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,
136 static void thread_free_batched(struct thread *td);
138 static struct mtx tid_lock;
139 static bitstr_t *tid_bitmap;
141 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
143 static int maxthread;
144 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
145 &maxthread, 0, "Maximum number of threads");
149 static LIST_HEAD(tidhashhead, thread) *tidhashtbl;
150 static u_long tidhash;
151 static u_long tidhashlock;
152 static struct rwlock *tidhashtbl_lock;
153 #define TIDHASH(tid) (&tidhashtbl[(tid) & tidhash])
154 #define TIDHASHLOCK(tid) (&tidhashtbl_lock[(tid) & tidhashlock])
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;
166 static lwpid_t trytid;
170 if (nthreads + 1 >= maxthread - 100) {
171 if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
172 nthreads + 1 >= maxthread) {
173 mtx_unlock(&tid_lock);
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);
185 * It is an invariant that the bitmap is big enough to hold maxthread
186 * IDs. If we got to this point there has to be at least one free.
188 if (trytid >= maxthread)
190 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
192 KASSERT(trytid != 0, ("unexpectedly ran out of IDs"));
194 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
195 KASSERT(tid != -1, ("unexpectedly ran out of IDs"));
197 bit_set(tid_bitmap, tid);
199 mtx_unlock(&tid_lock);
200 return (tid + NO_PID);
204 tid_free_locked(lwpid_t rtid)
208 mtx_assert(&tid_lock, MA_OWNED);
209 KASSERT(rtid >= NO_PID,
210 ("%s: invalid tid %d\n", __func__, rtid));
212 KASSERT(bit_test(tid_bitmap, tid) != 0,
213 ("thread ID %d not allocated\n", rtid));
214 bit_clear(tid_bitmap, tid);
219 tid_free(lwpid_t rtid)
223 tid_free_locked(rtid);
224 mtx_unlock(&tid_lock);
228 tid_free_batch(lwpid_t *batch, int n)
233 for (i = 0; i < n; i++) {
234 tid_free_locked(batch[i]);
236 mtx_unlock(&tid_lock);
240 * Prepare a thread for use.
243 thread_ctor(void *mem, int size, void *arg, int flags)
247 td = (struct thread *)mem;
248 td->td_state = TDS_INACTIVE;
249 td->td_lastcpu = td->td_oncpu = NOCPU;
252 * Note that td_critnest begins life as 1 because the thread is not
253 * running and is thereby implicitly waiting to be on the receiving
254 * end of a context switch.
257 td->td_lend_user_pri = PRI_MAX;
259 audit_thread_alloc(td);
261 umtx_thread_alloc(td);
266 * Reclaim a thread after use.
269 thread_dtor(void *mem, int size, void *arg)
273 td = (struct thread *)mem;
276 /* Verify that this thread is in a safe state to free. */
277 switch (td->td_state) {
283 * We must never unlink a thread that is in one of
284 * these states, because it is currently active.
286 panic("bad state for thread unlinking");
291 panic("bad thread state");
296 audit_thread_free(td);
298 /* Free all OSD associated to this thread. */
300 td_softdep_cleanup(td);
301 MPASS(td->td_su == NULL);
305 * Initialize type-stable parts of a thread (when newly created).
308 thread_init(void *mem, int size, int flags)
312 td = (struct thread *)mem;
314 td->td_sleepqueue = sleepq_alloc();
315 td->td_turnstile = turnstile_alloc();
317 EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
318 umtx_thread_init(td);
325 * Tear down type-stable parts of a thread (just before being discarded).
328 thread_fini(void *mem, int size)
332 td = (struct thread *)mem;
333 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
334 rlqentry_free(td->td_rlqe);
335 turnstile_free(td->td_turnstile);
336 sleepq_free(td->td_sleepqueue);
337 umtx_thread_fini(td);
342 * For a newly created process,
343 * link up all the structures and its initial threads etc.
345 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc.
346 * proc_dtor() (should go away)
350 proc_linkup0(struct proc *p, struct thread *td)
352 TAILQ_INIT(&p->p_threads); /* all threads in proc */
357 proc_linkup(struct proc *p, struct thread *td)
360 sigqueue_init(&p->p_sigqueue, p);
361 p->p_ksi = ksiginfo_alloc(1);
362 if (p->p_ksi != NULL) {
363 /* XXX p_ksi may be null if ksiginfo zone is not ready */
364 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
366 LIST_INIT(&p->p_mqnotifier);
371 extern int max_threads_per_proc;
374 * Initialize global thread allocation resources.
