread_reap_task_cb, t);
769 taskqueue_enqueue(taskqueue_thread, t);
770 taskqueue_drain(taskqueue_thread, t);
778 thread_alloc(int pages)
783 if (!thread_count_inc()) {
788 td = uma_zalloc(thread_zone, M_WAITOK);
789 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
790 if (!vm_thread_new(td, pages)) {
791 uma_zfree(thread_zone, td);
797 bzero(&td->td_sa.args, sizeof(td->td_sa.args));
798 kmsan_thread_alloc(td);
799 cpu_thread_alloc(td);
800 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
805 thread_alloc_stack(struct thread *td, int pages)
808 KASSERT(td->td_kstack == 0,
809 ("thread_alloc_stack called on a thread with kstack"));
810 if (!vm_thread_new(td, pages))
812 cpu_thread_alloc(td);
817 * Deallocate a thread.
820 thread_free_batched(struct thread *td)
823 lock_profile_thread_exit(td);
825 cpuset_rel(td->td_cpuset);
826 td->td_cpuset = NULL;
828 if (td->td_kstack != 0)
829 vm_thread_dispose(td);
830 callout_drain(&td->td_slpcallout);
832 * Freeing handled by the caller.
835 kmsan_thread_free(td);
836 uma_zfree(thread_zone, td);
840 thread_free(struct thread *td)
844 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
846 thread_free_batched(td);
852 thread_cow_get_proc(struct thread *newtd, struct proc *p)
855 PROC_LOCK_ASSERT(p, MA_OWNED);
856 newtd->td_realucred = crcowget(p->p_ucred);
857 newtd->td_ucred = newtd->td_realucred;
858 newtd->td_limit = lim_hold(p->p_limit);
859 newtd->td_cowgen = p->p_cowgen;
863 thread_cow_get(struct thread *newtd, struct thread *td)
866 MPASS(td->td_realucred == td->td_ucred);
867 newtd->td_realucred = crcowget(td->td_realucred);
868 newtd->td_ucred = newtd->td_realucred;
869 newtd->td_limit = lim_hold(td->td_limit);
870 newtd->td_cowgen = td->td_cowgen;
874 thread_cow_free(struct thread *td)
877 if (td->td_realucred != NULL)
879 if (td->td_limit != NULL)
880 lim_free(td->td_limit);
884 thread_cow_update(struct thread *td)
887 struct ucred *oldcred;
888 struct plimit *oldlimit;
892 oldcred = crcowsync();
893 oldlimit = lim_cowsync();
894 td->td_cowgen = p->p_cowgen;
898 if (oldlimit != NULL)
903 thread_cow_synced(struct thread *td)
908 PROC_LOCK_ASSERT(p, MA_OWNED);
909 MPASS(td->td_cowgen != p->p_cowgen);
910 MPASS(td->td_ucred == p->p_ucred);
911 MPASS(td->td_limit == p->p_limit);
912 td->td_cowgen = p->p_cowgen;
916 * Discard the current thread and exit from its context.
917 * Always called with scheduler locked.
919 * Because we can't free a thread while we're operating under its context,
920 * push the current thread into our CPU's deadthread holder. This means
921 * we needn't worry about someone else grabbing our context before we
927 uint64_t runtime, new_switchtime;
936 PROC_SLOCK_ASSERT(p, MA_OWNED);
937 mtx_assert(&Giant, MA_NOTOWNED);
939 PROC_LOCK_ASSERT(p, MA_OWNED);
940 KASSERT(p != NULL, ("thread exiting without a process"));
941 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
942 (long)p->p_pid, td->td_name);
943 SDT_PROBE0(proc, , , lwp__exit);
944 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
945 MPASS(td->td_realucred == td->td_ucred);
948 * drop FPU & debug register state storage, or any other
949 * architecture specific resources that
950 * would not be on a new untouched process.
955 * The last thread is left attached to the process
956 * So that the whole bundle gets recycled. Skip
957 * all this stuff if we never had threads.
958 * EXIT clears all sign of other threads when
959 * it goes to single threading, so the last thread always
960 * takes the short path.
