1/*
2 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice(s), this list of conditions and the following disclaimer as
10 * the first lines of this file unmodified other than the possible
11 * addition of one or more copyright notices.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice(s), this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
17 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
20 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
21 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
26 * DAMAGE.
27 */
28
29#include <sys/cdefs.h>
30__FBSDID("$FreeBSD: head/sys/kern/kern_thread.c 118486 2003-08-05 12:00:55Z davidxu $");
31
32#include <sys/param.h>
33#include <sys/systm.h>
34#include <sys/kernel.h>
35#include <sys/lock.h>
36#include <sys/malloc.h>
37#include <sys/mutex.h>
38#include <sys/proc.h>
39#include <sys/smp.h>
40#include <sys/sysctl.h>
41#include <sys/sysproto.h>
42#include <sys/filedesc.h>
43#include <sys/sched.h>
44#include <sys/signalvar.h>
45#include <sys/sx.h>
46#include <sys/tty.h>
47#include <sys/user.h>
48#include <sys/jail.h>
49#include <sys/kse.h>
50#include <sys/ktr.h>
51#include <sys/ucontext.h>
52
53#include <vm/vm.h>
54#include <vm/vm_extern.h>
55#include <vm/vm_object.h>
56#include <vm/pmap.h>
57#include <vm/uma.h>
58#include <vm/vm_map.h>
59
60#include <machine/frame.h>
61
62/*
63 * KSEGRP related storage.
64 */
65static uma_zone_t ksegrp_zone;
66static uma_zone_t kse_zone;
67static uma_zone_t thread_zone;
68static uma_zone_t upcall_zone;
69
70/* DEBUG ONLY */
71SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
72static int thread_debug = 0;
73SYSCTL_INT(_kern_threads, OID_AUTO, debug, CTLFLAG_RW,
74 &thread_debug, 0, "thread debug");
75
76static int max_threads_per_proc = 150;
77SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
78 &max_threads_per_proc, 0, "Limit on threads per proc");
79
80static int max_groups_per_proc = 50;
81SYSCTL_INT(_kern_threads, OID_AUTO, max_groups_per_proc, CTLFLAG_RW,
82 &max_groups_per_proc, 0, "Limit on thread groups per proc");
83
84static int max_threads_hits;
85SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
86 &max_threads_hits, 0, "");
87
88static int virtual_cpu;
89
90#define RANGEOF(type, start, end) (offsetof(type, end) - offsetof(type, start))
91
92TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
93TAILQ_HEAD(, kse) zombie_kses = TAILQ_HEAD_INITIALIZER(zombie_kses);
94TAILQ_HEAD(, ksegrp) zombie_ksegrps = TAILQ_HEAD_INITIALIZER(zombie_ksegrps);
95TAILQ_HEAD(, kse_upcall) zombie_upcalls =
96 TAILQ_HEAD_INITIALIZER(zombie_upcalls);
97struct mtx kse_zombie_lock;
98MTX_SYSINIT(kse_zombie_lock, &kse_zombie_lock, "kse zombie lock", MTX_SPIN);
99
100static void kse_purge(struct proc *p, struct thread *td);
101static void kse_purge_group(struct thread *td);
102static int thread_update_usr_ticks(struct thread *td, int user);
103static void thread_alloc_spare(struct thread *td, struct thread *spare);
104
105static int
106sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
107{
108 int error, new_val;
109 int def_val;
110
111#ifdef SMP
112 def_val = mp_ncpus;
113#else
114 def_val = 1;
115#endif
116 if (virtual_cpu == 0)
117 new_val = def_val;
118 else
119 new_val = virtual_cpu;
120 error = sysctl_handle_int(oidp, &new_val, 0, req);
121 if (error != 0 || req->newptr == NULL)
122 return (error);
123 if (new_val < 0)
124 return (EINVAL);
125 virtual_cpu = new_val;
126 return (0);
127}
128
129/* DEBUG ONLY */
130SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
131 0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
132 "debug virtual cpus");
133
134/*
135 * Prepare a thread for use.
136 */
137static void
138thread_ctor(void *mem, int size, void *arg)
139{
140 struct thread *td;
141
142 td = (struct thread *)mem;
143 td->td_state = TDS_INACTIVE;
144 td->td_oncpu = NOCPU;
145 td->td_critnest = 1;
146}
147
148/*
149 * Reclaim a thread after use.
150 */
151static void
152thread_dtor(void *mem, int size, void *arg)
153{
154 struct thread *td;
155
156 td = (struct thread *)mem;
157
158#ifdef INVARIANTS
159 /* Verify that this thread is in a safe state to free. */
160 switch (td->td_state) {
161 case TDS_INHIBITED:
162 case TDS_RUNNING:
163 case TDS_CAN_RUN:
164 case TDS_RUNQ:
165 /*
166 * We must never unlink a thread that is in one of
167 * these states, because it is currently active.
168 */
169 panic("bad state for thread unlinking");
170 /* NOTREACHED */
171 case TDS_INACTIVE:
172 break;
173 default:
174 panic("bad thread state");
175 /* NOTREACHED */
176 }
177#endif
178}
179
180/*
181 * Initialize type-stable parts of a thread (when newly created).
182 */
183static void
184thread_init(void *mem, int size)
185{
186 struct thread *td;
187
188 td = (struct thread *)mem;
189 mtx_lock(&Giant);
190 vm_thread_new(td, 0);
191 mtx_unlock(&Giant);
192 cpu_thread_setup(td);
193 td->td_sched = (struct td_sched *)&td[1];
194}
195
196/*
197 * Tear down type-stable parts of a thread (just before being discarded).
198 */
199static void
200thread_fini(void *mem, int size)
201{
202 struct thread *td;
203
204 td = (struct thread *)mem;
205 vm_thread_dispose(td);
206}
207
208/*
209 * Initialize type-stable parts of a kse (when newly created).
210 */
211static void
212kse_init(void *mem, int size)
213{
214 struct kse *ke;
215
216 ke = (struct kse *)mem;
217 ke->ke_sched = (struct ke_sched *)&ke[1];
218}
219
220/*
221 * Initialize type-stable parts of a ksegrp (when newly created).
222 */
223static void
224ksegrp_init(void *mem, int size)
225{
226 struct ksegrp *kg;
227
228 kg = (struct ksegrp *)mem;
229 kg->kg_sched = (struct kg_sched *)&kg[1];
230}
231
232/*
233 * KSE is linked into kse group.
