2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice(s), this list of conditions and the following disclaimer as
12 * the first lines of this file unmodified other than the possible
13 * addition of one or more copyright notices.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice(s), this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
19 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
22 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
25 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
31 #include "opt_witness.h"
32 #include "opt_hwpmc_hooks.h"
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
41 #include <sys/mutex.h>
43 #include <sys/epoch.h>
44 #include <sys/rangelock.h>
45 #include <sys/resourcevar.h>
48 #include <sys/sched.h>
49 #include <sys/sleepqueue.h>
50 #include <sys/selinfo.h>
51 #include <sys/syscallsubr.h>
52 #include <sys/sysent.h>
53 #include <sys/turnstile.h>
55 #include <sys/rwlock.h>
57 #include <sys/vmmeter.h>
58 #include <sys/cpuset.h>
60 #include <sys/pmckern.h>
64 #include <security/audit/audit.h>
67 #include <vm/vm_extern.h>
69 #include <sys/eventhandler.h>
72 * Asserts below verify the stability of struct thread and struct proc
73 * layout, as exposed by KBI to modules. On head, the KBI is allowed
74 * to drift, change to the structures must be accompanied by the
77 * On the stable branches after KBI freeze, conditions must not be
78 * violated. Typically new fields are moved to the end of the
82 _Static_assert(offsetof(struct thread, td_flags) == 0xfc,
83 "struct thread KBI td_flags");
84 _Static_assert(offsetof(struct thread, td_pflags) == 0x104,
85 "struct thread KBI td_pflags");
86 _Static_assert(offsetof(struct thread, td_frame) == 0x4a0,
87 "struct thread KBI td_frame");
88 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6b0,
89 "struct thread KBI td_emuldata");
90 _Static_assert(offsetof(struct proc, p_flag) == 0xb0,
91 "struct proc KBI p_flag");
92 _Static_assert(offsetof(struct proc, p_pid) == 0xbc,
93 "struct proc KBI p_pid");
94 _Static_assert(offsetof(struct proc, p_filemon) == 0x3b8,
95 "struct proc KBI p_filemon");
96 _Static_assert(offsetof(struct proc, p_comm) == 0x3d0,
97 "struct proc KBI p_comm");
98 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4b0,
99 "struct proc KBI p_emuldata");
102 _Static_assert(offsetof(struct thread, td_flags) == 0x98,
103 "struct thread KBI td_flags");
104 _Static_assert(offsetof(struct thread, td_pflags) == 0xa0,
105 "struct thread KBI td_pflags");
106 _Static_assert(offsetof(struct thread, td_frame) == 0x300,
107 "struct thread KBI td_frame");
108 _Static_assert(offsetof(struct thread, td_emuldata) == 0x344,
109 "struct thread KBI td_emuldata");
110 _Static_assert(offsetof(struct proc, p_flag) == 0x68,
111 "struct proc KBI p_flag");
112 _Static_assert(offsetof(struct proc, p_pid) == 0x74,
113 "struct proc KBI p_pid");
114 _Static_assert(offsetof(struct proc, p_filemon) == 0x268,
115 "struct proc KBI p_filemon");
116 _Static_assert(offsetof(struct proc, p_comm) == 0x27c,
117 "struct proc KBI p_comm");
118 _Static_assert(offsetof(struct proc, p_emuldata) == 0x308,
119 "struct proc KBI p_emuldata");
122 SDT_PROVIDER_DECLARE(proc);
123 SDT_PROBE_DEFINE(proc, , , lwp__exit);
126 * thread related storage.