962 if (p->p_flag & P_HADTHREADS) {
963 if (p->p_numthreads > 1) {
964 atomic_add_int(&td->td_proc->p_exitthreads, 1);
966 td2 = FIRST_THREAD_IN_PROC(p);
967 sched_exit_thread(td2, td);
970 * The test below is NOT true if we are the
971 * sole exiting thread. P_STOPPED_SINGLE is unset
972 * in exit1() after it is the only survivor.
974 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
975 if (p->p_numthreads == p->p_suspcount) {
976 thread_lock(p->p_singlethread);
977 wakeup_swapper = thread_unsuspend_one(
978 p->p_singlethread, p, false);
984 PCPU_SET(deadthread, td);
987 * The last thread is exiting.. but not through exit()
989 panic ("thread_exit: Last thread exiting on its own");
994 * If this thread is part of a process that is being tracked by hwpmc(4),
995 * inform the module of the thread's impending exit.
997 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
998 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
999 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
1000 } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
1001 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
1008 /* Do the same timestamp bookkeeping that mi_switch() would do. */
1009 new_switchtime = cpu_ticks();
1010 runtime = new_switchtime - PCPU_GET(switchtime);
1011 td->td_runtime += runtime;
1012 td->td_incruntime += runtime;
1013 PCPU_SET(switchtime, new_switchtime);
1014 PCPU_SET(switchticks, ticks);
1015 VM_CNT_INC(v_swtch);
1017 /* Save our resource usage in our process. */
1018 td->td_ru.ru_nvcsw++;
1019 ruxagg_locked(p, td);
1020 rucollect(&p->p_ru, &td->td_ru);
1023 TD_SET_STATE(td, TDS_INACTIVE);
1025 witness_thread_exit(td);
1027 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
1029 panic("I'm a teapot!");
1034 * Do any thread specific cleanups that may be needed in wait()
1035 * called with Giant, proc and schedlock not held.
1038 thread_wait(struct proc *p)
1042 mtx_assert(&Giant, MA_NOTOWNED);
1043 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
1044 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
1045 td = FIRST_THREAD_IN_PROC(p);
1046 /* Lock the last thread so we spin until it exits cpu_throw(). */
1049 lock_profile_thread_exit(td);
1050 cpuset_rel(td->td_cpuset);
1051 td->td_cpuset = NULL;
1052 cpu_thread_clean(td);
1053 thread_cow_free(td);
1054 callout_drain(&td->td_slpcallout);
1055 thread_reap(); /* check for zombie threads etc. */
1059 * Link a thread to a process.
1060 * set up anything that needs to be initialized for it to
1061 * be used by the process.
1064 thread_link(struct thread *td, struct proc *p)
1068 * XXX This can't be enabled because it's called for proc0 before
1069 * its lock has been created.
1070 * PROC_LOCK_ASSERT(p, MA_OWNED);
1072 TD_SET_STATE(td, TDS_INACTIVE);
1074 td->td_flags = TDF_INMEM;
1076 LIST_INIT(&td->td_contested);
1077 LIST_INIT(&td->td_lprof[0]);
1078 LIST_INIT(&td->td_lprof[1]);
1080 SLIST_INIT(&td->td_epochs);
1082 sigqueue_init(&td->td_sigqueue, p);
1083 callout_init(&td->td_slpcallout, 1);
1084 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
1093 thread_unlink(struct thread *td)
1095 struct proc *p = td->td_proc;
1097 PROC_LOCK_ASSERT(p, MA_OWNED);
1099 MPASS(SLIST_EMPTY(&td->td_epochs));
1102 TAILQ_REMOVE(&p->p_threads, td, td_plist);
1104 /* could clear a few other things here */
1105 /* Must NOT clear links to proc! */
1109 calc_remaining(struct proc *p, int mode)
1113 PROC_LOCK_ASSERT(p, MA_OWNED);
1114 PROC_SLOCK_ASSERT(p, MA_OWNED);
1115 if (mode == SINGLE_EXIT)
1116 remaining = p->p_numthreads;
1117 else if (mode == SINGLE_BOUNDARY)
1118 remaining = p->p_numthreads - p->p_boundary_count;
1119 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
1120 remaining = p->p_numthreads - p->p_suspcount;
1122 panic("calc_remaining: wrong mode %d", mode);
1127 remain_for_mode(int mode)
1130 return (mode == SINGLE_ALLPROC ? 0 : 1);
1134 weed_inhib(int mode, struct thread *td2, struct proc *p)
1138 PROC_LOCK_ASSERT(p, MA_OWNED);
1139 PROC_SLOCK_ASSERT(p, MA_OWNED);
1140 THREAD_LOCK_ASSERT(td2, MA_OWNED);
1145 * Since the thread lock is dropped by the scheduler we have
1146 * to retry to check for races.