234 */
235void
236kse_link(struct kse *ke, struct ksegrp *kg)
237{
238 struct proc *p = kg->kg_proc;
239
240 TAILQ_INSERT_HEAD(&kg->kg_kseq, ke, ke_kglist);
241 kg->kg_kses++;
242 ke->ke_state = KES_UNQUEUED;
243 ke->ke_proc = p;
244 ke->ke_ksegrp = kg;
245 ke->ke_thread = NULL;
246 ke->ke_oncpu = NOCPU;
247 ke->ke_flags = 0;
248}
249
250void
251kse_unlink(struct kse *ke)
252{
253 struct ksegrp *kg;
254
255 mtx_assert(&sched_lock, MA_OWNED);
256 kg = ke->ke_ksegrp;
257 TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist);
258 if (ke->ke_state == KES_IDLE) {
259 TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist);
260 kg->kg_idle_kses--;
261 }
262 if (--kg->kg_kses == 0)
263 ksegrp_unlink(kg);
264 /*
265 * Aggregate stats from the KSE
266 */
267 kse_stash(ke);
268}
269
270void
271ksegrp_link(struct ksegrp *kg, struct proc *p)
272{
273
274 TAILQ_INIT(&kg->kg_threads);
275 TAILQ_INIT(&kg->kg_runq); /* links with td_runq */
276 TAILQ_INIT(&kg->kg_slpq); /* links with td_runq */
277 TAILQ_INIT(&kg->kg_kseq); /* all kses in ksegrp */
278 TAILQ_INIT(&kg->kg_iq); /* all idle kses in ksegrp */
279 TAILQ_INIT(&kg->kg_upcalls); /* all upcall structure in ksegrp */
280 kg->kg_proc = p;
281 /*
282 * the following counters are in the -zero- section
283 * and may not need clearing
284 */
285 kg->kg_numthreads = 0;
286 kg->kg_runnable = 0;
287 kg->kg_kses = 0;
288 kg->kg_runq_kses = 0; /* XXXKSE change name */
289 kg->kg_idle_kses = 0;
290 kg->kg_numupcalls = 0;
291 /* link it in now that it's consistent */
292 p->p_numksegrps++;
293 TAILQ_INSERT_HEAD(&p->p_ksegrps, kg, kg_ksegrp);
294}
295
296void
297ksegrp_unlink(struct ksegrp *kg)
298{
299 struct proc *p;
300
301 mtx_assert(&sched_lock, MA_OWNED);
302 KASSERT((kg->kg_numthreads == 0), ("ksegrp_unlink: residual threads"));
303 KASSERT((kg->kg_kses == 0), ("ksegrp_unlink: residual kses"));
304 KASSERT((kg->kg_numupcalls == 0), ("ksegrp_unlink: residual upcalls"));
305
306 p = kg->kg_proc;
307 TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
308 p->p_numksegrps--;
309 /*
310 * Aggregate stats from the KSE
311 */
312 ksegrp_stash(kg);
313}
314
315struct kse_upcall *
316upcall_alloc(void)
317{
318 struct kse_upcall *ku;
319
320 ku = uma_zalloc(upcall_zone, M_WAITOK);
321 bzero(ku, sizeof(*ku));
322 return (ku);
323}
324
325void
326upcall_free(struct kse_upcall *ku)
327{
328
329 uma_zfree(upcall_zone, ku);
330}
331
332void
333upcall_link(struct kse_upcall *ku, struct ksegrp *kg)
334{
335
336 mtx_assert(&sched_lock, MA_OWNED);
337 TAILQ_INSERT_TAIL(&kg->kg_upcalls, ku, ku_link);
338 ku->ku_ksegrp = kg;
339 kg->kg_numupcalls++;
340}
341
342void
343upcall_unlink(struct kse_upcall *ku)
344{
345 struct ksegrp *kg = ku->ku_ksegrp;
346
347 mtx_assert(&sched_lock, MA_OWNED);
348 KASSERT(ku->ku_owner == NULL, ("%s: have owner", __func__));
349 TAILQ_REMOVE(&kg->kg_upcalls, ku, ku_link);
350 kg->kg_numupcalls--;
351 upcall_stash(ku);
352}
353
354void
355upcall_remove(struct thread *td)
356{
357
358 if (td->td_upcall) {
359 td->td_upcall->ku_owner = NULL;
360 upcall_unlink(td->td_upcall);
361 td->td_upcall = 0;
362 }
363}
364
365/*
366 * For a newly created process,
367 * link up all the structures and its initial threads etc.
368 */
369void
370proc_linkup(struct proc *p, struct ksegrp *kg,
371 struct kse *ke, struct thread *td)
372{
373
374 TAILQ_INIT(&p->p_ksegrps); /* all ksegrps in proc */
375 TAILQ_INIT(&p->p_threads); /* all threads in proc */
376 TAILQ_INIT(&p->p_suspended); /* Threads suspended */
377 p->p_numksegrps = 0;
378 p->p_numthreads = 0;
379
380 ksegrp_link(kg, p);
381 kse_link(ke, kg);
382 thread_link(td, kg);
383}
384
385/*
386struct kse_thr_interrupt_args {
387 struct kse_thr_mailbox * tmbx;
388 int cmd;
389 long data;
390};
391*/
392int
393kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap)
394{
395 struct proc *p;
396 struct thread *td2;
397
398 p = td->td_proc;
399 if (!(p->p_flag & P_SA))
400 return (EINVAL);
401
402 switch (uap->cmd) {
403 case KSE_INTR_SENDSIG:
404 if (uap->data < 0 || uap->data > _SIG_MAXSIG)
405 return (EINVAL);
406 case KSE_INTR_INTERRUPT:
407 case KSE_INTR_RESTART:
408 PROC_LOCK(p);
409 mtx_lock_spin(&sched_lock);
410 FOREACH_THREAD_IN_PROC(p, td2) {
411 if (td2->td_mailbox == uap->tmbx)
412 break;
413 }
414 if (td2 == NULL) {
415 mtx_unlock_spin(&sched_lock);
416 PROC_UNLOCK(p);
417 return (ESRCH);
418 }
419 if (uap->cmd == KSE_INTR_SENDSIG) {
420 if (uap->data > 0) {
421 td2->td_flags &= ~TDF_INTERRUPT;
422 mtx_unlock_spin(&sched_lock);
423 tdsignal(td2, (int)uap->data, SIGTARGET_TD);
424 } else {
425 mtx_unlock_spin(&sched_lock);
426 }
427 } else {
428 td2->td_flags |= TDF_INTERRUPT | TDF_ASTPENDING;
429 if (TD_CAN_UNBIND(td2))
430 td2->td_upcall->ku_flags |= KUF_DOUPCALL;
431 if (uap->cmd == KSE_INTR_INTERRUPT)
432 td2->td_intrval = EINTR;
433 else
434 td2->td_intrval = ERESTART;
435 if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR)) {
436 if (td2->td_flags & TDF_CVWAITQ)
437 cv_abort(td2);
438 else
439 abortsleep(td2);
440 }
441 mtx_unlock_spin(&sched_lock);
442 }
443 PROC_UNLOCK(p);
444 break;
445 case KSE_INTR_SIGEXIT:
446 if (uap->data < 1 || uap->data > _SIG_MAXSIG)
447 return (EINVAL);
448 PROC_LOCK(p);
449 sigexit(td, (int)uap->data);
450 break;
451 default:
452 return (EINVAL);
453 }
454 return (0);
455}
456
457/*
458struct kse_exit_args {
459 register_t dummy;
460};
461*/
462int
463kse_exit(struct thread *td, struct kse_exit_args *uap)
464{
465 struct proc *p;
466 struct ksegrp *kg;
467 struct kse *ke;
468 struct kse_upcall *ku, *ku2;
469 int error, count;
470
471 p = td->td_proc;
472 if ((ku = td->td_upcall) == NULL || TD_CAN_UNBIND(td))
473 return (EINVAL);
474 kg = td->td_ksegrp;
475 count = 0;
476 PROC_LOCK(p);
477 mtx_lock_spin(&sched_lock);
478 FOREACH_UPCALL_IN_GROUP(kg, ku2) {
479 if (ku2->ku_flags & KUF_EXITING)
480 count++;
481 }
482 if ((kg->kg_numupcalls - count) == 1 &&
483 (kg->kg_numthreads > 1)) {
484 mtx_unlock_spin(&sched_lock);
485 PROC_UNLOCK(p);
486 return (EDEADLK);
487 }
488 ku->ku_flags |= KUF_EXITING;
489 mtx_unlock_spin(&sched_lock);
490 PROC_UNLOCK(p);
491 error = suword(&ku->ku_mailbox->km_flags, ku->ku_mflags|KMF_DONE);
492 PROC_LOCK(p);
493 if (error)
494 psignal(p, SIGSEGV);
495 mtx_lock_spin(&sched_lock);
496 upcall_remove(td);
497 ke = td->td_kse;
498 if (p->p_numthreads == 1) {
499 kse_purge(p, td);
500 p->p_flag &= ~P_SA;
501 mtx_unlock_spin(&sched_lock);
502 PROC_UNLOCK(p);
503 } else {
504 if (kg->kg_numthreads == 1) { /* Shutdown a group */
505 kse_purge_group(td);
506 ke->ke_flags |= KEF_EXIT;
507 }
508 thread_stopped(p);
509 thread_exit();
510 /* NOTREACHED */
511 }
512 return (0);
513}
514
515/*
516 * Either becomes an upcall or waits for an awakening event and
517 * then becomes an upcall. Only error cases return.
