2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice(s), this list of conditions and the following disclaimer as
12 * the first lines of this file unmodified other than the possible
13 * addition of one or more copyright notices.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice(s), this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
19 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
22 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
25 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
31 #include "opt_witness.h"
32 #include "opt_hwpmc_hooks.h"
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
41 #include <sys/mutex.h>
43 #include <sys/bitstring.h>
44 #include <sys/epoch.h>
45 #include <sys/rangelock.h>
46 #include <sys/resourcevar.h>
49 #include <sys/sched.h>
50 #include <sys/sleepqueue.h>
51 #include <sys/selinfo.h>
52 #include <sys/syscallsubr.h>
53 #include <sys/sysent.h>
54 #include <sys/turnstile.h>
56 #include <sys/rwlock.h>
58 #include <sys/vmmeter.h>
59 #include <sys/cpuset.h>
61 #include <sys/pmckern.h>
65 #include <security/audit/audit.h>
68 #include <vm/vm_extern.h>
70 #include <sys/eventhandler.h>
73 * Asserts below verify the stability of struct thread and struct proc
74 * layout, as exposed by KBI to modules. On head, the KBI is allowed
75 * to drift, change to the structures must be accompanied by the
78 * On the stable branches after KBI freeze, conditions must not be
79 * violated. Typically new fields are moved to the end of the
83 _Static_assert(offsetof(struct thread, td_flags) == 0xfc,
84 "struct thread KBI td_flags");
85 _Static_assert(offsetof(struct thread, td_pflags) == 0x104,
86 "struct thread KBI td_pflags");
87 _Static_assert(offsetof(struct thread, td_frame) == 0x4a0,
88 "struct thread KBI td_frame");
89 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6b0,
90 "struct thread KBI td_emuldata");
91 _Static_assert(offsetof(struct proc, p_flag) == 0xb0,
92 "struct proc KBI p_flag");
93 _Static_assert(offsetof(struct proc, p_pid) == 0xbc,
94 "struct proc KBI p_pid");
95 _Static_assert(offsetof(struct proc, p_filemon) == 0x3b8,
96 "struct proc KBI p_filemon");
97 _Static_assert(offsetof(struct proc, p_comm) == 0x3d0,
98 "struct proc KBI p_comm");
99 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4b0,
100 "struct proc KBI p_emuldata");
103 _Static_assert(offsetof(struct thread, td_flags) == 0x98,
104 "struct thread KBI td_flags");
105 _Static_assert(offsetof(struct thread, td_pflags) == 0xa0,
106 "struct thread KBI td_pflags");
107 _Static_assert(offsetof(struct thread, td_frame) == 0x300,
108 "struct thread KBI td_frame");
109 _Static_assert(offsetof(struct thread, td_emuldata) == 0x344,
110 "struct thread KBI td_emuldata");
111 _Static_assert(offsetof(struct proc, p_flag) == 0x68,
112 "struct proc KBI p_flag");
113 _Static_assert(offsetof(struct proc, p_pid) == 0x74,
114 "struct proc KBI p_pid");
115 _Static_assert(offsetof(struct proc, p_filemon) == 0x268,
116 "struct proc KBI p_filemon");
117 _Static_assert(offsetof(struct proc, p_comm) == 0x27c,
118 "struct proc KBI p_comm");
119 _Static_assert(offsetof(struct proc, p_emuldata) == 0x308,
120 "struct proc KBI p_emuldata");
123 SDT_PROVIDER_DECLARE(proc);
124 SDT_PROBE_DEFINE(proc, , , lwp__exit);
127 * thread related storage.