128 static uma_zone_t thread_zone;
130 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
131 static struct mtx zombie_lock;
132 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
134 static void thread_zombie(struct thread *);
135 static int thread_unsuspend_one(struct thread *td, struct proc *p,
138 #define TID_BUFFER_SIZE 1024
141 static struct unrhdr *tid_unrhdr;
142 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
143 static int tid_head, tid_tail;
144 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
146 static int maxthread;
147 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
148 &maxthread, 0, "Maximum number of threads");
152 struct tidhashhead *tidhashtbl;
154 struct rwlock tidhash_lock;
156 EVENTHANDLER_LIST_DEFINE(thread_ctor);
157 EVENTHANDLER_LIST_DEFINE(thread_dtor);
158 EVENTHANDLER_LIST_DEFINE(thread_init);
159 EVENTHANDLER_LIST_DEFINE(thread_fini);
164 static struct timeval lastfail;
169 nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1;
170 if (nthreads_new >= maxthread - 100) {
171 if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
172 nthreads_new >= maxthread) {
173 atomic_subtract_int(&nthreads, 1);
174 if (ppsratecheck(&lastfail, &curfail, 1)) {
175 printf("maxthread limit exceeded by uid %u "
176 "(pid %d); consider increasing kern.maxthread\n",
177 curthread->td_ucred->cr_ruid, curproc->p_pid);
183 tid = alloc_unr(tid_unrhdr);
187 if (tid_head == tid_tail) {
188 mtx_unlock(&tid_lock);
191 tid = tid_buffer[tid_head];
192 tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
193 mtx_unlock(&tid_lock);
198 tid_free(lwpid_t tid)
200 lwpid_t tmp_tid = -1;
203 if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
204 tmp_tid = tid_buffer[tid_head];
205 tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
207 tid_buffer[tid_tail] = tid;
208 tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE;
209 mtx_unlock(&tid_lock);
211 free_unr(tid_unrhdr, tmp_tid);
212 atomic_subtract_int(&nthreads, 1);
216 * Prepare a thread for use.
219 thread_ctor(void *mem, int size, void *arg, int flags)
223 td = (struct thread *)mem;
224 td->td_state = TDS_INACTIVE;
225 td->td_lastcpu = td->td_oncpu = NOCPU;
228 * Note that td_critnest begins life as 1 because the thread is not
229 * running and is thereby implicitly waiting to be on the receiving
230 * end of a context switch.
233 td->td_lend_user_pri = PRI_MAX;
235 audit_thread_alloc(td);
237 umtx_thread_alloc(td);
242 * Reclaim a thread after use.
245 thread_dtor(void *mem, int size, void *arg)
249 td = (struct thread *)mem;
252 /* Verify that this thread is in a safe state to free. */
253 switch (td->td_state) {
259 * We must never unlink a thread that is in one of
260 * these states, because it is currently active.
262 panic("bad state for thread unlinking");
267 panic("bad thread state");
272 audit_thread_free(td);
274 /* Free all OSD associated to this thread. */
276 td_softdep_cleanup(td);
277 MPASS(td->td_su == NULL);
281 * Initialize type-stable parts of a thread (when newly created).
284 thread_init(void *mem, int size, int flags)
288 td = (struct thread *)mem;
290 td->td_sleepqueue = sleepq_alloc();
291 td->td_turnstile = turnstile_alloc();
293 EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
294 umtx_thread_init(td);
301 * Tear down type-stable parts of a thread (just before being discarded).
304 thread_fini(void *mem, int size)
308 td = (struct thread *)mem;
309 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
310 rlqentry_free(td->td_rlqe);
311 turnstile_free(td->td_turnstile);
312 sleepq_free(td->td_sleepqueue);
313 umtx_thread_fini(td);
318 * For a newly created process,
319 * link up all the structures and its initial threads etc.
321 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc.
322 * proc_dtor() (should go away)
326 proc_linkup0(struct proc *p, struct thread *td)
328 TAILQ_INIT(&p->p_threads); /* all threads in proc */
333 proc_linkup(struct proc *p, struct thread *td)
336 sigqueue_init(&p->p_sigqueue, p);
337 p->p_ksi = ksiginfo_alloc(1);
338 if (p->p_ksi != NULL) {
339 /* XXX p_ksi may be null if ksiginfo zone is not ready */
340 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
342 LIST_INIT(&p->p_mqnotifier);
347 extern int max_threads_per_proc;
350 * Initialize global thread allocation resources.
358 * Place an upper limit on threads which can be allocated.
360 * Note that other factors may make the de facto limit much lower.