518 */
519/*
520struct kse_release_args {
521 struct timespec *timeout;
522};
523*/
524int
525kse_release(struct thread *td, struct kse_release_args *uap)
526{
527 struct proc *p;
528 struct ksegrp *kg;
529 struct kse_upcall *ku;
530 struct timespec timeout;
531 struct timeval tv;
532 sigset_t sigset;
533 int error;
534
535 p = td->td_proc;
536 kg = td->td_ksegrp;
537 if ((ku = td->td_upcall) == NULL || TD_CAN_UNBIND(td))
538 return (EINVAL);
539 if (uap->timeout != NULL) {
540 if ((error = copyin(uap->timeout, &timeout, sizeof(timeout))))
541 return (error);
542 TIMESPEC_TO_TIMEVAL(&tv, &timeout);
543 }
544 if (td->td_flags & TDF_SA)
545 td->td_pflags |= TDP_UPCALLING;
546 else {
547 ku->ku_mflags = fuword(&ku->ku_mailbox->km_flags);
548 if (ku->ku_mflags == -1) {
549 PROC_LOCK(p);
550 sigexit(td, SIGSEGV);
551 }
552 }
553 PROC_LOCK(p);
554 if (ku->ku_mflags & KMF_WAITSIGEVENT) {
555 /* UTS wants to wait for signal event */
556 if (!(p->p_flag & P_SIGEVENT) && !(ku->ku_flags & KUF_DOUPCALL))
557 error = msleep(&p->p_siglist, &p->p_mtx, PPAUSE|PCATCH,
558 "ksesigwait", (uap->timeout ? tvtohz(&tv) : 0));
559 p->p_flag &= ~P_SIGEVENT;
560 sigset = p->p_siglist;
561 PROC_UNLOCK(p);
562 error = copyout(&sigset, &ku->ku_mailbox->km_sigscaught,
563 sizeof(sigset));
564 } else {
565 if (! kg->kg_completed && !(ku->ku_flags & KUF_DOUPCALL)) {
566 kg->kg_upsleeps++;
567 error = msleep(&kg->kg_completed, &p->p_mtx,
568 PPAUSE|PCATCH, "kserel",
569 (uap->timeout ? tvtohz(&tv) : 0));
570 kg->kg_upsleeps--;
571 }
572 PROC_UNLOCK(p);
573 }
574 if (ku->ku_flags & KUF_DOUPCALL) {
575 mtx_lock_spin(&sched_lock);
576 ku->ku_flags &= ~KUF_DOUPCALL;
577 mtx_unlock_spin(&sched_lock);
578 }
579 return (0);
580}
581
582/* struct kse_wakeup_args {
583 struct kse_mailbox *mbx;
584}; */
585int
586kse_wakeup(struct thread *td, struct kse_wakeup_args *uap)
587{
588 struct proc *p;
589 struct ksegrp *kg;
590 struct kse_upcall *ku;
591 struct thread *td2;
592
593 p = td->td_proc;
594 td2 = NULL;
595 ku = NULL;
596 /* KSE-enabled processes only, please. */
597 if (!(p->p_flag & P_SA))
598 return (EINVAL);
599 PROC_LOCK(p);
600 mtx_lock_spin(&sched_lock);
601 if (uap->mbx) {
602 FOREACH_KSEGRP_IN_PROC(p, kg) {
603 FOREACH_UPCALL_IN_GROUP(kg, ku) {
604 if (ku->ku_mailbox == uap->mbx)
605 break;
606 }
607 if (ku)
608 break;
609 }
610 } else {
611 kg = td->td_ksegrp;
612 if (kg->kg_upsleeps) {
613 wakeup_one(&kg->kg_completed);
614 mtx_unlock_spin(&sched_lock);
615 PROC_UNLOCK(p);
616 return (0);
617 }
618 ku = TAILQ_FIRST(&kg->kg_upcalls);
619 }
620 if (ku) {
621 if ((td2 = ku->ku_owner) == NULL) {
622 panic("%s: no owner", __func__);
623 } else if (TD_ON_SLEEPQ(td2) &&
624 ((td2->td_wchan == &kg->kg_completed) ||
625 (td2->td_wchan == &p->p_siglist &&
626 (ku->ku_mflags & KMF_WAITSIGEVENT)))) {
627 abortsleep(td2);
628 } else {
629 ku->ku_flags |= KUF_DOUPCALL;
630 }
631 mtx_unlock_spin(&sched_lock);
632 PROC_UNLOCK(p);
633 return (0);
634 }
635 mtx_unlock_spin(&sched_lock);
636 PROC_UNLOCK(p);
637 return (ESRCH);
638}
639
640/*
641 * No new KSEG: first call: use current KSE, don't schedule an upcall
642 * All other situations, do allocate max new KSEs and schedule an upcall.
643 */
644/* struct kse_create_args {
645 struct kse_mailbox *mbx;
646 int newgroup;
647}; */
648int
649kse_create(struct thread *td, struct kse_create_args *uap)
650{
651 struct kse *newke;
652 struct ksegrp *newkg;
653 struct ksegrp *kg;
654 struct proc *p;
655 struct kse_mailbox mbx;
656 struct kse_upcall *newku;
657 int err, ncpus, sa = 0, first = 0;
658 struct thread *newtd;
659
660 p = td->td_proc;
661 if ((err = copyin(uap->mbx, &mbx, sizeof(mbx))))
662 return (err);
663
664 /* Too bad, why hasn't kernel always a cpu counter !? */
665#ifdef SMP
666 ncpus = mp_ncpus;
667#else
668 ncpus = 1;
669#endif
670 if (virtual_cpu != 0)
671 ncpus = virtual_cpu;
672 if (!(mbx.km_flags & KMF_BOUND))
673 sa = TDF_SA;
674 else
675 ncpus = 1;
676 PROC_LOCK(p);
677 if (!(p->p_flag & P_SA)) {
678 first = 1;
679 p->p_flag |= P_SA;
680 }
681 PROC_UNLOCK(p);
682 if (!sa && !uap->newgroup && !first)
683 return (EINVAL);
684 kg = td->td_ksegrp;
685 if (uap->newgroup) {
686 /* Have race condition but it is cheap */
687 if (p->p_numksegrps >= max_groups_per_proc)
688 return (EPROCLIM);
689 /*
690 * If we want a new KSEGRP it doesn't matter whether
691 * we have already fired up KSE mode before or not.
692 * We put the process in KSE mode and create a new KSEGRP.