1151 if (TD_IS_SUSPENDED(td2)) {
1152 wakeup_swapper |= thread_unsuspend_one(td2, p, true);
1156 if (TD_CAN_ABORT(td2)) {
1157 wakeup_swapper |= sleepq_abort(td2, EINTR);
1158 return (wakeup_swapper);
1161 case SINGLE_BOUNDARY:
1162 case SINGLE_NO_EXIT:
1163 if (TD_IS_SUSPENDED(td2) &&
1164 (td2->td_flags & TDF_BOUNDARY) == 0) {
1165 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1169 if (TD_CAN_ABORT(td2)) {
1170 wakeup_swapper |= sleepq_abort(td2, ERESTART);
1171 return (wakeup_swapper);
1174 case SINGLE_ALLPROC:
1176 * ALLPROC suspend tries to avoid spurious EINTR for
1177 * threads sleeping interruptable, by suspending the
1178 * thread directly, similarly to sig_suspend_threads().
1179 * Since such sleep is not neccessary performed at the user
1180 * boundary, TDF_ALLPROCSUSP is used to avoid immediate
1183 if (TD_IS_SUSPENDED(td2) &&
1184 (td2->td_flags & TDF_ALLPROCSUSP) == 0) {
1185 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1189 if (TD_CAN_ABORT(td2)) {
1190 td2->td_flags |= TDF_ALLPROCSUSP;
1191 wakeup_swapper |= sleepq_abort(td2, ERESTART);
1192 return (wakeup_swapper);
1199 return (wakeup_swapper);
1203 * Enforce single-threading.
1205 * Returns 1 if the caller must abort (another thread is waiting to
1206 * exit the process or similar). Process is locked!
1207 * Returns 0 when you are successfully the only thread running.
1208 * A process has successfully single threaded in the suspend mode when
1209 * There are no threads in user mode. Threads in the kernel must be
1210 * allowed to continue until they get to the user boundary. They may even
1211 * copy out their return values and data before suspending. They may however be
1212 * accelerated in reaching the user boundary as we will wake up
1213 * any sleeping threads that are interruptable. (PCATCH).
1216 thread_single(struct proc *p, int mode)
1220 int remaining, wakeup_swapper;
1223 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1224 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1225 ("invalid mode %d", mode));
1227 * If allowing non-ALLPROC singlethreading for non-curproc
1228 * callers, calc_remaining() and remain_for_mode() should be
1229 * adjusted to also account for td->td_proc != p. For now
1230 * this is not implemented because it is not used.
1232 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
1233 (mode != SINGLE_ALLPROC && td->td_proc == p),
1234 ("mode %d proc %p curproc %p", mode, p, td->td_proc));
1235 mtx_assert(&Giant, MA_NOTOWNED);
1236 PROC_LOCK_ASSERT(p, MA_OWNED);
1239 * Is someone already single threading?
1240 * Or may be singlethreading is not needed at all.