362 * Platform limits are somewhat arbitrary but deemed "more than good
363 * enough" for the foreseable future.
365 if (maxthread == 0) {
367 maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
369 maxthread = MIN(maxproc * max_threads_per_proc, 100000);
373 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
376 * pid_max cannot be greater than PID_MAX.
377 * leave one number for thread0.
379 tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
381 flags = UMA_ZONE_NOFREE;
384 * Force thread structures to be allocated from the direct map.
385 * Otherwise, superpage promotions and demotions may temporarily
386 * invalidate thread structure mappings. For most dynamically allocated
387 * structures this is not a problem, but translation faults cannot be
388 * handled without accessing curthread.
390 flags |= UMA_ZONE_CONTIG;
392 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
393 thread_ctor, thread_dtor, thread_init, thread_fini,
395 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
396 rw_init(&tidhash_lock, "tidhash");
400 * Place an unused thread on the zombie list.
401 * Use the slpq as that must be unused by now.
404 thread_zombie(struct thread *td)
406 mtx_lock_spin(&zombie_lock);
407 TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
408 mtx_unlock_spin(&zombie_lock);
412 * Release a thread that has exited after cpu_throw().
415 thread_stash(struct thread *td)
417 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
422 * Reap zombie resources.
427 struct thread *td_first, *td_next;
430 * Don't even bother to lock if none at this instant,
431 * we really don't care about the next instant.
433 if (!TAILQ_EMPTY(&zombie_threads)) {
434 mtx_lock_spin(&zombie_lock);
435 td_first = TAILQ_FIRST(&zombie_threads);
437 TAILQ_INIT(&zombie_threads);
438 mtx_unlock_spin(&zombie_lock);
440 td_next = TAILQ_NEXT(td_first, td_slpq);
441 thread_cow_free(td_first);
442 thread_free(td_first);
452 thread_alloc(int pages)
457 thread_reap(); /* check if any zombies to get */
464 td = uma_zalloc(thread_zone, M_WAITOK);
465 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
466 if (!vm_thread_new(td, pages)) {
467 uma_zfree(thread_zone, td);
472 cpu_thread_alloc(td);
473 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
478 thread_alloc_stack(struct thread *td, int pages)
481 KASSERT(td->td_kstack == 0,
482 ("thread_alloc_stack called on a thread with kstack"));
483 if (!vm_thread_new(td, pages))
485 cpu_thread_alloc(td);
490 * Deallocate a thread.
493 thread_free(struct thread *td)
496 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
497 lock_profile_thread_exit(td);
499 cpuset_rel(td->td_cpuset);
500 td->td_cpuset = NULL;
502 if (td->td_kstack != 0)
503 vm_thread_dispose(td);
504 callout_drain(&td->td_slpcallout);
505 tid_free(td->td_tid);
507 uma_zfree(thread_zone, td);
511 thread_cow_get_proc(struct thread *newtd, struct proc *p)
514 PROC_LOCK_ASSERT(p, MA_OWNED);
515 newtd->td_realucred = crcowget(p->p_ucred);
516 newtd->td_ucred = newtd->td_realucred;
517 newtd->td_limit = lim_hold(p->p_limit);
518 newtd->td_cowgen = p->p_cowgen;
522 thread_cow_get(struct thread *newtd, struct thread *td)
525 MPASS(td->td_realucred == td->td_ucred);
526 newtd->td_realucred = crcowget(td->td_realucred);
527 newtd->td_ucred = newtd->td_realucred;
528 newtd->td_limit = lim_hold(td->td_limit);
529 newtd->td_cowgen = td->td_cowgen;
533 thread_cow_free(struct thread *td)
536 if (td->td_realucred != NULL)
538 if (td->td_limit != NULL)
539 lim_free(td->td_limit);
543 thread_cow_update(struct thread *td)
546 struct ucred *oldcred;
547 struct plimit *oldlimit;
552 oldcred = crcowsync();
553 if (td->td_limit != p->p_limit) {
554 oldlimit = td->td_limit;
555 td->td_limit = lim_hold(p->p_limit);
557 td->td_cowgen = p->p_cowgen;
561 if (oldlimit != NULL)
566 * Discard the current thread and exit from its context.