693 */
694 newkg = ksegrp_alloc();
695 bzero(&newkg->kg_startzero, RANGEOF(struct ksegrp,
696 kg_startzero, kg_endzero));
697 bcopy(&kg->kg_startcopy, &newkg->kg_startcopy,
698 RANGEOF(struct ksegrp, kg_startcopy, kg_endcopy));
699 mtx_lock_spin(&sched_lock);
700 if (p->p_numksegrps >= max_groups_per_proc) {
701 mtx_unlock_spin(&sched_lock);
702 ksegrp_free(newkg);
703 return (EPROCLIM);
704 }
705 ksegrp_link(newkg, p);
706 mtx_unlock_spin(&sched_lock);
707 } else {
708 if (!first && ((td->td_flags & TDF_SA) ^ sa) != 0)
709 return (EINVAL);
710 newkg = kg;
711 }
712
713 /*
714 * Creating upcalls more than number of physical cpu does
715 * not help performance.
716 */
717 if (newkg->kg_numupcalls >= ncpus)
718 return (EPROCLIM);
719
720 if (newkg->kg_numupcalls == 0) {
721 /*
722 * Initialize KSE group
723 *
724 * For multiplxed group, create KSEs as many as physical
725 * cpus. This increases concurrent even if userland
726 * is not MP safe and can only run on single CPU.
727 * In ideal world, every physical cpu should execute a thread.
728 * If there is enough KSEs, threads in kernel can be
729 * executed parallel on different cpus with full speed,
730 * Concurrent in kernel shouldn't be restricted by number of
731 * upcalls userland provides. Adding more upcall structures
732 * only increases concurrent in userland.
733 *
734 * For bound thread group, because there is only thread in the
735 * group, we only create one KSE for the group. Thread in this
736 * kind of group will never schedule an upcall when blocked,
737 * this intends to simulate pthread system scope thread.
738 */
739 while (newkg->kg_kses < ncpus) {
740 newke = kse_alloc();
741 bzero(&newke->ke_startzero, RANGEOF(struct kse,
742 ke_startzero, ke_endzero));
743#if 0
744 mtx_lock_spin(&sched_lock);
745 bcopy(&ke->ke_startcopy, &newke->ke_startcopy,
746 RANGEOF(struct kse, ke_startcopy, ke_endcopy));
747 mtx_unlock_spin(&sched_lock);
748#endif
749 mtx_lock_spin(&sched_lock);
750 kse_link(newke, newkg);
751 /* Add engine */
752 kse_reassign(newke);
753 mtx_unlock_spin(&sched_lock);
754 }
755 }
756 newku = upcall_alloc();
757 newku->ku_mailbox = uap->mbx;
758 newku->ku_func = mbx.km_func;
759 bcopy(&mbx.km_stack, &newku->ku_stack, sizeof(stack_t));
760
761 /* For the first call this may not have been set */
762 if (td->td_standin == NULL)
763 thread_alloc_spare(td, NULL);
764
765 PROC_LOCK(p);
766 if (newkg->kg_numupcalls >= ncpus) {
767 PROC_UNLOCK(p);
768 upcall_free(newku);
769 return (EPROCLIM);
770 }
771 if (first && sa) {
772 SIGSETOR(p->p_siglist, td->td_siglist);
773 SIGEMPTYSET(td->td_siglist);
774 SIGFILLSET(td->td_sigmask);
775 SIG_CANTMASK(td->td_sigmask);
776 }
777 mtx_lock_spin(&sched_lock);
778 PROC_UNLOCK(p);
779 upcall_link(newku, newkg);
780 if (mbx.km_quantum)
781 newkg->kg_upquantum = max(1, mbx.km_quantum/tick);
782
783 /*
784 * Each upcall structure has an owner thread, find which
785 * one owns it.
786 */
787 if (uap->newgroup) {
788 /*
789 * Because new ksegrp hasn't thread,
790 * create an initial upcall thread to own it.
791 */
792 newtd = thread_schedule_upcall(td, newku);
793 } else {
794 /*
795 * If current thread hasn't an upcall structure,
796 * just assign the upcall to it.
797 */
798 if (td->td_upcall == NULL) {
799 newku->ku_owner = td;
800 td->td_upcall = newku;
801 newtd = td;
802 } else {
803 /*
804 * Create a new upcall thread to own it.
805 */
806 newtd = thread_schedule_upcall(td, newku);
807 }
808 }
809 if (!sa) {
810 newtd->td_mailbox = mbx.km_curthread;
811 newtd->td_flags &= ~TDF_SA;
812 if (newtd != td) {
813 mtx_unlock_spin(&sched_lock);
814 cpu_set_upcall_kse(newtd, newku);
815 mtx_lock_spin(&sched_lock);
816 }
817 } else {
818 newtd->td_flags |= TDF_SA;
819 }
820 if (newtd != td)
821 setrunqueue(newtd);
822 mtx_unlock_spin(&sched_lock);
823 return (0);
824}
825
826/*
827 * Initialize global thread allocation resources.
828 */
829void
830threadinit(void)
831{
832
833 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
834 thread_ctor, thread_dtor, thread_init, thread_fini,
835 UMA_ALIGN_CACHE, 0);
836 ksegrp_zone = uma_zcreate("KSEGRP", sched_sizeof_ksegrp(),
837 NULL, NULL, ksegrp_init, NULL,
838 UMA_ALIGN_CACHE, 0);
839 kse_zone = uma_zcreate("KSE", sched_sizeof_kse(),
840 NULL, NULL, kse_init, NULL,
841 UMA_ALIGN_CACHE, 0);
842 upcall_zone = uma_zcreate("UPCALL", sizeof(struct kse_upcall),
843 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
844}
845
846/*
847 * Stash an embarasingly extra thread into the zombie thread queue.
848 */
849void
850thread_stash(struct thread *td)
851{
852 mtx_lock_spin(&kse_zombie_lock);
853 TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq);
854 mtx_unlock_spin(&kse_zombie_lock);
855}
856
857/*
858 * Stash an embarasingly extra kse into the zombie kse queue.
859 */
860void
861kse_stash(struct kse *ke)
862{
863 mtx_lock_spin(&kse_zombie_lock);
864 TAILQ_INSERT_HEAD(&zombie_kses, ke, ke_procq);
865 mtx_unlock_spin(&kse_zombie_lock);
866}
867
868/*
869 * Stash an embarasingly extra upcall into the zombie upcall queue.
870 */
871
872void
873upcall_stash(struct kse_upcall *ku)
874{
875 mtx_lock_spin(&kse_zombie_lock);
876 TAILQ_INSERT_HEAD(&zombie_upcalls, ku, ku_link);
877 mtx_unlock_spin(&kse_zombie_lock);
878}
879
880/*
881 * Stash an embarasingly extra ksegrp into the zombie ksegrp queue.
882 */
883void
884ksegrp_stash(struct ksegrp *kg)
885{
886 mtx_lock_spin(&kse_zombie_lock);
887 TAILQ_INSERT_HEAD(&zombie_ksegrps, kg, kg_ksegrp);
888 mtx_unlock_spin(&kse_zombie_lock);
889}
890
891/*
892 * Reap zombie kse resource.
893 */
894void
895thread_reap(void)
896{
897 struct thread *td_first, *td_next;
898 struct kse *ke_first, *ke_next;
899 struct ksegrp *kg_first, * kg_next;
900 struct kse_upcall *ku_first, *ku_next;
901
902 /*
903 * Don't even bother to lock if none at this instant,
904 * we really don't care about the next instant..