1242 if (mode == SINGLE_ALLPROC) {
1243 while ((p->p_flag & P_STOPPED_SINGLE) != 0) {
1244 if ((p->p_flag2 & P2_WEXIT) != 0)
1246 msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0);
1248 } else if ((p->p_flag & P_HADTHREADS) == 0)
1250 if (p->p_singlethread != NULL && p->p_singlethread != td)
1253 if (mode == SINGLE_EXIT) {
1254 p->p_flag |= P_SINGLE_EXIT;
1255 p->p_flag &= ~P_SINGLE_BOUNDARY;
1257 p->p_flag &= ~P_SINGLE_EXIT;
1258 if (mode == SINGLE_BOUNDARY)
1259 p->p_flag |= P_SINGLE_BOUNDARY;
1261 p->p_flag &= ~P_SINGLE_BOUNDARY;
1263 if (mode == SINGLE_ALLPROC)
1264 p->p_flag |= P_TOTAL_STOP;
1265 p->p_flag |= P_STOPPED_SINGLE;
1267 p->p_singlethread = td;
1268 remaining = calc_remaining(p, mode);
1269 while (remaining != remain_for_mode(mode)) {
1270 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
1273 FOREACH_THREAD_IN_PROC(p, td2) {
1277 ast_sched_locked(td2, TDA_SUSPEND);
1278 if (TD_IS_INHIBITED(td2)) {
1279 wakeup_swapper |= weed_inhib(mode, td2, p);
1281 } else if (TD_IS_RUNNING(td2)) {
1282 forward_signal(td2);
1290 remaining = calc_remaining(p, mode);
1293 * Maybe we suspended some threads.. was it enough?
1295 if (remaining == remain_for_mode(mode))
1300 * Wake us up when everyone else has suspended.
1301 * In the mean time we suspend as well.
1303 thread_suspend_switch(td, p);
1304 remaining = calc_remaining(p, mode);
1306 if (mode == SINGLE_EXIT) {
1308 * Convert the process to an unthreaded process. The
1309 * SINGLE_EXIT is called by exit1() or execve(), in
1310 * both cases other threads must be retired.
1312 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1313 p->p_singlethread = NULL;
1314 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1317 * Wait for any remaining threads to exit cpu_throw().
1319 while (p->p_exitthreads != 0) {
1322 sched_relinquish(td);
1326 } else if (mode == SINGLE_BOUNDARY) {
1328 * Wait until all suspended threads are removed from
1329 * the processors. The thread_suspend_check()
1330 * increments p_boundary_count while it is still
1331 * running, which makes it possible for the execve()
1332 * to destroy vmspace while our other threads are
1333 * still using the address space.
1335 * We lock the thread, which is only allowed to
1336 * succeed after context switch code finished using
1337 * the address space.
1339 FOREACH_THREAD_IN_PROC(p, td2) {
1343 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1344 ("td %p not on boundary", td2));
1345 KASSERT(TD_IS_SUSPENDED(td2),
1346 ("td %p is not suspended", td2));
1355 thread_suspend_check_needed(void)
1362 PROC_LOCK_ASSERT(p, MA_OWNED);
1363 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1364 (td->td_dbgflags & TDB_SUSPEND) != 0));
1368 * Called in from locations that can safely check to see
1369 * whether we have to suspend or at least throttle for a
1370 * single-thread event (e.g. fork).
1372 * Such locations include userret().
1373 * If the "return_instead" argument is non zero, the thread must be able to
1374 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1376 * The 'return_instead' argument tells the function if it may do a
1377 * thread_exit() or suspend, or whether the caller must abort and back
1380 * If the thread that set the single_threading request has set the
1381 * P_SINGLE_EXIT bit in the process flags then this call will never return
1382 * if 'return_instead' is false, but will exit.
1384 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1385 *---------------+--------------------+---------------------
1386 * 0 | returns 0 | returns 0 or 1
1387 * | when ST ends | immediately
1388 *---------------+--------------------+---------------------
1389 * 1 | thread exits | returns 1
1391 * 0 = thread_exit() or suspension ok,
1392 * other = return error instead of stopping the thread.
1394 * While a full suspension is under effect, even a single threading
1395 * thread would be suspended if it made this call (but it shouldn't).
1396 * This call should only be made from places where
1397 * thread_exit() would be safe as that may be the outcome unless
1398 * return_instead is set.
1401 thread_suspend_check(int return_instead)
1409 mtx_assert(&Giant, MA_NOTOWNED);
1410 PROC_LOCK_ASSERT(p, MA_OWNED);
1411 while (thread_suspend_check_needed()) {
1412 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1413 KASSERT(p->p_singlethread != NULL,
1414 ("singlethread not set"));
1416 * The only suspension in action is a
1417 * single-threading. Single threader need not stop.