129 static uma_zone_t thread_zone;
131 static __exclusive_cache_line struct thread *thread_zombies;
133 static void thread_zombie(struct thread *);
134 static int thread_unsuspend_one(struct thread *td, struct proc *p,
137 static struct mtx tid_lock;
138 static bitstr_t *tid_bitmap;
140 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
142 static int maxthread;
143 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
144 &maxthread, 0, "Maximum number of threads");
148 static LIST_HEAD(tidhashhead, thread) *tidhashtbl;
149 static u_long tidhash;
150 static u_long tidhashlock;
151 static struct rwlock *tidhashtbl_lock;
152 #define TIDHASH(tid) (&tidhashtbl[(tid) & tidhash])
153 #define TIDHASHLOCK(tid) (&tidhashtbl_lock[(tid) & tidhashlock])
155 EVENTHANDLER_LIST_DEFINE(thread_ctor);
156 EVENTHANDLER_LIST_DEFINE(thread_dtor);
157 EVENTHANDLER_LIST_DEFINE(thread_init);
158 EVENTHANDLER_LIST_DEFINE(thread_fini);
163 static struct timeval lastfail;
165 static lwpid_t trytid;
169 if (nthreads + 1 >= maxthread - 100) {
170 if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
171 nthreads + 1 >= maxthread) {
172 mtx_unlock(&tid_lock);
173 if (ppsratecheck(&lastfail, &curfail, 1)) {
174 printf("maxthread limit exceeded by uid %u "
175 "(pid %d); consider increasing kern.maxthread\n",
176 curthread->td_ucred->cr_ruid, curproc->p_pid);
184 * It is an invariant that the bitmap is big enough to hold maxthread
185 * IDs. If we got to this point there has to be at least one free.
187 if (trytid >= maxthread)
189 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
191 KASSERT(trytid != 0, ("unexpectedly ran out of IDs"));
193 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
194 KASSERT(tid != -1, ("unexpectedly ran out of IDs"));
196 bit_set(tid_bitmap, tid);
198 mtx_unlock(&tid_lock);
199 return (tid + NO_PID);
203 tid_free(lwpid_t rtid)
207 KASSERT(rtid >= NO_PID,
208 ("%s: invalid tid %d\n", __func__, rtid));
211 KASSERT(bit_test(tid_bitmap, tid) != 0,
212 ("thread ID %d not allocated\n", rtid));
213 bit_clear(tid_bitmap, tid);
215 mtx_unlock(&tid_lock);
219 * Prepare a thread for use.
222 thread_ctor(void *mem, int size, void *arg, int flags)
226 td = (struct thread *)mem;
227 td->td_state = TDS_INACTIVE;
228 td->td_lastcpu = td->td_oncpu = NOCPU;
231 * Note that td_critnest begins life as 1 because the thread is not
232 * running and is thereby implicitly waiting to be on the receiving
233 * end of a context switch.
236 td->td_lend_user_pri = PRI_MAX;
238 audit_thread_alloc(td);
240 umtx_thread_alloc(td);
245 * Reclaim a thread after use.
248 thread_dtor(void *mem, int size, void *arg)
252 td = (struct thread *)mem;
255 /* Verify that this thread is in a safe state to free. */
256 switch (td->td_state) {
262 * We must never unlink a thread that is in one of
263 * these states, because it is currently active.
265 panic("bad state for thread unlinking");
270 panic("bad thread state");
275 audit_thread_free(td);
277 /* Free all OSD associated to this thread. */
279 td_softdep_cleanup(td);
280 MPASS(td->td_su == NULL);
284 * Initialize type-stable parts of a thread (when newly created).
287 thread_init(void *mem, int size, int flags)
291 td = (struct thread *)mem;
293 td->td_sleepqueue = sleepq_alloc();
294 td->td_turnstile = turnstile_alloc();
296 EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
297 umtx_thread_init(td);
304 * Tear down type-stable parts of a thread (just before being discarded).
307 thread_fini(void *mem, int size)
311 td = (struct thread *)mem;
312 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
313 rlqentry_free(td->td_rlqe);
314 turnstile_free(td->td_turnstile);
315 sleepq_free(td->td_sleepqueue);
316 umtx_thread_fini(td);
321 * For a newly created process,
322 * link up all the structures and its initial threads etc.
324 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc.
325 * proc_dtor() (should go away)
329 proc_linkup0(struct proc *p, struct thread *td)
331 TAILQ_INIT(&p->p_threads); /* all threads in proc */
336 proc_linkup(struct proc *p, struct thread *td)
339 sigqueue_init(&p->p_sigqueue, p);
340 p->p_ksi = ksiginfo_alloc(1);
341 if (p->p_ksi != NULL) {
342 /* XXX p_ksi may be null if ksiginfo zone is not ready */
343 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
345 LIST_INIT(&p->p_mqnotifier);
350 extern int max_threads_per_proc;
353 * Initialize global thread allocation resources.