905 */
906 if ((!TAILQ_EMPTY(&zombie_threads))
907 || (!TAILQ_EMPTY(&zombie_kses))
908 || (!TAILQ_EMPTY(&zombie_ksegrps))
909 || (!TAILQ_EMPTY(&zombie_upcalls))) {
910 mtx_lock_spin(&kse_zombie_lock);
911 td_first = TAILQ_FIRST(&zombie_threads);
912 ke_first = TAILQ_FIRST(&zombie_kses);
913 kg_first = TAILQ_FIRST(&zombie_ksegrps);
914 ku_first = TAILQ_FIRST(&zombie_upcalls);
915 if (td_first)
916 TAILQ_INIT(&zombie_threads);
917 if (ke_first)
918 TAILQ_INIT(&zombie_kses);
919 if (kg_first)
920 TAILQ_INIT(&zombie_ksegrps);
921 if (ku_first)
922 TAILQ_INIT(&zombie_upcalls);
923 mtx_unlock_spin(&kse_zombie_lock);
924 while (td_first) {
925 td_next = TAILQ_NEXT(td_first, td_runq);
926 if (td_first->td_ucred)
927 crfree(td_first->td_ucred);
928 thread_free(td_first);
929 td_first = td_next;
930 }
931 while (ke_first) {
932 ke_next = TAILQ_NEXT(ke_first, ke_procq);
933 kse_free(ke_first);
934 ke_first = ke_next;
935 }
936 while (kg_first) {
937 kg_next = TAILQ_NEXT(kg_first, kg_ksegrp);
938 ksegrp_free(kg_first);
939 kg_first = kg_next;
940 }
941 while (ku_first) {
942 ku_next = TAILQ_NEXT(ku_first, ku_link);
943 upcall_free(ku_first);
944 ku_first = ku_next;
945 }
946 }
947}
948
949/*
950 * Allocate a ksegrp.
951 */
952struct ksegrp *
953ksegrp_alloc(void)
954{
955 return (uma_zalloc(ksegrp_zone, M_WAITOK));
956}
957
958/*
959 * Allocate a kse.
960 */
961struct kse *
962kse_alloc(void)
963{
964 return (uma_zalloc(kse_zone, M_WAITOK));
965}
966
967/*
968 * Allocate a thread.
969 */
970struct thread *
971thread_alloc(void)
972{
973 thread_reap(); /* check if any zombies to get */
974 return (uma_zalloc(thread_zone, M_WAITOK));
975}
976
977/*
978 * Deallocate a ksegrp.
979 */
980void
981ksegrp_free(struct ksegrp *td)
982{
983 uma_zfree(ksegrp_zone, td);
984}
985
986/*
987 * Deallocate a kse.
988 */
989void
990kse_free(struct kse *td)
991{
992 uma_zfree(kse_zone, td);
993}
994
995/*
996 * Deallocate a thread.
997 */
998void
999thread_free(struct thread *td)
1000{
1001
1002 cpu_thread_clean(td);
1003 uma_zfree(thread_zone, td);
1004}
1005
1006/*
1007 * Store the thread context in the UTS's mailbox.
1008 * then add the mailbox at the head of a list we are building in user space.
1009 * The list is anchored in the ksegrp structure.
1010 */
1011int
1012thread_export_context(struct thread *td, int willexit)
1013{
1014 struct proc *p;
1015 struct ksegrp *kg;
1016 uintptr_t mbx;
1017 void *addr;
1018 int error = 0, temp, sig;
1019 mcontext_t mc;
1020
1021 p = td->td_proc;
1022 kg = td->td_ksegrp;
1023
1024 /* Export the user/machine context. */
1025 get_mcontext(td, &mc, 0);
1026 addr = (void *)(&td->td_mailbox->tm_context.uc_mcontext);
1027 error = copyout(&mc, addr, sizeof(mcontext_t));
1028 if (error)
1029 goto bad;
1030
1031 /* Exports clock ticks in kernel mode */
1032 addr = (caddr_t)(&td->td_mailbox->tm_sticks);
1033 temp = fuword32(addr) + td->td_usticks;
1034 if (suword32(addr, temp)) {
1035 error = EFAULT;
1036 goto bad;
1037 }
1038
1039 /*
1040 * Post sync signal, or process SIGKILL and SIGSTOP.
1041 * For sync signal, it is only possible when the signal is not
1042 * caught by userland or process is being debugged.
1043 */
1044 PROC_LOCK(p);
1045 if (td->td_flags & TDF_NEEDSIGCHK) {
1046 mtx_lock_spin(&sched_lock);
1047 td->td_flags &= ~TDF_NEEDSIGCHK;
1048 mtx_unlock_spin(&sched_lock);
1049 mtx_lock(&p->p_sigacts->ps_mtx);
1050 while ((sig = cursig(td)) != 0)
1051 postsig(sig);
1052 mtx_unlock(&p->p_sigacts->ps_mtx);
1053 }
1054 if (willexit)
1055 SIGFILLSET(td->td_sigmask);
1056 PROC_UNLOCK(p);
1057
1058 /* Get address in latest mbox of list pointer */
1059 addr = (void *)(&td->td_mailbox->tm_next);
1060 /*
1061 * Put the saved address of the previous first
1062 * entry into this one
1063 */
1064 for (;;) {
1065 mbx = (uintptr_t)kg->kg_completed;
1066 if (suword(addr, mbx)) {
1067 error = EFAULT;
1068 goto bad;
1069 }
1070 PROC_LOCK(p);
1071 if (mbx == (uintptr_t)kg->kg_completed) {
1072 kg->kg_completed = td->td_mailbox;
1073 /*
1074 * The thread context may be taken away by
1075 * other upcall threads when we unlock
1076 * process lock. it's no longer valid to
1077 * use it again in any other places.
1078 */
1079 td->td_mailbox = NULL;
1080 PROC_UNLOCK(p);
1081 break;
1082 }
1083 PROC_UNLOCK(p);
1084 }
1085 td->td_usticks = 0;
1086 return (0);
1087
1088bad:
1089 PROC_LOCK(p);
1090 sigexit(td, SIGILL);
1091 return (error);
1092}
1093
1094/*
1095 * Take the list of completed mailboxes for this KSEGRP and put them on this
1096 * upcall's mailbox as it's the next one going up.
1097 */
1098static int
1099thread_link_mboxes(struct ksegrp *kg, struct kse_upcall *ku)
1100{
1101 struct proc *p = kg->kg_proc;
1102 void *addr;
1103 uintptr_t mbx;
1104
1105 addr = (void *)(&ku->ku_mailbox->km_completed);
1106 for (;;) {
1107 mbx = (uintptr_t)kg->kg_completed;
1108 if (suword(addr, mbx)) {
1109 PROC_LOCK(p);
1110 psignal(p, SIGSEGV);
1111 PROC_UNLOCK(p);
1112 return (EFAULT);
1113 }
1114 PROC_LOCK(p);
1115 if (mbx == (uintptr_t)kg->kg_completed) {
1116 kg->kg_completed = NULL;
1117 PROC_UNLOCK(p);
1118 break;
1119 }
1120 PROC_UNLOCK(p);
1121 }
1122 return (0);
1123}
1124
1125/*
1126 * This function should be called at statclock interrupt time
1127 */
1128int
1129thread_statclock(int user)
1130{
1131 struct thread *td = curthread;
1132 struct ksegrp *kg = td->td_ksegrp;
1133
1134 if (kg->kg_numupcalls == 0 || !(td->td_flags & TDF_SA))
1135 return (0);
1136 if (user) {
1137 /* Current always do via ast() */
1138 mtx_lock_spin(&sched_lock);
1139 td->td_flags |= (TDF_USTATCLOCK|TDF_ASTPENDING);
1140 mtx_unlock_spin(&sched_lock);
1141 td->td_uuticks++;
1142 } else {
1143 if (td->td_mailbox != NULL)
1144 td->td_usticks++;
1145 else {
1146 /* XXXKSE
1147 * We will call thread_user_enter() for every
1148 * kernel entry in future, so if the thread mailbox
1149 * is NULL, it must be a UTS kernel, don't account
1150 * clock ticks for it.