567 * Always called with scheduler locked.
569 * Because we can't free a thread while we're operating under its context,
570 * push the current thread into our CPU's deadthread holder. This means
571 * we needn't worry about someone else grabbing our context before we
577 uint64_t runtime, new_switchtime;
586 PROC_SLOCK_ASSERT(p, MA_OWNED);
587 mtx_assert(&Giant, MA_NOTOWNED);
589 PROC_LOCK_ASSERT(p, MA_OWNED);
590 KASSERT(p != NULL, ("thread exiting without a process"));
591 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
592 (long)p->p_pid, td->td_name);
593 SDT_PROBE0(proc, , , lwp__exit);
594 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
595 MPASS(td->td_realucred == td->td_ucred);
598 * drop FPU & debug register state storage, or any other
599 * architecture specific resources that
600 * would not be on a new untouched process.
605 * The last thread is left attached to the process
606 * So that the whole bundle gets recycled. Skip
607 * all this stuff if we never had threads.
608 * EXIT clears all sign of other threads when
609 * it goes to single threading, so the last thread always
610 * takes the short path.
612 if (p->p_flag & P_HADTHREADS) {
613 if (p->p_numthreads > 1) {
614 atomic_add_int(&td->td_proc->p_exitthreads, 1);
616 td2 = FIRST_THREAD_IN_PROC(p);
617 sched_exit_thread(td2, td);
620 * The test below is NOT true if we are the
621 * sole exiting thread. P_STOPPED_SINGLE is unset
622 * in exit1() after it is the only survivor.
624 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
625 if (p->p_numthreads == p->p_suspcount) {
626 thread_lock(p->p_singlethread);
627 wakeup_swapper = thread_unsuspend_one(
628 p->p_singlethread, p, false);
634 PCPU_SET(deadthread, td);
637 * The last thread is exiting.. but not through exit()
639 panic ("thread_exit: Last thread exiting on its own");
644 * If this thread is part of a process that is being tracked by hwpmc(4),
645 * inform the module of the thread's impending exit.
647 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
648 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
649 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
650 } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
651 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
658 /* Do the same timestamp bookkeeping that mi_switch() would do. */
659 new_switchtime = cpu_ticks();
660 runtime = new_switchtime - PCPU_GET(switchtime);
661 td->td_runtime += runtime;
662 td->td_incruntime += runtime;
663 PCPU_SET(switchtime, new_switchtime);
664 PCPU_SET(switchticks, ticks);
667 /* Save our resource usage in our process. */
668 td->td_ru.ru_nvcsw++;
669 ruxagg_locked(p, td);
670 rucollect(&p->p_ru, &td->td_ru);
673 td->td_state = TDS_INACTIVE;
675 witness_thread_exit(td);
677 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
679 panic("I'm a teapot!");
684 * Do any thread specific cleanups that may be needed in wait()
685 * called with Giant, proc and schedlock not held.
688 thread_wait(struct proc *p)
692 mtx_assert(&Giant, MA_NOTOWNED);
693 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
694 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
695 td = FIRST_THREAD_IN_PROC(p);
696 /* Lock the last thread so we spin until it exits cpu_throw(). */
699 lock_profile_thread_exit(td);
700 cpuset_rel(td->td_cpuset);
701 td->td_cpuset = NULL;
702 cpu_thread_clean(td);
704 callout_drain(&td->td_slpcallout);
705 thread_reap(); /* check for zombie threads etc. */
709 * Link a thread to a process.
710 * set up anything that needs to be initialized for it to
711 * be used by the process.
714 thread_link(struct thread *td, struct proc *p)
718 * XXX This can't be enabled because it's called for proc0 before
719 * its lock has been created.