363 * Place an upper limit on threads which can be allocated.
365 * Note that other factors may make the de facto limit much lower.
367 * Platform limits are somewhat arbitrary but deemed "more than good
368 * enough" for the foreseable future.
370 if (maxthread == 0) {
372 maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
374 maxthread = MIN(maxproc * max_threads_per_proc, 100000);
378 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
379 tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
381 if (tid0 != THREAD0_TID)
382 panic("tid0 %d != %d\n", tid0, THREAD0_TID);
384 flags = UMA_ZONE_NOFREE;
387 * Force thread structures to be allocated from the direct map.
388 * Otherwise, superpage promotions and demotions may temporarily
389 * invalidate thread structure mappings. For most dynamically allocated
390 * structures this is not a problem, but translation faults cannot be
391 * handled without accessing curthread.
393 flags |= UMA_ZONE_CONTIG;
395 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
396 thread_ctor, thread_dtor, thread_init, thread_fini,
398 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
399 tidhashlock = (tidhash + 1) / 64;
402 tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
403 M_TIDHASH, M_WAITOK | M_ZERO);
404 for (i = 0; i < tidhashlock + 1; i++)
405 rw_init(&tidhashtbl_lock[i], "tidhash");
409 * Place an unused thread on the zombie list.
412 thread_zombie(struct thread *td)
416 ztd = atomic_load_ptr(&thread_zombies);
419 if (atomic_fcmpset_rel_ptr((uintptr_t *)&thread_zombies,
420 (uintptr_t *)&ztd, (uintptr_t)td))
427 * Release a thread that has exited after cpu_throw().
430 thread_stash(struct thread *td)
432 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
437 * Reap zombie threads.
442 struct thread *itd, *ntd;
445 * Reading upfront is pessimal if followed by concurrent atomic_swap,
446 * but most of the time the list is empty.
448 if (thread_zombies == NULL)
451 itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&thread_zombies,
453 while (itd != NULL) {
454 ntd = itd->td_zombie;
455 thread_cow_free(itd);
465 thread_alloc(int pages)
470 thread_reap(); /* check if any zombies to get */
477 td = uma_zalloc(thread_zone, M_WAITOK);
478 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
479 if (!vm_thread_new(td, pages)) {
480 uma_zfree(thread_zone, td);
485 cpu_thread_alloc(td);
486 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
491 thread_alloc_stack(struct thread *td, int pages)
494 KASSERT(td->td_kstack == 0,
495 ("thread_alloc_stack called on a thread with kstack"));
496 if (!vm_thread_new(td, pages))
498 cpu_thread_alloc(td);
503 * Deallocate a thread.
506 thread_free(struct thread *td)
509 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
510 lock_profile_thread_exit(td);
512 cpuset_rel(td->td_cpuset);
513 td->td_cpuset = NULL;
515 if (td->td_kstack != 0)
516 vm_thread_dispose(td);
517 callout_drain(&td->td_slpcallout);
518 tid_free(td->td_tid);
520 uma_zfree(thread_zone, td);
524 thread_cow_get_proc(struct thread *newtd, struct proc *p)
527 PROC_LOCK_ASSERT(p, MA_OWNED);
528 newtd->td_realucred = crcowget(p->p_ucred);
529 newtd->td_ucred = newtd->td_realucred;
530 newtd->td_limit = lim_hold(p->p_limit);
531 newtd->td_cowgen = p->p_cowgen;
535 thread_cow_get(struct thread *newtd, struct thread *td)
538 MPASS(td->td_realucred == td->td_ucred);
539 newtd->td_realucred = crcowget(td->td_realucred);
540 newtd->td_ucred = newtd->td_realucred;
541 newtd->td_limit = lim_hold(td->td_limit);
542 newtd->td_cowgen = td->td_cowgen;
546 thread_cow_free(struct thread *td)
549 if (td->td_realucred != NULL)
551 if (td->td_limit != NULL)
552 lim_free(td->td_limit);
556 thread_cow_update(struct thread *td)
559 struct ucred *oldcred;
560 struct plimit *oldlimit;
565 oldcred = crcowsync();
566 if (td->td_limit != p->p_limit) {
567 oldlimit = td->td_limit;
568 td->td_limit = lim_hold(p->p_limit);
570 td->td_cowgen = p->p_cowgen;
574 if (oldlimit != NULL)
579 * Discard the current thread and exit from its context.