1151 */
1152 }
1153 }
1154 return (0);
1155}
1156
1157/*
1158 * Export state clock ticks for userland
1159 */
1160static int
1161thread_update_usr_ticks(struct thread *td, int user)
1162{
1163 struct proc *p = td->td_proc;
1164 struct kse_thr_mailbox *tmbx;
1165 struct kse_upcall *ku;
1166 struct ksegrp *kg;
1167 caddr_t addr;
1168 uint uticks;
1169
1170 if ((ku = td->td_upcall) == NULL)
1171 return (-1);
1172
1173 tmbx = (void *)fuword((void *)&ku->ku_mailbox->km_curthread);
1174 if ((tmbx == NULL) || (tmbx == (void *)-1))
1175 return (-1);
1176 if (user) {
1177 uticks = td->td_uuticks;
1178 td->td_uuticks = 0;
1179 addr = (caddr_t)&tmbx->tm_uticks;
1180 } else {
1181 uticks = td->td_usticks;
1182 td->td_usticks = 0;
1183 addr = (caddr_t)&tmbx->tm_sticks;
1184 }
1185 if (uticks) {
1186 if (suword32(addr, uticks+fuword32(addr))) {
1187 PROC_LOCK(p);
1188 psignal(p, SIGSEGV);
1189 PROC_UNLOCK(p);
1190 return (-2);
1191 }
1192 }
1193 kg = td->td_ksegrp;
1194 if (kg->kg_upquantum && ticks >= kg->kg_nextupcall) {
1195 mtx_lock_spin(&sched_lock);
1196 td->td_upcall->ku_flags |= KUF_DOUPCALL;
1197 mtx_unlock_spin(&sched_lock);
1198 }
1199 return (0);
1200}
1201
1202/*
1203 * Discard the current thread and exit from its context.
1204 *
1205 * Because we can't free a thread while we're operating under its context,
1206 * push the current thread into our CPU's deadthread holder. This means
1207 * we needn't worry about someone else grabbing our context before we
1208 * do a cpu_throw().
1209 */
1210void
1211thread_exit(void)
1212{
1213 struct thread *td;
1214 struct kse *ke;
1215 struct proc *p;
1216 struct ksegrp *kg;
1217
1218 td = curthread;
1219 kg = td->td_ksegrp;
1220 p = td->td_proc;
1221 ke = td->td_kse;
1222
1223 mtx_assert(&sched_lock, MA_OWNED);
1224 KASSERT(p != NULL, ("thread exiting without a process"));
1225 KASSERT(ke != NULL, ("thread exiting without a kse"));
1226 KASSERT(kg != NULL, ("thread exiting without a kse group"));
1227 PROC_LOCK_ASSERT(p, MA_OWNED);
1228 CTR1(KTR_PROC, "thread_exit: thread %p", td);
1229 KASSERT(!mtx_owned(&Giant), ("dying thread owns giant"));
1230
1231 if (td->td_standin != NULL) {
1232 thread_stash(td->td_standin);
1233 td->td_standin = NULL;
1234 }
1235
1236 cpu_thread_exit(td); /* XXXSMP */
1237
1238 /*
1239 * The last thread is left attached to the process
1240 * So that the whole bundle gets recycled. Skip
1241 * all this stuff.
1242 */
1243 if (p->p_numthreads > 1) {
1244 thread_unlink(td);
1245 if (p->p_maxthrwaits)
1246 wakeup(&p->p_numthreads);
1247 /*
1248 * The test below is NOT true if we are the
1249 * sole exiting thread. P_STOPPED_SNGL is unset
1250 * in exit1() after it is the only survivor.
1251 */
1252 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1253 if (p->p_numthreads == p->p_suspcount) {
1254 thread_unsuspend_one(p->p_singlethread);
1255 }
1256 }
1257
1258 /*
1259 * Because each upcall structure has an owner thread,
1260 * owner thread exits only when process is in exiting
1261 * state, so upcall to userland is no longer needed,
1262 * deleting upcall structure is safe here.
1263 * So when all threads in a group is exited, all upcalls
1264 * in the group should be automatically freed.
1265 */
1266 if (td->td_upcall)
1267 upcall_remove(td);
1268
1269 ke->ke_state = KES_UNQUEUED;
1270 ke->ke_thread = NULL;
1271 /*
1272 * Decide what to do with the KSE attached to this thread.
1273 */
1274 if (ke->ke_flags & KEF_EXIT)
1275 kse_unlink(ke);
1276 else
1277 kse_reassign(ke);
1278 PROC_UNLOCK(p);
1279 td->td_kse = NULL;
1280 td->td_state = TDS_INACTIVE;
1281#if 0
1282 td->td_proc = NULL;
1283#endif
1284 td->td_ksegrp = NULL;
1285 td->td_last_kse = NULL;
1286 PCPU_SET(deadthread, td);
1287 } else {
1288 PROC_UNLOCK(p);
1289 }
1290 /* XXX Shouldn't cpu_throw() here. */
1291 mtx_assert(&sched_lock, MA_OWNED);
1292#if !defined(__alpha__) && !defined(__powerpc__)
1293 cpu_throw(td, choosethread());
1294#else
1295 cpu_throw();
1296#endif
1297 panic("I'm a teapot!");
1298 /* NOTREACHED */
1299}
1300
1301/*
1302 * Do any thread specific cleanups that may be needed in wait()
1303 * called with Giant held, proc and schedlock not held.
1304 */
1305void
1306thread_wait(struct proc *p)
1307{
1308 struct thread *td;
1309
1310 KASSERT((p->p_numthreads == 1), ("Muliple threads in wait1()"));
1311 KASSERT((p->p_numksegrps == 1), ("Muliple ksegrps in wait1()"));
1312 FOREACH_THREAD_IN_PROC(p, td) {
1313 if (td->td_standin != NULL) {
1314 thread_free(td->td_standin);
1315 td->td_standin = NULL;
1316 }
1317 cpu_thread_clean(td);
1318 }
1319 thread_reap(); /* check for zombie threads etc. */
1320}
1321
1322/*
1323 * Link a thread to a process.
1324 * set up anything that needs to be initialized for it to
1325 * be used by the process.
1326 *
1327 * Note that we do not link to the proc's ucred here.
1328 * The thread is linked as if running but no KSE assigned.