720 * PROC_LOCK_ASSERT(p, MA_OWNED);
722 td->td_state = TDS_INACTIVE;
724 td->td_flags = TDF_INMEM;
726 LIST_INIT(&td->td_contested);
727 LIST_INIT(&td->td_lprof[0]);
728 LIST_INIT(&td->td_lprof[1]);
730 SLIST_INIT(&td->td_epochs);
732 sigqueue_init(&td->td_sigqueue, p);
733 callout_init(&td->td_slpcallout, 1);
734 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
743 thread_unlink(struct thread *td)
745 struct proc *p = td->td_proc;
747 PROC_LOCK_ASSERT(p, MA_OWNED);
749 MPASS(SLIST_EMPTY(&td->td_epochs));
752 TAILQ_REMOVE(&p->p_threads, td, td_plist);
754 /* could clear a few other things here */
755 /* Must NOT clear links to proc! */
759 calc_remaining(struct proc *p, int mode)
763 PROC_LOCK_ASSERT(p, MA_OWNED);
764 PROC_SLOCK_ASSERT(p, MA_OWNED);
765 if (mode == SINGLE_EXIT)
766 remaining = p->p_numthreads;
767 else if (mode == SINGLE_BOUNDARY)
768 remaining = p->p_numthreads - p->p_boundary_count;
769 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
770 remaining = p->p_numthreads - p->p_suspcount;
772 panic("calc_remaining: wrong mode %d", mode);
777 remain_for_mode(int mode)
780 return (mode == SINGLE_ALLPROC ? 0 : 1);
784 weed_inhib(int mode, struct thread *td2, struct proc *p)
788 PROC_LOCK_ASSERT(p, MA_OWNED);
789 PROC_SLOCK_ASSERT(p, MA_OWNED);
790 THREAD_LOCK_ASSERT(td2, MA_OWNED);
795 * Since the thread lock is dropped by the scheduler we have
796 * to retry to check for races.
801 if (TD_IS_SUSPENDED(td2)) {
802 wakeup_swapper |= thread_unsuspend_one(td2, p, true);
806 if (TD_CAN_ABORT(td2)) {
807 wakeup_swapper |= sleepq_abort(td2, EINTR);
808 return (wakeup_swapper);
811 case SINGLE_BOUNDARY:
813 if (TD_IS_SUSPENDED(td2) &&
814 (td2->td_flags & TDF_BOUNDARY) == 0) {
815 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
819 if (TD_CAN_ABORT(td2)) {
820 wakeup_swapper |= sleepq_abort(td2, ERESTART);
821 return (wakeup_swapper);
826 * ALLPROC suspend tries to avoid spurious EINTR for
827 * threads sleeping interruptable, by suspending the
828 * thread directly, similarly to sig_suspend_threads().
829 * Since such sleep is not performed at the user
830 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
831 * is used to avoid immediate un-suspend.
833 if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
834 TDF_ALLPROCSUSP)) == 0) {
835 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
839 if (TD_CAN_ABORT(td2)) {
840 if ((td2->td_flags & TDF_SBDRY) == 0) {
841 thread_suspend_one(td2);
842 td2->td_flags |= TDF_ALLPROCSUSP;
844 wakeup_swapper |= sleepq_abort(td2, ERESTART);
845 return (wakeup_swapper);
853 return (wakeup_swapper);
857 * Enforce single-threading.
859 * Returns 1 if the caller must abort (another thread is waiting to
860 * exit the process or similar). Process is locked!
861 * Returns 0 when you are successfully the only thread running.
862 * A process has successfully single threaded in the suspend mode when
863 * There are no threads in user mode. Threads in the kernel must be
864 * allowed to continue until they get to the user boundary. They may even
865 * copy out their return values and data before suspending. They may however be
866 * accelerated in reaching the user boundary as we will wake up
867 * any sleeping threads that are interruptable. (PCATCH).
870 thread_single(struct proc *p, int mode)
874 int remaining, wakeup_swapper;
877 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
878 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
879 ("invalid mode %d", mode));
881 * If allowing non-ALLPROC singlethreading for non-curproc
882 * callers, calc_remaining() and remain_for_mode() should be
883 * adjusted to also account for td->td_proc != p. For now
884 * this is not implemented because it is not used.