384 * Place an upper limit on threads which can be allocated.
386 * Note that other factors may make the de facto limit much lower.
388 * Platform limits are somewhat arbitrary but deemed "more than good
389 * enough" for the foreseable future.
391 if (maxthread == 0) {
393 maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
395 maxthread = MIN(maxproc * max_threads_per_proc, 100000);
399 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
400 tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
402 if (tid0 != THREAD0_TID)
403 panic("tid0 %d != %d\n", tid0, THREAD0_TID);
405 flags = UMA_ZONE_NOFREE;
408 * Force thread structures to be allocated from the direct map.
409 * Otherwise, superpage promotions and demotions may temporarily
410 * invalidate thread structure mappings. For most dynamically allocated
411 * structures this is not a problem, but translation faults cannot be
412 * handled without accessing curthread.
414 flags |= UMA_ZONE_CONTIG;
416 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
417 thread_ctor, thread_dtor, thread_init, thread_fini,
419 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
420 tidhashlock = (tidhash + 1) / 64;
423 tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
424 M_TIDHASH, M_WAITOK | M_ZERO);
425 for (i = 0; i < tidhashlock + 1; i++)
426 rw_init(&tidhashtbl_lock[i], "tidhash");
430 * Place an unused thread on the zombie list.
433 thread_zombie(struct thread *td)
437 ztd = atomic_load_ptr(&thread_zombies);
440 if (atomic_fcmpset_rel_ptr((uintptr_t *)&thread_zombies,
441 (uintptr_t *)&ztd, (uintptr_t)td))
448 * Release a thread that has exited after cpu_throw().
451 thread_stash(struct thread *td)
453 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
458 * Reap zombie threads.
463 struct thread *itd, *ntd;
464 lwpid_t tidbatch[16];
468 * Reading upfront is pessimal if followed by concurrent atomic_swap,
469 * but most of the time the list is empty.
471 if (thread_zombies == NULL)
474 itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&thread_zombies,
477 while (itd != NULL) {
478 ntd = itd->td_zombie;
479 tidbatch[tidbatchn] = itd->td_tid;
481 thread_cow_free(itd);
482 thread_free_batched(itd);
483 if (tidbatchn == nitems(tidbatch)) {
484 tid_free_batch(tidbatch, tidbatchn);
490 if (tidbatchn != 0) {
491 tid_free_batch(tidbatch, tidbatchn);
499 thread_alloc(int pages)
504 thread_reap(); /* check if any zombies to get */
511 td = uma_zalloc(thread_zone, M_WAITOK);
512 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
513 if (!vm_thread_new(td, pages)) {
514 uma_zfree(thread_zone, td);
519 cpu_thread_alloc(td);
520 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
525 thread_alloc_stack(struct thread *td, int pages)
528 KASSERT(td->td_kstack == 0,
529 ("thread_alloc_stack called on a thread with kstack"));
530 if (!vm_thread_new(td, pages))
532 cpu_thread_alloc(td);
537 * Deallocate a thread.
540 thread_free_batched(struct thread *td)
543 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
544 lock_profile_thread_exit(td);
546 cpuset_rel(td->td_cpuset);
547 td->td_cpuset = NULL;
549 if (td->td_kstack != 0)
550 vm_thread_dispose(td);
551 callout_drain(&td->td_slpcallout);
553 * Freeing handled by the caller.