1329 */
1330void
1331thread_link(struct thread *td, struct ksegrp *kg)
1332{
1333 struct proc *p;
1334
1335 p = kg->kg_proc;
1336 td->td_state = TDS_INACTIVE;
1337 td->td_proc = p;
1338 td->td_ksegrp = kg;
1339 td->td_last_kse = NULL;
1340 td->td_flags = 0;
1341 td->td_kse = NULL;
1342
1343 LIST_INIT(&td->td_contested);
1344 callout_init(&td->td_slpcallout, 1);
1345 TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
1346 TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist);
1347 p->p_numthreads++;
1348 kg->kg_numthreads++;
1349}
1350
1351void
1352thread_unlink(struct thread *td)
1353{
1354 struct proc *p = td->td_proc;
1355 struct ksegrp *kg = td->td_ksegrp;
1356
1357 mtx_assert(&sched_lock, MA_OWNED);
1358 TAILQ_REMOVE(&p->p_threads, td, td_plist);
1359 p->p_numthreads--;
1360 TAILQ_REMOVE(&kg->kg_threads, td, td_kglist);
1361 kg->kg_numthreads--;
1362 /* could clear a few other things here */
1363}
1364
1365/*
1366 * Purge a ksegrp resource. When a ksegrp is preparing to
1367 * exit, it calls this function.
1368 */
1369static void
1370kse_purge_group(struct thread *td)
1371{
1372 struct ksegrp *kg;
1373 struct kse *ke;
1374
1375 kg = td->td_ksegrp;
1376 KASSERT(kg->kg_numthreads == 1, ("%s: bad thread number", __func__));
1377 while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) {
1378 KASSERT(ke->ke_state == KES_IDLE,
1379 ("%s: wrong idle KSE state", __func__));
1380 kse_unlink(ke);
1381 }
1382 KASSERT((kg->kg_kses == 1),
1383 ("%s: ksegrp still has %d KSEs", __func__, kg->kg_kses));
1384 KASSERT((kg->kg_numupcalls == 0),
1385 ("%s: ksegrp still has %d upcall datas",
1386 __func__, kg->kg_numupcalls));
1387}
1388
1389/*
1390 * Purge a process's KSE resource. When a process is preparing to
1391 * exit, it calls kse_purge to release any extra KSE resources in
1392 * the process.
1393 */
1394static void
1395kse_purge(struct proc *p, struct thread *td)
1396{
1397 struct ksegrp *kg;
1398 struct kse *ke;
1399
1400 KASSERT(p->p_numthreads == 1, ("bad thread number"));
1401 while ((kg = TAILQ_FIRST(&p->p_ksegrps)) != NULL) {
1402 TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
1403 p->p_numksegrps--;
1404 /*
1405 * There is no ownership for KSE, after all threads
1406 * in the group exited, it is possible that some KSEs
1407 * were left in idle queue, gc them now.
1408 */
1409 while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) {
1410 KASSERT(ke->ke_state == KES_IDLE,
1411 ("%s: wrong idle KSE state", __func__));
1412 TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist);
1413 kg->kg_idle_kses--;
1414 TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist);
1415 kg->kg_kses--;
1416 kse_stash(ke);
1417 }
1418 KASSERT(((kg->kg_kses == 0) && (kg != td->td_ksegrp)) ||
1419 ((kg->kg_kses == 1) && (kg == td->td_ksegrp)),
1420 ("ksegrp has wrong kg_kses: %d", kg->kg_kses));
1421 KASSERT((kg->kg_numupcalls == 0),
1422 ("%s: ksegrp still has %d upcall datas",
1423 __func__, kg->kg_numupcalls));
1424
1425 if (kg != td->td_ksegrp)
1426 ksegrp_stash(kg);
1427 }
1428 TAILQ_INSERT_HEAD(&p->p_ksegrps, td->td_ksegrp, kg_ksegrp);
1429 p->p_numksegrps++;
1430}
1431
1432/*
1433 * This function is intended to be used to initialize a spare thread
1434 * for upcall. Initialize thread's large data area outside sched_lock
1435 * for thread_schedule_upcall().
1436 */
1437void
1438thread_alloc_spare(struct thread *td, struct thread *spare)
1439{
1440 if (td->td_standin)
1441 return;
1442 if (spare == NULL)
1443 spare = thread_alloc();
1444 td->td_standin = spare;
1445 bzero(&spare->td_startzero,
1446 (unsigned)RANGEOF(struct thread, td_startzero, td_endzero));
1447 spare->td_proc = td->td_proc;
1448 spare->td_ucred = crhold(td->td_ucred);
1449}
1450
1451/*
1452 * Create a thread and schedule it for upcall on the KSE given.
1453 * Use our thread's standin so that we don't have to allocate one.
1454 */
1455struct thread *
1456thread_schedule_upcall(struct thread *td, struct kse_upcall *ku)
1457{
1458 struct thread *td2;
1459
1460 mtx_assert(&sched_lock, MA_OWNED);
1461
1462 /*
1463 * Schedule an upcall thread on specified kse_upcall,
1464 * the kse_upcall must be free.
1465 * td must have a spare thread.
1466 */
1467 KASSERT(ku->ku_owner == NULL, ("%s: upcall has owner", __func__));
1468 if ((td2 = td->td_standin) != NULL) {
1469 td->td_standin = NULL;
1470 } else {
1471 panic("no reserve thread when scheduling an upcall");
1472 return (NULL);
1473 }
1474 CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)",
1475 td2, td->td_proc->p_pid, td->td_proc->p_comm);
1476 bcopy(&td->td_startcopy, &td2->td_startcopy,
1477 (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy));
1478 thread_link(td2, ku->ku_ksegrp);
1479 /* inherit blocked thread's context */
1480 cpu_set_upcall(td2, td);
1481 /* Let the new thread become owner of the upcall */
1482 ku->ku_owner = td2;
1483 td2->td_upcall = ku;
1484 td2->td_flags = TDF_SA;
1485 td2->td_pflags = TDP_UPCALLING;
1486 td2->td_kse = NULL;
1487 td2->td_state = TDS_CAN_RUN;
1488 td2->td_inhibitors = 0;
1489 SIGFILLSET(td2->td_sigmask);
1490 SIG_CANTMASK(td2->td_sigmask);
1491 return (td2); /* bogus.. should be a void function */
1492}
1493
1494/*
1495 * It is only used when thread generated a trap and process is being
1496 * debugged.
1497 */
1498void
1499thread_signal_add(struct thread *td, int sig)
1500{
1501 struct proc *p;
1502 siginfo_t siginfo;
1503 struct sigacts *ps;
1504 int error;
1505
1506 p = td->td_proc;
1507 PROC_LOCK_ASSERT(p, MA_OWNED);
1508 ps = p->p_sigacts;
1509 mtx_assert(&ps->ps_mtx, MA_OWNED);
1510
1511 cpu_thread_siginfo(sig, 0, &siginfo);
1512 mtx_unlock(&ps->ps_mtx);
1513 PROC_UNLOCK(p);
1514 error = copyout(&siginfo, &td->td_mailbox->tm_syncsig, sizeof(siginfo));
1515 if (error) {
1516 PROC_LOCK(p);
1517 sigexit(td, SIGILL);
1518 }
1519 PROC_LOCK(p);
1520 SIGADDSET(td->td_sigmask, sig);
1521 mtx_lock(&ps->ps_mtx);
1522}
1523
1524void
1525thread_switchout(struct thread *td)
1526{
1527 struct kse_upcall *ku;
1528 struct thread *td2;
1529
1530 mtx_assert(&sched_lock, MA_OWNED);
1531
1532 /*
1533 * If the outgoing thread is in threaded group and has never
1534 * scheduled an upcall, decide whether this is a short
1535 * or long term event and thus whether or not to schedule
1536 * an upcall.
1537 * If it is a short term event, just suspend it in
1538 * a way that takes its KSE with it.
1539 * Select the events for which we want to schedule upcalls.
1540 * For now it's just sleep.
1541 * XXXKSE eventually almost any inhibition could do.