886 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
887 (mode != SINGLE_ALLPROC && td->td_proc == p),
888 ("mode %d proc %p curproc %p", mode, p, td->td_proc));
889 mtx_assert(&Giant, MA_NOTOWNED);
890 PROC_LOCK_ASSERT(p, MA_OWNED);
892 if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
895 /* Is someone already single threading? */
896 if (p->p_singlethread != NULL && p->p_singlethread != td)
899 if (mode == SINGLE_EXIT) {
900 p->p_flag |= P_SINGLE_EXIT;
901 p->p_flag &= ~P_SINGLE_BOUNDARY;
903 p->p_flag &= ~P_SINGLE_EXIT;
904 if (mode == SINGLE_BOUNDARY)
905 p->p_flag |= P_SINGLE_BOUNDARY;
907 p->p_flag &= ~P_SINGLE_BOUNDARY;
909 if (mode == SINGLE_ALLPROC)
910 p->p_flag |= P_TOTAL_STOP;
911 p->p_flag |= P_STOPPED_SINGLE;
913 p->p_singlethread = td;
914 remaining = calc_remaining(p, mode);
915 while (remaining != remain_for_mode(mode)) {
916 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
919 FOREACH_THREAD_IN_PROC(p, td2) {
923 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
924 if (TD_IS_INHIBITED(td2)) {
925 wakeup_swapper |= weed_inhib(mode, td2, p);
927 } else if (TD_IS_RUNNING(td2) && td != td2) {
936 remaining = calc_remaining(p, mode);
939 * Maybe we suspended some threads.. was it enough?
941 if (remaining == remain_for_mode(mode))
946 * Wake us up when everyone else has suspended.
947 * In the mean time we suspend as well.
949 thread_suspend_switch(td, p);
950 remaining = calc_remaining(p, mode);
952 if (mode == SINGLE_EXIT) {
954 * Convert the process to an unthreaded process. The
955 * SINGLE_EXIT is called by exit1() or execve(), in
956 * both cases other threads must be retired.
958 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
959 p->p_singlethread = NULL;
960 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
963 * Wait for any remaining threads to exit cpu_throw().
965 while (p->p_exitthreads != 0) {
968 sched_relinquish(td);
972 } else if (mode == SINGLE_BOUNDARY) {
974 * Wait until all suspended threads are removed from
975 * the processors. The thread_suspend_check()
976 * increments p_boundary_count while it is still
977 * running, which makes it possible for the execve()
978 * to destroy vmspace while our other threads are
979 * still using the address space.
981 * We lock the thread, which is only allowed to
982 * succeed after context switch code finished using
985 FOREACH_THREAD_IN_PROC(p, td2) {
989 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
990 ("td %p not on boundary", td2));
991 KASSERT(TD_IS_SUSPENDED(td2),
992 ("td %p is not suspended", td2));
1001 thread_suspend_check_needed(void)
1008 PROC_LOCK_ASSERT(p, MA_OWNED);
1009 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1010 (td->td_dbgflags & TDB_SUSPEND) != 0));
1014 * Called in from locations that can safely check to see
1015 * whether we have to suspend or at least throttle for a
1016 * single-thread event (e.g. fork).
1018 * Such locations include userret().
1019 * If the "return_instead" argument is non zero, the thread must be able to
1020 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1022 * The 'return_instead' argument tells the function if it may do a
1023 * thread_exit() or suspend, or whether the caller must abort and back
1026 * If the thread that set the single_threading request has set the
1027 * P_SINGLE_EXIT bit in the process flags then this call will never return
1028 * if 'return_instead' is false, but will exit.
1030 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1031 *---------------+--------------------+---------------------
1032 * 0 | returns 0 | returns 0 or 1
1033 * | when ST ends | immediately
1034 *---------------+--------------------+---------------------
1035 * 1 | thread exits | returns 1
1037 * 0 = thread_exit() or suspension ok,
1038 * other = return error instead of stopping the thread.
1040 * While a full suspension is under effect, even a single threading
1041 * thread would be suspended if it made this call (but it shouldn't).
1042 * This call should only be made from places where
1043 * thread_exit() would be safe as that may be the outcome unless
1044 * return_instead is set.