580 * Always called with scheduler locked.
582 * Because we can't free a thread while we're operating under its context,
583 * push the current thread into our CPU's deadthread holder. This means
584 * we needn't worry about someone else grabbing our context before we
590 uint64_t runtime, new_switchtime;
599 PROC_SLOCK_ASSERT(p, MA_OWNED);
600 mtx_assert(&Giant, MA_NOTOWNED);
602 PROC_LOCK_ASSERT(p, MA_OWNED);
603 KASSERT(p != NULL, ("thread exiting without a process"));
604 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
605 (long)p->p_pid, td->td_name);
606 SDT_PROBE0(proc, , , lwp__exit);
607 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
608 MPASS(td->td_realucred == td->td_ucred);
611 * drop FPU & debug register state storage, or any other
612 * architecture specific resources that
613 * would not be on a new untouched process.
618 * The last thread is left attached to the process
619 * So that the whole bundle gets recycled. Skip
620 * all this stuff if we never had threads.
621 * EXIT clears all sign of other threads when
622 * it goes to single threading, so the last thread always
623 * takes the short path.
625 if (p->p_flag & P_HADTHREADS) {
626 if (p->p_numthreads > 1) {
627 atomic_add_int(&td->td_proc->p_exitthreads, 1);
629 td2 = FIRST_THREAD_IN_PROC(p);
630 sched_exit_thread(td2, td);
633 * The test below is NOT true if we are the
634 * sole exiting thread. P_STOPPED_SINGLE is unset
635 * in exit1() after it is the only survivor.
637 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
638 if (p->p_numthreads == p->p_suspcount) {
639 thread_lock(p->p_singlethread);
640 wakeup_swapper = thread_unsuspend_one(
641 p->p_singlethread, p, false);
647 PCPU_SET(deadthread, td);
650 * The last thread is exiting.. but not through exit()
652 panic ("thread_exit: Last thread exiting on its own");
657 * If this thread is part of a process that is being tracked by hwpmc(4),
658 * inform the module of the thread's impending exit.
660 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
661 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
662 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
663 } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
664 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
671 /* Do the same timestamp bookkeeping that mi_switch() would do. */
672 new_switchtime = cpu_ticks();
673 runtime = new_switchtime - PCPU_GET(switchtime);
674 td->td_runtime += runtime;
675 td->td_incruntime += runtime;
676 PCPU_SET(switchtime, new_switchtime);
677 PCPU_SET(switchticks, ticks);
680 /* Save our resource usage in our process. */
681 td->td_ru.ru_nvcsw++;
682 ruxagg_locked(p, td);
683 rucollect(&p->p_ru, &td->td_ru);
686 td->td_state = TDS_INACTIVE;
688 witness_thread_exit(td);
690 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
692 panic("I'm a teapot!");
697 * Do any thread specific cleanups that may be needed in wait()
698 * called with Giant, proc and schedlock not held.
701 thread_wait(struct proc *p)
705 mtx_assert(&Giant, MA_NOTOWNED);
706 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
707 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
708 td = FIRST_THREAD_IN_PROC(p);
709 /* Lock the last thread so we spin until it exits cpu_throw(). */
712 lock_profile_thread_exit(td);
713 cpuset_rel(td->td_cpuset);
714 td->td_cpuset = NULL;
715 cpu_thread_clean(td);
717 callout_drain(&td->td_slpcallout);
718 thread_reap(); /* check for zombie threads etc. */
722 * Link a thread to a process.
723 * set up anything that needs to be initialized for it to
724 * be used by the process.
727 thread_link(struct thread *td, struct proc *p)
731 * XXX This can't be enabled because it's called for proc0 before
732 * its lock has been created.