556 uma_zfree(thread_zone, td);
560 thread_free(struct thread *td)
565 thread_free_batched(td);
570 thread_cow_get_proc(struct thread *newtd, struct proc *p)
573 PROC_LOCK_ASSERT(p, MA_OWNED);
574 newtd->td_realucred = crcowget(p->p_ucred);
575 newtd->td_ucred = newtd->td_realucred;
576 newtd->td_limit = lim_hold(p->p_limit);
577 newtd->td_cowgen = p->p_cowgen;
581 thread_cow_get(struct thread *newtd, struct thread *td)
584 MPASS(td->td_realucred == td->td_ucred);
585 newtd->td_realucred = crcowget(td->td_realucred);
586 newtd->td_ucred = newtd->td_realucred;
587 newtd->td_limit = lim_hold(td->td_limit);
588 newtd->td_cowgen = td->td_cowgen;
592 thread_cow_free(struct thread *td)
595 if (td->td_realucred != NULL)
597 if (td->td_limit != NULL)
598 lim_free(td->td_limit);
602 thread_cow_update(struct thread *td)
605 struct ucred *oldcred;
606 struct plimit *oldlimit;
611 oldcred = crcowsync();
612 if (td->td_limit != p->p_limit) {
613 oldlimit = td->td_limit;
614 td->td_limit = lim_hold(p->p_limit);
616 td->td_cowgen = p->p_cowgen;
620 if (oldlimit != NULL)
625 * Discard the current thread and exit from its context.
626 * Always called with scheduler locked.
628 * Because we can't free a thread while we're operating under its context,
629 * push the current thread into our CPU's deadthread holder. This means
630 * we needn't worry about someone else grabbing our context before we
636 uint64_t runtime, new_switchtime;
645 PROC_SLOCK_ASSERT(p, MA_OWNED);
646 mtx_assert(&Giant, MA_NOTOWNED);
648 PROC_LOCK_ASSERT(p, MA_OWNED);
649 KASSERT(p != NULL, ("thread exiting without a process"));
650 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
651 (long)p->p_pid, td->td_name);
652 SDT_PROBE0(proc, , , lwp__exit);
653 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
654 MPASS(td->td_realucred == td->td_ucred);
657 * drop FPU & debug register state storage, or any other
658 * architecture specific resources that
659 * would not be on a new untouched process.
664 * The last thread is left attached to the process
665 * So that the whole bundle gets recycled. Skip
666 * all this stuff if we never had threads.
667 * EXIT clears all sign of other threads when
668 * it goes to single threading, so the last thread always
669 * takes the short path.
671 if (p->p_flag & P_HADTHREADS) {
672 if (p->p_numthreads > 1) {
673 atomic_add_int(&td->td_proc->p_exitthreads, 1);
675 td2 = FIRST_THREAD_IN_PROC(p);
676 sched_exit_thread(td2, td);
679 * The test below is NOT true if we are the
680 * sole exiting thread. P_STOPPED_SINGLE is unset
681 * in exit1() after it is the only survivor.
683 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
684 if (p->p_numthreads == p->p_suspcount) {
685 thread_lock(p->p_singlethread);
686 wakeup_swapper = thread_unsuspend_one(
687 p->p_singlethread, p, false);
693 PCPU_SET(deadthread, td);
696 * The last thread is exiting.. but not through exit()
698 panic ("thread_exit: Last thread exiting on its own");
703 * If this thread is part of a process that is being tracked by hwpmc(4),
704 * inform the module of the thread's impending exit.
706 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
707 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
708 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
709 } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
710 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
717 /* Do the same timestamp bookkeeping that mi_switch() would do. */
718 new_switchtime = cpu_ticks();
719 runtime = new_switchtime - PCPU_GET(switchtime);
720 td->td_runtime += runtime;
721 td->td_incruntime += runtime;
722 PCPU_SET(switchtime, new_switchtime);
723 PCPU_SET(switchticks, ticks);
726 /* Save our resource usage in our process. */
727 td->td_ru.ru_nvcsw++;
728 ruxagg_locked(p, td);
729 rucollect(&p->p_ru, &td->td_ru);
732 td->td_state = TDS_INACTIVE;
734 witness_thread_exit(td);
736 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
738 panic("I'm a teapot!");
743 * Do any thread specific cleanups that may be needed in wait()
744 * called with Giant, proc and schedlock not held.
747 thread_wait(struct proc *p)
751 mtx_assert(&Giant, MA_NOTOWNED);
752 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
753 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
754 td = FIRST_THREAD_IN_PROC(p);
755 /* Lock the last thread so we spin until it exits cpu_throw(). */
758 lock_profile_thread_exit(td);
759 cpuset_rel(td->td_cpuset);
760 td->td_cpuset = NULL;
761 cpu_thread_clean(td);
763 callout_drain(&td->td_slpcallout);
764 thread_reap(); /* check for zombie threads etc. */
768 * Link a thread to a process.
769 * set up anything that needs to be initialized for it to
770 * be used by the process.