1418 * It is safe to access p->p_singlethread unlocked
1419 * because it can only be set to our address by us.
1421 if (p->p_singlethread == td)
1422 return (0); /* Exempt from stopping. */
1424 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1427 /* Should we goto user boundary if we didn't come from there? */
1428 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1429 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1433 * Ignore suspend requests if they are deferred.
1435 if ((td->td_flags & TDF_SBDRY) != 0) {
1436 KASSERT(return_instead,
1437 ("TDF_SBDRY set for unsafe thread_suspend_check"));
1438 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1439 (TDF_SEINTR | TDF_SERESTART),
1440 ("both TDF_SEINTR and TDF_SERESTART"));
1441 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1445 * If the process is waiting for us to exit,
1446 * this thread should just suicide.
1447 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1449 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1453 * Allow Linux emulation layer to do some work
1454 * before thread suicide.
1456 if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1457 (p->p_sysent->sv_thread_detach)(td);
1458 umtx_thread_exit(td);
1460 panic("stopped thread did not exit");
1465 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1466 if (p->p_numthreads == p->p_suspcount + 1) {
1467 thread_lock(p->p_singlethread);
1468 wakeup_swapper = thread_unsuspend_one(
1469 p->p_singlethread, p, false);
1477 * When a thread suspends, it just
1478 * gets taken off all queues.
1480 thread_suspend_one(td);
1481 if (return_instead == 0) {
1482 p->p_boundary_count++;
1483 td->td_flags |= TDF_BOUNDARY;
1486 mi_switch(SW_INVOL | SWT_SUSPEND);
1493 * Check for possible stops and suspensions while executing a
1494 * casueword or similar transiently failing operation.
1496 * The sleep argument controls whether the function can handle a stop
1497 * request itself or it should return ERESTART and the request is
1498 * proceed at the kernel/user boundary in ast.
1500 * Typically, when retrying due to casueword(9) failure (rv == 1), we
1501 * should handle the stop requests there, with exception of cases when
1502 * the thread owns a kernel resource, for instance busied the umtx
1503 * key, or when functions return immediately if thread_check_susp()
1504 * returned non-zero. On the other hand, retrying the whole lock
1505 * operation, we better not stop there but delegate the handling to
1508 * If the request is for thread termination P_SINGLE_EXIT, we cannot
1509 * handle it at all, and simply return EINTR.
1512 thread_check_susp(struct thread *td, bool sleep)
1518 * The check for TDA_SUSPEND is racy, but it is enough to
1519 * eventually break the lockstep loop.
1521 if (!td_ast_pending(td, TDA_SUSPEND))
1526 if (p->p_flag & P_SINGLE_EXIT)
1528 else if (P_SHOULDSTOP(p) ||
1529 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1530 error = sleep ? thread_suspend_check(0) : ERESTART;
1536 thread_suspend_switch(struct thread *td, struct proc *p)
1539 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1540 PROC_LOCK_ASSERT(p, MA_OWNED);
1541 PROC_SLOCK_ASSERT(p, MA_OWNED);
1543 * We implement thread_suspend_one in stages here to avoid
1544 * dropping the proc lock while the thread lock is owned.
1546 if (p == td->td_proc) {
1552 ast_unsched_locked(td, TDA_SUSPEND);
1553 TD_SET_SUSPENDED(td);
1557 mi_switch(SW_VOL | SWT_SUSPEND);
1564 thread_suspend_one(struct thread *td)
1569 PROC_SLOCK_ASSERT(p, MA_OWNED);
1570 THREAD_LOCK_ASSERT(td, MA_OWNED);
1571 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1573 ast_unsched_locked(td, TDA_SUSPEND);
1574 TD_SET_SUSPENDED(td);
1579 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1582 THREAD_LOCK_ASSERT(td, MA_OWNED);
1583 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1584 TD_CLR_SUSPENDED(td);
1585 td->td_flags &= ~TDF_ALLPROCSUSP;
1586 if (td->td_proc == p) {
1587 PROC_SLOCK_ASSERT(p, MA_OWNED);
1589 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1590 td->td_flags &= ~TDF_BOUNDARY;
1591 p->p_boundary_count--;
1594 return (setrunnable(td, 0));
1598 thread_run_flash(struct thread *td)
1603 PROC_LOCK_ASSERT(p, MA_OWNED);
1605 if (TD_ON_SLEEPQ(td))
1606 sleepq_remove_nested(td);
1610 THREAD_LOCK_ASSERT(td, MA_OWNED);
1611 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1613 TD_CLR_SUSPENDED(td);
1615 MPASS(p->p_suspcount > 0);
1618 if (setrunnable(td, 0))
1623 * Allow all threads blocked by single threading to continue running.