1047 thread_suspend_check(int return_instead)
1055 mtx_assert(&Giant, MA_NOTOWNED);
1056 PROC_LOCK_ASSERT(p, MA_OWNED);
1057 while (thread_suspend_check_needed()) {
1058 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1059 KASSERT(p->p_singlethread != NULL,
1060 ("singlethread not set"));
1062 * The only suspension in action is a
1063 * single-threading. Single threader need not stop.
1064 * It is safe to access p->p_singlethread unlocked
1065 * because it can only be set to our address by us.
1067 if (p->p_singlethread == td)
1068 return (0); /* Exempt from stopping. */
1070 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1073 /* Should we goto user boundary if we didn't come from there? */
1074 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1075 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1079 * Ignore suspend requests if they are deferred.
1081 if ((td->td_flags & TDF_SBDRY) != 0) {
1082 KASSERT(return_instead,
1083 ("TDF_SBDRY set for unsafe thread_suspend_check"));
1084 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1085 (TDF_SEINTR | TDF_SERESTART),
1086 ("both TDF_SEINTR and TDF_SERESTART"));
1087 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1091 * If the process is waiting for us to exit,
1092 * this thread should just suicide.
1093 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1095 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1099 * Allow Linux emulation layer to do some work
1100 * before thread suicide.
1102 if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1103 (p->p_sysent->sv_thread_detach)(td);
1104 umtx_thread_exit(td);
1106 panic("stopped thread did not exit");
1111 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1112 if (p->p_numthreads == p->p_suspcount + 1) {
1113 thread_lock(p->p_singlethread);
1114 wakeup_swapper = thread_unsuspend_one(
1115 p->p_singlethread, p, false);
1123 * When a thread suspends, it just
1124 * gets taken off all queues.
1126 thread_suspend_one(td);
1127 if (return_instead == 0) {
1128 p->p_boundary_count++;
1129 td->td_flags |= TDF_BOUNDARY;
1132 mi_switch(SW_INVOL | SWT_SUSPEND);
1139 * Check for possible stops and suspensions while executing a
1140 * casueword or similar transiently failing operation.
1142 * The sleep argument controls whether the function can handle a stop
1143 * request itself or it should return ERESTART and the request is
1144 * proceed at the kernel/user boundary in ast.
1146 * Typically, when retrying due to casueword(9) failure (rv == 1), we
1147 * should handle the stop requests there, with exception of cases when
1148 * the thread owns a kernel resource, for instance busied the umtx
1149 * key, or when functions return immediately if thread_check_susp()
1150 * returned non-zero. On the other hand, retrying the whole lock
1151 * operation, we better not stop there but delegate the handling to
1154 * If the request is for thread termination P_SINGLE_EXIT, we cannot
1155 * handle it at all, and simply return EINTR.
1158 thread_check_susp(struct thread *td, bool sleep)
1164 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
1165 * eventually break the lockstep loop.
1167 if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
1172 if (p->p_flag & P_SINGLE_EXIT)
1174 else if (P_SHOULDSTOP(p) ||
1175 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1176 error = sleep ? thread_suspend_check(0) : ERESTART;
1182 thread_suspend_switch(struct thread *td, struct proc *p)
1185 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1186 PROC_LOCK_ASSERT(p, MA_OWNED);
1187 PROC_SLOCK_ASSERT(p, MA_OWNED);
1189 * We implement thread_suspend_one in stages here to avoid
1190 * dropping the proc lock while the thread lock is owned.
1192 if (p == td->td_proc) {
1198 td->td_flags &= ~TDF_NEEDSUSPCHK;
1199 TD_SET_SUSPENDED(td);
1203 mi_switch(SW_VOL | SWT_SUSPEND);
1210 thread_suspend_one(struct thread *td)
1215 PROC_SLOCK_ASSERT(p, MA_OWNED);
1216 THREAD_LOCK_ASSERT(td, MA_OWNED);
1217 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1219 td->td_flags &= ~TDF_NEEDSUSPCHK;
1220 TD_SET_SUSPENDED(td);
1225 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1228 THREAD_LOCK_ASSERT(td, MA_OWNED);
1229 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1230 TD_CLR_SUSPENDED(td);
1231 td->td_flags &= ~TDF_ALLPROCSUSP;
1232 if (td->td_proc == p) {
1233 PROC_SLOCK_ASSERT(p, MA_OWNED);
1235 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1236 td->td_flags &= ~TDF_BOUNDARY;
1237 p->p_boundary_count--;
1240 return (setrunnable(td, 0));
1244 * Allow all threads blocked by single threading to continue running.