773 thread_link(struct thread *td, struct proc *p)
777 * XXX This can't be enabled because it's called for proc0 before
778 * its lock has been created.
779 * PROC_LOCK_ASSERT(p, MA_OWNED);
781 td->td_state = TDS_INACTIVE;
783 td->td_flags = TDF_INMEM;
785 LIST_INIT(&td->td_contested);
786 LIST_INIT(&td->td_lprof[0]);
787 LIST_INIT(&td->td_lprof[1]);
789 SLIST_INIT(&td->td_epochs);
791 sigqueue_init(&td->td_sigqueue, p);
792 callout_init(&td->td_slpcallout, 1);
793 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
802 thread_unlink(struct thread *td)
804 struct proc *p = td->td_proc;
806 PROC_LOCK_ASSERT(p, MA_OWNED);
808 MPASS(SLIST_EMPTY(&td->td_epochs));
811 TAILQ_REMOVE(&p->p_threads, td, td_plist);
813 /* could clear a few other things here */
814 /* Must NOT clear links to proc! */
818 calc_remaining(struct proc *p, int mode)
822 PROC_LOCK_ASSERT(p, MA_OWNED);
823 PROC_SLOCK_ASSERT(p, MA_OWNED);
824 if (mode == SINGLE_EXIT)
825 remaining = p->p_numthreads;
826 else if (mode == SINGLE_BOUNDARY)
827 remaining = p->p_numthreads - p->p_boundary_count;
828 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
829 remaining = p->p_numthreads - p->p_suspcount;
831 panic("calc_remaining: wrong mode %d", mode);
836 remain_for_mode(int mode)
839 return (mode == SINGLE_ALLPROC ? 0 : 1);
843 weed_inhib(int mode, struct thread *td2, struct proc *p)
847 PROC_LOCK_ASSERT(p, MA_OWNED);
848 PROC_SLOCK_ASSERT(p, MA_OWNED);
849 THREAD_LOCK_ASSERT(td2, MA_OWNED);
854 * Since the thread lock is dropped by the scheduler we have
855 * to retry to check for races.
860 if (TD_IS_SUSPENDED(td2)) {
861 wakeup_swapper |= thread_unsuspend_one(td2, p, true);
865 if (TD_CAN_ABORT(td2)) {
866 wakeup_swapper |= sleepq_abort(td2, EINTR);
867 return (wakeup_swapper);
870 case SINGLE_BOUNDARY:
872 if (TD_IS_SUSPENDED(td2) &&
873 (td2->td_flags & TDF_BOUNDARY) == 0) {
874 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
878 if (TD_CAN_ABORT(td2)) {
879 wakeup_swapper |= sleepq_abort(td2, ERESTART);
880 return (wakeup_swapper);
885 * ALLPROC suspend tries to avoid spurious EINTR for
886 * threads sleeping interruptable, by suspending the
887 * thread directly, similarly to sig_suspend_threads().
888 * Since such sleep is not performed at the user
889 * boundary, TDF_BOUNDARY flag is not set, and TDF_ALLPROCSUSP
890 * is used to avoid immediate un-suspend.
892 if (TD_IS_SUSPENDED(td2) && (td2->td_flags & (TDF_BOUNDARY |
893 TDF_ALLPROCSUSP)) == 0) {
894 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
898 if (TD_CAN_ABORT(td2)) {
899 if ((td2->td_flags & TDF_SBDRY) == 0) {
900 thread_suspend_one(td2);
901 td2->td_flags |= TDF_ALLPROCSUSP;
903 wakeup_swapper |= sleepq_abort(td2, ERESTART);
904 return (wakeup_swapper);
912 return (wakeup_swapper);
916 * Enforce single-threading.
918 * Returns 1 if the caller must abort (another thread is waiting to
919 * exit the process or similar). Process is locked!
920 * Returns 0 when you are successfully the only thread running.
921 * A process has successfully single threaded in the suspend mode when
922 * There are no threads in user mode. Threads in the kernel must be
923 * allowed to continue until they get to the user boundary. They may even
924 * copy out their return values and data before suspending. They may however be
925 * accelerated in reaching the user boundary as we will wake up
926 * any sleeping threads that are interruptable. (PCATCH).