1626 thread_unsuspend(struct proc *p)
1631 PROC_LOCK_ASSERT(p, MA_OWNED);
1632 PROC_SLOCK_ASSERT(p, MA_OWNED);
1634 if (!P_SHOULDSTOP(p)) {
1635 FOREACH_THREAD_IN_PROC(p, td) {
1637 if (TD_IS_SUSPENDED(td))
1638 wakeup_swapper |= thread_unsuspend_one(td, p,
1643 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1644 p->p_numthreads == p->p_suspcount) {
1646 * Stopping everything also did the job for the single
1647 * threading request. Now we've downgraded to single-threaded,
1650 if (p->p_singlethread->td_proc == p) {
1651 thread_lock(p->p_singlethread);
1652 wakeup_swapper = thread_unsuspend_one(
1653 p->p_singlethread, p, false);
1661 * End the single threading mode..
1664 thread_single_end(struct proc *p, int mode)
1669 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1670 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1671 ("invalid mode %d", mode));
1672 PROC_LOCK_ASSERT(p, MA_OWNED);
1673 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1674 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1675 ("mode %d does not match P_TOTAL_STOP", mode));
1676 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1677 ("thread_single_end from other thread %p %p",
1678 curthread, p->p_singlethread));
1679 KASSERT(mode != SINGLE_BOUNDARY ||
1680 (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1681 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1682 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1685 p->p_singlethread = NULL;
1688 * If there are other threads they may now run,
1689 * unless of course there is a blanket 'stop order'
1690 * on the process. The single threader must be allowed
1691 * to continue however as this is a bad place to stop.
1693 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1694 FOREACH_THREAD_IN_PROC(p, td) {
1696 if (TD_IS_SUSPENDED(td)) {
1697 wakeup_swapper |= thread_unsuspend_one(td, p,
1703 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1704 ("inconsistent boundary count %d", p->p_boundary_count));
1712 * Locate a thread by number and return with proc lock held.
1714 * thread exit establishes proc -> tidhash lock ordering, but lookup
1715 * takes tidhash first and needs to return locked proc.
1717 * The problem is worked around by relying on type-safety of both
1718 * structures and doing the work in 2 steps:
1719 * - tidhash-locked lookup which saves both thread and proc pointers
1720 * - proc-locked verification that the found thread still matches
1723 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1725 #define RUN_THRESH 16
1732 rw_rlock(TIDHASHLOCK(tid));
1734 LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1735 if (td->td_tid != tid) {
1740 if (pid != -1 && p->p_pid != pid) {
1744 if (run > RUN_THRESH) {
1745 if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1746 LIST_REMOVE(td, td_hash);
1747 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1749 rw_wunlock(TIDHASHLOCK(tid));
1757 rw_runlock(TIDHASHLOCK(tid));
1766 tdfind(lwpid_t tid, pid_t pid)
1772 if (td->td_tid == tid) {
1773 if (pid != -1 && td->td_proc->p_pid != pid)
1775 PROC_LOCK(td->td_proc);
1780 if (!tdfind_hash(tid, pid, &p, &td))
1783 if (td->td_tid != tid) {
1787 if (td->td_proc != p) {
1791 if (p->p_state == PRS_NEW) {
1800 tidhash_add(struct thread *td)
1802 rw_wlock(TIDHASHLOCK(td->td_tid));
1803 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1804 rw_wunlock(TIDHASHLOCK(td->td_tid));
1808 tidhash_remove(struct thread *td)
1811 rw_wlock(TIDHASHLOCK(td->td_tid));
1812 LIST_REMOVE(td, td_hash);
1813 rw_wunlock(TIDHASHLOCK(td->td_tid));