733 * PROC_LOCK_ASSERT(p, MA_OWNED);
735 td->td_state = TDS_INACTIVE;
737 td->td_flags = TDF_INMEM;
739 LIST_INIT(&td->td_contested);
740 LIST_INIT(&td->td_lprof[0]);
741 LIST_INIT(&td->td_lprof[1]);
743 SLIST_INIT(&td->td_epochs);
745 sigqueue_init(&td->td_sigqueue, p);
746 callout_init(&td->td_slpcallout, 1);
747 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
756 thread_unlink(struct thread *td)
758 struct proc *p = td->td_proc;
760 PROC_LOCK_ASSERT(p, MA_OWNED);
762 MPASS(SLIST_EMPTY(&td->td_epochs));
765 TAILQ_REMOVE(&p->p_threads, td, td_plist);
767 /* could clear a few other things here */
768 /* Must NOT clear links to proc! */
772 calc_remaining(struct proc *p, int mode)
776 PROC_LOCK_ASSERT(p, MA_OWNED);
777 PROC_SLOCK_ASSERT(p, MA_OWNED);
778 if (mode == SINGLE_EXIT)
779 remaining = p->p_numthreads;
780 else if (mode == SINGLE_BOUNDARY)
781 remaining = p->p_numthreads - p->p_boundary_count;
782 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
783 remaining = p->p_numthreads - p->p_suspcount;
785 panic("calc_remaining: wrong mode %d", mode);
790 remain_for_mode(int mode)
793 return (mode == SINGLE_ALLPROC ? 0 : 1);
797 weed_inhib(int mode, struct thread *td2, struct proc *p)
801 PROC_LOCK_ASSERT(p, MA_OWNED);
802 PROC_SLOCK_ASSERT(p, MA_OWNED);
803 THREAD_LOCK_ASSERT(td2, MA_OWNED);
808 * Since the thread lock is dropped by the scheduler we have
809 * to retry to check for races.
814 if (TD_IS_SUSPENDED(td2)) {
815 wakeup_swapper |= thread_unsuspend_one(td2, p, true);
819 if (TD_CAN_ABORT(td2)) {
820 wakeup_swapper |= sleepq_abort(td2, EINTR);
821 return (wakeup_swapper);
824 case SINGLE_BOUNDARY:
826 if (TD_IS_SUSPENDED(td2) &&
827 (td2->td_flags & TDF_BOUNDARY) == 0) {
828 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
832 if (TD_CAN_ABORT(td2)) {
833 wakeup_swapper |= sleepq_abort(td2, ERESTART);
834 return (wakeup_swapper);
839 * ALLPROC suspend tries to avoid spurious EINTR for
840 * threads sleeping interruptable, by suspending the
841 * thread directly, similarly to sig_suspend_threads().
842 * Since such sleep is not performed at the user
843 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
844 * is used to avoid immediate un-suspend.
846 if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
847 TDF_ALLPROCSUSP)) == 0) {
848 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
852 if (TD_CAN_ABORT(td2)) {
853 if ((td2->td_flags & TDF_SBDRY) == 0) {
854 thread_suspend_one(td2);
855 td2->td_flags |= TDF_ALLPROCSUSP;
857 wakeup_swapper |= sleepq_abort(td2, ERESTART);
858 return (wakeup_swapper);
866 return (wakeup_swapper);
870 * Enforce single-threading.
872 * Returns 1 if the caller must abort (another thread is waiting to
873 * exit the process or similar). Process is locked!
874 * Returns 0 when you are successfully the only thread running.
875 * A process has successfully single threaded in the suspend mode when
876 * There are no threads in user mode. Threads in the kernel must be
877 * allowed to continue until they get to the user boundary. They may even
878 * copy out their return values and data before suspending. They may however be
879 * accelerated in reaching the user boundary as we will wake up
880 * any sleeping threads that are interruptable. (PCATCH).
883 thread_single(struct proc *p, int mode)
887 int remaining, wakeup_swapper;
890 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
891 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
892 ("invalid mode %d", mode));
894 * If allowing non-ALLPROC singlethreading for non-curproc
895 * callers, calc_remaining() and remain_for_mode() should be
896 * adjusted to also account for td->td_proc != p. For now
897 * this is not implemented because it is not used.