929 thread_single(struct proc *p, int mode)
933 int remaining, wakeup_swapper;
936 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
937 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
938 ("invalid mode %d", mode));
940 * If allowing non-ALLPROC singlethreading for non-curproc
941 * callers, calc_remaining() and remain_for_mode() should be
942 * adjusted to also account for td->td_proc != p. For now
943 * this is not implemented because it is not used.
945 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
946 (mode != SINGLE_ALLPROC && td->td_proc == p),
947 ("mode %d proc %p curproc %p", mode, p, td->td_proc));
948 mtx_assert(&Giant, MA_NOTOWNED);
949 PROC_LOCK_ASSERT(p, MA_OWNED);
951 if ((p->p_flag & P_HADTHREADS) == 0 && mode != SINGLE_ALLPROC)
954 /* Is someone already single threading? */
955 if (p->p_singlethread != NULL && p->p_singlethread != td)
958 if (mode == SINGLE_EXIT) {
959 p->p_flag |= P_SINGLE_EXIT;
960 p->p_flag &= ~P_SINGLE_BOUNDARY;
962 p->p_flag &= ~P_SINGLE_EXIT;
963 if (mode == SINGLE_BOUNDARY)
964 p->p_flag |= P_SINGLE_BOUNDARY;
966 p->p_flag &= ~P_SINGLE_BOUNDARY;
968 if (mode == SINGLE_ALLPROC)
969 p->p_flag |= P_TOTAL_STOP;
970 p->p_flag |= P_STOPPED_SINGLE;
972 p->p_singlethread = td;
973 remaining = calc_remaining(p, mode);
974 while (remaining != remain_for_mode(mode)) {
975 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
978 FOREACH_THREAD_IN_PROC(p, td2) {
982 td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
983 if (TD_IS_INHIBITED(td2)) {
984 wakeup_swapper |= weed_inhib(mode, td2, p);
986 } else if (TD_IS_RUNNING(td2) && td != td2) {
995 remaining = calc_remaining(p, mode);
998 * Maybe we suspended some threads.. was it enough?
1000 if (remaining == remain_for_mode(mode))
1005 * Wake us up when everyone else has suspended.
1006 * In the mean time we suspend as well.
1008 thread_suspend_switch(td, p);
1009 remaining = calc_remaining(p, mode);
1011 if (mode == SINGLE_EXIT) {
1013 * Convert the process to an unthreaded process. The
1014 * SINGLE_EXIT is called by exit1() or execve(), in
1015 * both cases other threads must be retired.
1017 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1018 p->p_singlethread = NULL;
1019 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1022 * Wait for any remaining threads to exit cpu_throw().
1024 while (p->p_exitthreads != 0) {
1027 sched_relinquish(td);
1031 } else if (mode == SINGLE_BOUNDARY) {
1033 * Wait until all suspended threads are removed from
1034 * the processors. The thread_suspend_check()
1035 * increments p_boundary_count while it is still
1036 * running, which makes it possible for the execve()
1037 * to destroy vmspace while our other threads are
1038 * still using the address space.
1040 * We lock the thread, which is only allowed to
1041 * succeed after context switch code finished using
1042 * the address space.
1044 FOREACH_THREAD_IN_PROC(p, td2) {
1048 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1049 ("td %p not on boundary", td2));
1050 KASSERT(TD_IS_SUSPENDED(td2),
1051 ("td %p is not suspended", td2));
1060 thread_suspend_check_needed(void)
1067 PROC_LOCK_ASSERT(p, MA_OWNED);
1068 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1069 (td->td_dbgflags & TDB_SUSPEND) != 0));
1073 * Called in from locations that can safely check to see
1074 * whether we have to suspend or at least throttle for a
1075 * single-thread event (e.g. fork).
1077 * Such locations include userret().
1078 * If the "return_instead" argument is non zero, the thread must be able to
1079 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1081 * The 'return_instead' argument tells the function if it may do a
1082 * thread_exit() or suspend, or whether the caller must abort and back
1085 * If the thread that set the single_threading request has set the
1086 * P_SINGLE_EXIT bit in the process flags then this call will never return
1087 * if 'return_instead' is false, but will exit.
1089 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1090 *---------------+--------------------+---------------------
1091 * 0 | returns 0 | returns 0 or 1
1092 * | when ST ends | immediately
1093 *---------------+--------------------+---------------------
1094 * 1 | thread exits | returns 1
1096 * 0 = thread_exit() or suspension ok,
1097 * other = return error instead of stopping the thread.