1247 thread_unsuspend(struct proc *p)
1252 PROC_LOCK_ASSERT(p, MA_OWNED);
1253 PROC_SLOCK_ASSERT(p, MA_OWNED);
1255 if (!P_SHOULDSTOP(p)) {
1256 FOREACH_THREAD_IN_PROC(p, td) {
1258 if (TD_IS_SUSPENDED(td)) {
1259 wakeup_swapper |= thread_unsuspend_one(td, p,
1264 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1265 p->p_numthreads == p->p_suspcount) {
1267 * Stopping everything also did the job for the single
1268 * threading request. Now we've downgraded to single-threaded,
1271 if (p->p_singlethread->td_proc == p) {
1272 thread_lock(p->p_singlethread);
1273 wakeup_swapper = thread_unsuspend_one(
1274 p->p_singlethread, p, false);
1282 * End the single threading mode..
1285 thread_single_end(struct proc *p, int mode)
1290 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1291 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1292 ("invalid mode %d", mode));
1293 PROC_LOCK_ASSERT(p, MA_OWNED);
1294 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1295 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1296 ("mode %d does not match P_TOTAL_STOP", mode));
1297 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1298 ("thread_single_end from other thread %p %p",
1299 curthread, p->p_singlethread));
1300 KASSERT(mode != SINGLE_BOUNDARY ||
1301 (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1302 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1303 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1306 p->p_singlethread = NULL;
1309 * If there are other threads they may now run,
1310 * unless of course there is a blanket 'stop order'
1311 * on the process. The single threader must be allowed
1312 * to continue however as this is a bad place to stop.
1314 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1315 FOREACH_THREAD_IN_PROC(p, td) {
1317 if (TD_IS_SUSPENDED(td)) {
1318 wakeup_swapper |= thread_unsuspend_one(td, p,
1319 mode == SINGLE_BOUNDARY);
1324 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1325 ("inconsistent boundary count %d", p->p_boundary_count));
1332 thread_find(struct proc *p, lwpid_t tid)
1336 PROC_LOCK_ASSERT(p, MA_OWNED);
1337 FOREACH_THREAD_IN_PROC(p, td) {
1338 if (td->td_tid == tid)
1344 /* Locate a thread by number; return with proc lock held. */
1346 tdfind(lwpid_t tid, pid_t pid)
1348 #define RUN_THRESH 16
1353 if (td->td_tid == tid) {
1354 if (pid != -1 && td->td_proc->p_pid != pid)
1356 PROC_LOCK(td->td_proc);
1360 rw_rlock(&tidhash_lock);
1361 LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1362 if (td->td_tid == tid) {
1363 if (pid != -1 && td->td_proc->p_pid != pid) {
1367 PROC_LOCK(td->td_proc);
1368 if (td->td_proc->p_state == PRS_NEW) {
1369 PROC_UNLOCK(td->td_proc);
1373 if (run > RUN_THRESH) {
1374 if (rw_try_upgrade(&tidhash_lock)) {
1375 LIST_REMOVE(td, td_hash);
1376 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1378 rw_wunlock(&tidhash_lock);
1386 rw_runlock(&tidhash_lock);
1391 tidhash_add(struct thread *td)
1393 rw_wlock(&tidhash_lock);
1394 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1395 rw_wunlock(&tidhash_lock);
1399 tidhash_remove(struct thread *td)
1401 rw_wlock(&tidhash_lock);
1402 LIST_REMOVE(td, td_hash);
1403 rw_wunlock(&tidhash_lock);