899 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
900 (mode != SINGLE_ALLPROC && td->td_proc == p),
901 ("mode %d proc %p curproc %p", mode, p, td->td_proc));
902 mtx_assert(&Giant, MA_NOTOWNED);
903 PROC_LOCK_ASSERT(p, MA_OWNED);
905 if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
908 /* Is someone already single threading? */
909 if (p->p_singlethread != NULL && p->p_singlethread != td)
912 if (mode == SINGLE_EXIT) {
913 p->p_flag |= P_SINGLE_EXIT;
914 p->p_flag &= ~P_SINGLE_BOUNDARY;
916 p->p_flag &= ~P_SINGLE_EXIT;
917 if (mode == SINGLE_BOUNDARY)
918 p->p_flag |= P_SINGLE_BOUNDARY;
920 p->p_flag &= ~P_SINGLE_BOUNDARY;
922 if (mode == SINGLE_ALLPROC)
923 p->p_flag |= P_TOTAL_STOP;
924 p->p_flag |= P_STOPPED_SINGLE;
926 p->p_singlethread = td;
927 remaining = calc_remaining(p, mode);
928 while (remaining != remain_for_mode(mode)) {
929 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
932 FOREACH_THREAD_IN_PROC(p, td2) {
936 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
937 if (TD_IS_INHIBITED(td2)) {
938 wakeup_swapper |= weed_inhib(mode, td2, p);
940 } else if (TD_IS_RUNNING(td2) && td != td2) {
949 remaining = calc_remaining(p, mode);
952 * Maybe we suspended some threads.. was it enough?
954 if (remaining == remain_for_mode(mode))
959 * Wake us up when everyone else has suspended.
960 * In the mean time we suspend as well.
962 thread_suspend_switch(td, p);
963 remaining = calc_remaining(p, mode);
965 if (mode == SINGLE_EXIT) {
967 * Convert the process to an unthreaded process. The
968 * SINGLE_EXIT is called by exit1() or execve(), in
969 * both cases other threads must be retired.
971 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
972 p->p_singlethread = NULL;
973 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
976 * Wait for any remaining threads to exit cpu_throw().
978 while (p->p_exitthreads != 0) {
981 sched_relinquish(td);
985 } else if (mode == SINGLE_BOUNDARY) {
987 * Wait until all suspended threads are removed from
988 * the processors. The thread_suspend_check()
989 * increments p_boundary_count while it is still
990 * running, which makes it possible for the execve()
991 * to destroy vmspace while our other threads are
992 * still using the address space.
994 * We lock the thread, which is only allowed to
995 * succeed after context switch code finished using
998 FOREACH_THREAD_IN_PROC(p, td2) {
1002 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1003 ("td %p not on boundary", td2));
1004 KASSERT(TD_IS_SUSPENDED(td2),
1005 ("td %p is not suspended", td2));
1014 thread_suspend_check_needed(void)
1021 PROC_LOCK_ASSERT(p, MA_OWNED);
1022 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1023 (td->td_dbgflags & TDB_SUSPEND) != 0));
1027 * Called in from locations that can safely check to see
1028 * whether we have to suspend or at least throttle for a
1029 * single-thread event (e.g. fork).
1031 * Such locations include userret().
1032 * If the "return_instead" argument is non zero, the thread must be able to
1033 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1035 * The 'return_instead' argument tells the function if it may do a
1036 * thread_exit() or suspend, or whether the caller must abort and back
1039 * If the thread that set the single_threading request has set the
1040 * P_SINGLE_EXIT bit in the process flags then this call will never return
1041 * if 'return_instead' is false, but will exit.
1043 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1044 *---------------+--------------------+---------------------
1045 * 0 | returns 0 | returns 0 or 1
1046 * | when ST ends | immediately
1047 *---------------+--------------------+---------------------
1048 * 1 | thread exits | returns 1
1050 * 0 = thread_exit() or suspension ok,
1051 * other = return error instead of stopping the thread.
1053 * While a full suspension is under effect, even a single threading
1054 * thread would be suspended if it made this call (but it shouldn't).
1055 * This call should only be made from places where
1056 * thread_exit() would be safe as that may be the outcome unless
1057 * return_instead is set.