1099 * While a full suspension is under effect, even a single threading
1100 * thread would be suspended if it made this call (but it shouldn't).
1101 * This call should only be made from places where
1102 * thread_exit() would be safe as that may be the outcome unless
1103 * return_instead is set.
1106 thread_suspend_check(int return_instead)
1114 mtx_assert(&Giant, MA_NOTOWNED);
1115 PROC_LOCK_ASSERT(p, MA_OWNED);
1116 while (thread_suspend_check_needed()) {
1117 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1118 KASSERT(p->p_singlethread != NULL,
1119 ("singlethread not set"));
1121 * The only suspension in action is a
1122 * single-threading. Single threader need not stop.
1123 * It is safe to access p->p_singlethread unlocked
1124 * because it can only be set to our address by us.
1126 if (p->p_singlethread == td)
1127 return (0); /* Exempt from stopping. */
1129 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1132 /* Should we goto user boundary if we didn't come from there? */
1133 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1134 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1138 * Ignore suspend requests if they are deferred.
1140 if ((td->td_flags & TDF_SBDRY) != 0) {
1141 KASSERT(return_instead,
1142 ("TDF_SBDRY set for unsafe thread_suspend_check"));
1143 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1144 (TDF_SEINTR | TDF_SERESTART),
1145 ("both TDF_SEINTR and TDF_SERESTART"));
1146 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1150 * If the process is waiting for us to exit,
1151 * this thread should just suicide.
1152 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1154 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1158 * Allow Linux emulation layer to do some work
1159 * before thread suicide.
1161 if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1162 (p->p_sysent->sv_thread_detach)(td);
1163 umtx_thread_exit(td);
1165 panic("stopped thread did not exit");
1170 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1171 if (p->p_numthreads == p->p_suspcount + 1) {
1172 thread_lock(p->p_singlethread);
1173 wakeup_swapper = thread_unsuspend_one(
1174 p->p_singlethread, p, false);
1182 * When a thread suspends, it just
1183 * gets taken off all queues.
1185 thread_suspend_one(td);
1186 if (return_instead == 0) {
1187 p->p_boundary_count++;
1188 td->td_flags |= TDF_BOUNDARY;
1191 mi_switch(SW_INVOL | SWT_SUSPEND);
1198 * Check for possible stops and suspensions while executing a
1199 * casueword or similar transiently failing operation.
1201 * The sleep argument controls whether the function can handle a stop
1202 * request itself or it should return ERESTART and the request is
1203 * proceed at the kernel/user boundary in ast.
1205 * Typically, when retrying due to casueword(9) failure (rv == 1), we
1206 * should handle the stop requests there, with exception of cases when
1207 * the thread owns a kernel resource, for instance busied the umtx
1208 * key, or when functions return immediately if thread_check_susp()
1209 * returned non-zero. On the other hand, retrying the whole lock
1210 * operation, we better not stop there but delegate the handling to
1213 * If the request is for thread termination P_SINGLE_EXIT, we cannot
1214 * handle it at all, and simply return EINTR.
1217 thread_check_susp(struct thread *td, bool sleep)
1223 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
1224 * eventually break the lockstep loop.
1226 if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
1231 if (p->p_flag & P_SINGLE_EXIT)
1233 else if (P_SHOULDSTOP(p) ||
1234 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1235 error = sleep ? thread_suspend_check(0) : ERESTART;
1241 thread_suspend_switch(struct thread *td, struct proc *p)
1244 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1245 PROC_LOCK_ASSERT(p, MA_OWNED);
1246 PROC_SLOCK_ASSERT(p, MA_OWNED);
1248 * We implement thread_suspend_one in stages here to avoid
1249 * dropping the proc lock while the thread lock is owned.