1060 thread_suspend_check(int return_instead)
1068 mtx_assert(&Giant, MA_NOTOWNED);
1069 PROC_LOCK_ASSERT(p, MA_OWNED);
1070 while (thread_suspend_check_needed()) {
1071 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1072 KASSERT(p->p_singlethread != NULL,
1073 ("singlethread not set"));
1075 * The only suspension in action is a
1076 * single-threading. Single threader need not stop.
1077 * It is safe to access p->p_singlethread unlocked
1078 * because it can only be set to our address by us.
1080 if (p->p_singlethread == td)
1081 return (0); /* Exempt from stopping. */
1083 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1086 /* Should we goto user boundary if we didn't come from there? */
1087 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1088 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1092 * Ignore suspend requests if they are deferred.
1094 if ((td->td_flags & TDF_SBDRY) != 0) {
1095 KASSERT(return_instead,
1096 ("TDF_SBDRY set for unsafe thread_suspend_check"));
1097 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1098 (TDF_SEINTR | TDF_SERESTART),
1099 ("both TDF_SEINTR and TDF_SERESTART"));
1100 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1104 * If the process is waiting for us to exit,
1105 * this thread should just suicide.
1106 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1108 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1112 * Allow Linux emulation layer to do some work
1113 * before thread suicide.
1115 if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1116 (p->p_sysent->sv_thread_detach)(td);
1117 umtx_thread_exit(td);
1119 panic("stopped thread did not exit");
1124 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1125 if (p->p_numthreads == p->p_suspcount + 1) {
1126 thread_lock(p->p_singlethread);
1127 wakeup_swapper = thread_unsuspend_one(
1128 p->p_singlethread, p, false);
1136 * When a thread suspends, it just
1137 * gets taken off all queues.
1139 thread_suspend_one(td);
1140 if (return_instead == 0) {
1141 p->p_boundary_count++;
1142 td->td_flags |= TDF_BOUNDARY;
1145 mi_switch(SW_INVOL | SWT_SUSPEND);
1152 * Check for possible stops and suspensions while executing a
1153 * casueword or similar transiently failing operation.
1155 * The sleep argument controls whether the function can handle a stop
1156 * request itself or it should return ERESTART and the request is
1157 * proceed at the kernel/user boundary in ast.
1159 * Typically, when retrying due to casueword(9) failure (rv == 1), we
1160 * should handle the stop requests there, with exception of cases when
1161 * the thread owns a kernel resource, for instance busied the umtx
1162 * key, or when functions return immediately if thread_check_susp()
1163 * returned non-zero. On the other hand, retrying the whole lock
1164 * operation, we better not stop there but delegate the handling to
1167 * If the request is for thread termination P_SINGLE_EXIT, we cannot
1168 * handle it at all, and simply return EINTR.
1171 thread_check_susp(struct thread *td, bool sleep)
1177 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
1178 * eventually break the lockstep loop.
1180 if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
1185 if (p->p_flag & P_SINGLE_EXIT)
1187 else if (P_SHOULDSTOP(p) ||
1188 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1189 error = sleep ? thread_suspend_check(0) : ERESTART;
1195 thread_suspend_switch(struct thread *td, struct proc *p)
1198 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1199 PROC_LOCK_ASSERT(p, MA_OWNED);
1200 PROC_SLOCK_ASSERT(p, MA_OWNED);
1202 * We implement thread_suspend_one in stages here to avoid
1203 * dropping the proc lock while the thread lock is owned.
1205 if (p == td->td_proc) {
1211 td->td_flags &= ~TDF_NEEDSUSPCHK;
1212 TD_SET_SUSPENDED(td);
1216 mi_switch(SW_VOL | SWT_SUSPEND);
1223 thread_suspend_one(struct thread *td)
1228 PROC_SLOCK_ASSERT(p, MA_OWNED);
1229 THREAD_LOCK_ASSERT(td, MA_OWNED);
1230 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1232 td->td_flags &= ~TDF_NEEDSUSPCHK;
1233 TD_SET_SUSPENDED(td);
1238 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1241 THREAD_LOCK_ASSERT(td, MA_OWNED);
1242 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1243 TD_CLR_SUSPENDED(td);
1244 td->td_flags &= ~TDF_ALLPROCSUSP;
1245 if (td->td_proc == p) {
1246 PROC_SLOCK_ASSERT(p, MA_OWNED);
1248 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1249 td->td_flags &= ~TDF_BOUNDARY;
1250 p->p_boundary_count--;
1253 return (setrunnable(td, 0));
1257 * Allow all threads blocked by single threading to continue running.
1260 thread_unsuspend(struct proc *p)
1265 PROC_LOCK_ASSERT(p, MA_OWNED);
1266 PROC_SLOCK_ASSERT(p, MA_OWNED);
1268 if (!P_SHOULDSTOP(p)) {
1269 FOREACH_THREAD_IN_PROC(p, td) {
1271 if (TD_IS_SUSPENDED(td)) {
1272 wakeup_swapper |= thread_unsuspend_one(td, p,
1277 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1278 p->p_numthreads == p->p_suspcount) {
1280 * Stopping everything also did the job for the single
1281 * threading request. Now we've downgraded to single-threaded,
1284 if (p->p_singlethread->td_proc == p) {
1285 thread_lock(p->p_singlethread);
1286 wakeup_swapper = thread_unsuspend_one(
1287 p->p_singlethread, p, false);
1295 * End the single threading mode..
1298 thread_single_end(struct proc *p, int mode)
1303 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1304 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1305 ("invalid mode %d", mode));
1306 PROC_LOCK_ASSERT(p, MA_OWNED);
1307 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1308 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1309 ("mode %d does not match P_TOTAL_STOP", mode));
1310 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1311 ("thread_single_end from other thread %p %p",
1312 curthread, p->p_singlethread));
1313 KASSERT(mode != SINGLE_BOUNDARY ||
1314 (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1315 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1316 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1319 p->p_singlethread = NULL;
1322 * If there are other threads they may now run,
1323 * unless of course there is a blanket 'stop order'
1324 * on the process. The single threader must be allowed
1325 * to continue however as this is a bad place to stop.
1327 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1328 FOREACH_THREAD_IN_PROC(p, td) {
1330 if (TD_IS_SUSPENDED(td)) {
1331 wakeup_swapper |= thread_unsuspend_one(td, p,
1332 mode == SINGLE_BOUNDARY);
1337 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1338 ("inconsistent boundary count %d", p->p_boundary_count));
1345 * Locate a thread by number and return with proc lock held.
1347 * thread exit establishes proc -> tidhash lock ordering, but lookup
1348 * takes tidhash first and needs to return locked proc.
1350 * The problem is worked around by relying on type-safety of both
1351 * structures and doing the work in 2 steps:
1352 * - tidhash-locked lookup which saves both thread and proc pointers
1353 * - proc-locked verification that the found thread still matches
1356 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1358 #define RUN_THRESH 16
1365 rw_rlock(TIDHASHLOCK(tid));
1367 LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1368 if (td->td_tid != tid) {
1373 if (pid != -1 && p->p_pid != pid) {
1377 if (run > RUN_THRESH) {
1378 if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1379 LIST_REMOVE(td, td_hash);
1380 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1382 rw_wunlock(TIDHASHLOCK(tid));
1390 rw_runlock(TIDHASHLOCK(tid));
1399 tdfind(lwpid_t tid, pid_t pid)
1405 if (td->td_tid == tid) {
1406 if (pid != -1 && td->td_proc->p_pid != pid)
1408 PROC_LOCK(td->td_proc);
1413 if (!tdfind_hash(tid, pid, &p, &td))
1416 if (td->td_tid != tid) {
1420 if (td->td_proc != p) {
1424 if (p->p_state == PRS_NEW) {
1433 tidhash_add(struct thread *td)
1435 rw_wlock(TIDHASHLOCK(td->td_tid));
1436 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1437 rw_wunlock(TIDHASHLOCK(td->td_tid));
1441 tidhash_remove(struct thread *td)
1444 rw_wlock(TIDHASHLOCK(td->td_tid));
1445 LIST_REMOVE(td, td_hash);
1446 rw_wunlock(TIDHASHLOCK(td->td_tid));