1251 if (p == td->td_proc) {
1257 td->td_flags &= ~TDF_NEEDSUSPCHK;
1258 TD_SET_SUSPENDED(td);
1262 mi_switch(SW_VOL | SWT_SUSPEND);
1269 thread_suspend_one(struct thread *td)
1274 PROC_SLOCK_ASSERT(p, MA_OWNED);
1275 THREAD_LOCK_ASSERT(td, MA_OWNED);
1276 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1278 td->td_flags &= ~TDF_NEEDSUSPCHK;
1279 TD_SET_SUSPENDED(td);
1284 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1287 THREAD_LOCK_ASSERT(td, MA_OWNED);
1288 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1289 TD_CLR_SUSPENDED(td);
1290 td->td_flags &= ~TDF_ALLPROCSUSP;
1291 if (td->td_proc == p) {
1292 PROC_SLOCK_ASSERT(p, MA_OWNED);
1294 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1295 td->td_flags &= ~TDF_BOUNDARY;
1296 p->p_boundary_count--;
1299 return (setrunnable(td, 0));
1303 * Allow all threads blocked by single threading to continue running.
1306 thread_unsuspend(struct proc *p)
1311 PROC_LOCK_ASSERT(p, MA_OWNED);
1312 PROC_SLOCK_ASSERT(p, MA_OWNED);
1314 if (!P_SHOULDSTOP(p)) {
1315 FOREACH_THREAD_IN_PROC(p, td) {
1317 if (TD_IS_SUSPENDED(td)) {
1318 wakeup_swapper |= thread_unsuspend_one(td, p,
1323 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1324 p->p_numthreads == p->p_suspcount) {
1326 * Stopping everything also did the job for the single
1327 * threading request. Now we've downgraded to single-threaded,
1330 if (p->p_singlethread->td_proc == p) {
1331 thread_lock(p->p_singlethread);
1332 wakeup_swapper = thread_unsuspend_one(
1333 p->p_singlethread, p, false);
1341 * End the single threading mode..
1344 thread_single_end(struct proc *p, int mode)
1349 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1350 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1351 ("invalid mode %d", mode));
1352 PROC_LOCK_ASSERT(p, MA_OWNED);
1353 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1354 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1355 ("mode %d does not match P_TOTAL_STOP", mode));
1356 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1357 ("thread_single_end from other thread %p %p",
1358 curthread, p->p_singlethread));
1359 KASSERT(mode != SINGLE_BOUNDARY ||
1360 (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1361 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1362 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1365 p->p_singlethread = NULL;
1368 * If there are other threads they may now run,
1369 * unless of course there is a blanket 'stop order'
1370 * on the process. The single threader must be allowed
1371 * to continue however as this is a bad place to stop.
1373 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1374 FOREACH_THREAD_IN_PROC(p, td) {
1376 if (TD_IS_SUSPENDED(td)) {
1377 wakeup_swapper |= thread_unsuspend_one(td, p,
1378 mode == SINGLE_BOUNDARY);
1383 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1384 ("inconsistent boundary count %d", p->p_boundary_count));
1391 * Locate a thread by number and return with proc lock held.
1393 * thread exit establishes proc -> tidhash lock ordering, but lookup
1394 * takes tidhash first and needs to return locked proc.
1396 * The problem is worked around by relying on type-safety of both
1397 * structures and doing the work in 2 steps:
1398 * - tidhash-locked lookup which saves both thread and proc pointers
1399 * - proc-locked verification that the found thread still matches
1402 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1404 #define RUN_THRESH 16
1411 rw_rlock(TIDHASHLOCK(tid));
1413 LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1414 if (td->td_tid != tid) {
1419 if (pid != -1 && p->p_pid != pid) {
1423 if (run > RUN_THRESH) {
1424 if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1425 LIST_REMOVE(td, td_hash);
1426 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1428 rw_wunlock(TIDHASHLOCK(tid));
1436 rw_runlock(TIDHASHLOCK(tid));
1445 tdfind(lwpid_t tid, pid_t pid)
1451 if (td->td_tid == tid) {
1452 if (pid != -1 && td->td_proc->p_pid != pid)
1454 PROC_LOCK(td->td_proc);
1459 if (!tdfind_hash(tid, pid, &p, &td))
1462 if (td->td_tid != tid) {
1466 if (td->td_proc != p) {
1470 if (p->p_state == PRS_NEW) {
1479 tidhash_add(struct thread *td)
1481 rw_wlock(TIDHASHLOCK(td->td_tid));
1482 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1483 rw_wunlock(TIDHASHLOCK(td->td_tid));
1487 tidhash_remove(struct thread *td)
1490 rw_wlock(TIDHASHLOCK(td->td_tid));
1491 LIST_REMOVE(td, td_hash);
1492 rw_wunlock(TIDHASHLOCK(td->td_tid));