1542 */
1543 if (TD_CAN_UNBIND(td) && (td->td_standin) && TD_ON_SLEEPQ(td)) {
1544 /*
1545 * Release ownership of upcall, and schedule an upcall
1546 * thread, this new upcall thread becomes the owner of
1547 * the upcall structure.
1548 */
1549 ku = td->td_upcall;
1550 ku->ku_owner = NULL;
1551 td->td_upcall = NULL;
1552 td->td_flags &= ~TDF_CAN_UNBIND;
1553 td2 = thread_schedule_upcall(td, ku);
1554 setrunqueue(td2);
1555 }
1556}
1557
1558/*
1559 * Setup done on the thread when it enters the kernel.
1560 * XXXKSE Presently only for syscalls but eventually all kernel entries.
1561 */
1562void
1563thread_user_enter(struct proc *p, struct thread *td)
1564{
1565 struct ksegrp *kg;
1566 struct kse_upcall *ku;
1567 struct kse_thr_mailbox *tmbx;
1568 uint32_t tflags;
1569
1570 kg = td->td_ksegrp;
1571
1572 /*
1573 * First check that we shouldn't just abort.
1574 * But check if we are the single thread first!
1575 */
1576 if (p->p_flag & P_SINGLE_EXIT) {
1577 PROC_LOCK(p);
1578 mtx_lock_spin(&sched_lock);
1579 thread_stopped(p);
1580 thread_exit();
1581 /* NOTREACHED */
1582 }
1583
1584 /*
1585 * If we are doing a syscall in a KSE environment,
1586 * note where our mailbox is. There is always the
1587 * possibility that we could do this lazily (in kse_reassign()),
1588 * but for now do it every time.
1589 */
1590 kg = td->td_ksegrp;
1591 if (td->td_flags & TDF_SA) {
1592 ku = td->td_upcall;
1593 KASSERT(ku, ("%s: no upcall owned", __func__));
1594 KASSERT((ku->ku_owner == td), ("%s: wrong owner", __func__));
1595 KASSERT(!TD_CAN_UNBIND(td), ("%s: can unbind", __func__));
1596 ku->ku_mflags = fuword32((void *)&ku->ku_mailbox->km_flags);
1597 tmbx = (void *)fuword((void *)&ku->ku_mailbox->km_curthread);
1598 if ((tmbx == NULL) || (tmbx == (void *)-1L) ||
1599 (ku->ku_mflags & KMF_NOUPCALL)) {
1600 td->td_mailbox = NULL;
1601 } else {
1602 if (td->td_standin == NULL)
1603 thread_alloc_spare(td, NULL);
1604 tflags = fuword32(tmbx);
1605 /*
1606 * On some architectures, TP register points to thread
1607 * mailbox but not points to kse mailbox, and userland
1608 * can not atomically clear km_curthread, but can
1609 * use TP register, and set TMF_NOUPCALL in thread
1610 * flag to indicate a critical region.
1611 */
1612 if (tflags & TMF_NOUPCALL) {
1613 td->td_mailbox = NULL;
1614 } else {
1615 td->td_mailbox = tmbx;
1616 mtx_lock_spin(&sched_lock);
1617 td->td_flags |= TDF_CAN_UNBIND;
1618 mtx_unlock_spin(&sched_lock);
1619 }
1620 }
1621 }
1622}
1623
1624/*
1625 * The extra work we go through if we are a threaded process when we
1626 * return to userland.
1627 *
1628 * If we are a KSE process and returning to user mode, check for
1629 * extra work to do before we return (e.g. for more syscalls
1630 * to complete first). If we were in a critical section, we should
1631 * just return to let it finish. Same if we were in the UTS (in
1632 * which case the mailbox's context's busy indicator will be set).
1633 * The only traps we suport will have set the mailbox.
1634 * We will clear it here.
1635 */
1636int
1637thread_userret(struct thread *td, struct trapframe *frame)
1638{
1639 int error = 0, upcalls, uts_crit;
1640 struct kse_upcall *ku;
1641 struct ksegrp *kg, *kg2;
1642 struct proc *p;
1643 struct timespec ts;
1644
1645 p = td->td_proc;
1646 kg = td->td_ksegrp;
1647 ku = td->td_upcall;
1648
1649 /* Nothing to do with bound thread */
1650 if (!(td->td_flags & TDF_SA))
1651 return (0);
1652
1653 /*
1654 * Stat clock interrupt hit in userland, it
1655 * is returning from interrupt, charge thread's
1656 * userland time for UTS.
1657 */
1658 if (td->td_flags & TDF_USTATCLOCK) {
1659 thread_update_usr_ticks(td, 1);
1660 mtx_lock_spin(&sched_lock);
1661 td->td_flags &= ~TDF_USTATCLOCK;
1662 mtx_unlock_spin(&sched_lock);
1663 if (kg->kg_completed ||
1664 (td->td_upcall->ku_flags & KUF_DOUPCALL))
1665 thread_user_enter(p, td);
1666 }
1667
1668 uts_crit = (td->td_mailbox == NULL);
1669 /*
1670 * Optimisation:
1671 * This thread has not started any upcall.
1672 * If there is no work to report other than ourself,
1673 * then it can return direct to userland.
1674 */
1675 if (TD_CAN_UNBIND(td)) {
1676 mtx_lock_spin(&sched_lock);
1677 td->td_flags &= ~TDF_CAN_UNBIND;
1678 if ((td->td_flags & TDF_NEEDSIGCHK) == 0 &&
1679 (kg->kg_completed == NULL) &&
1680 (ku->ku_flags & KUF_DOUPCALL) == 0 &&
1681 (kg->kg_upquantum && ticks < kg->kg_nextupcall)) {
1682 mtx_unlock_spin(&sched_lock);
1683 thread_update_usr_ticks(td, 0);
1684 nanotime(&ts);
1685 error = copyout(&ts,
1686 (caddr_t)&ku->ku_mailbox->km_timeofday,
1687 sizeof(ts));
1688 td->td_mailbox = 0;
1689 ku->ku_mflags = 0;
1690 if (error)
1691 goto out;
1692 return (0);
1693 }
1694 mtx_unlock_spin(&sched_lock);
1695 thread_export_context(td, 0);
1696 /*
1697 * There is something to report, and we own an upcall
1698 * strucuture, we can go to userland.
1699 * Turn ourself into an upcall thread.
1700 */
1701 td->td_pflags |= TDP_UPCALLING;
1702 } else if (td->td_mailbox && (ku == NULL)) {
1703 thread_export_context(td, 1);
1704 PROC_LOCK(p);
1705 /*
1706 * There are upcall threads waiting for
1707 * work to do, wake one of them up.
1708 * XXXKSE Maybe wake all of them up.
1709 */
1710 if (kg->kg_upsleeps)
1711 wakeup_one(&kg->kg_completed);
1712 mtx_lock_spin(&sched_lock);
1713 thread_stopped(p);
1714 thread_exit();
1715 /* NOTREACHED */
1716 }
1717
1718 KASSERT(ku != NULL, ("upcall is NULL\n"));
1719 KASSERT(TD_CAN_UNBIND(td) == 0, ("can unbind"));
1720
1721 if (p->p_numthreads > max_threads_per_proc) {
1722 max_threads_hits++;
1723 PROC_LOCK(p);
1724 mtx_lock_spin(&sched_lock);
1725 p->p_maxthrwaits++;
1726 while (p->p_numthreads > max_threads_per_proc) {
1727 upcalls = 0;
1728 FOREACH_KSEGRP_IN_PROC(p, kg2) {
1729 if (kg2->kg_numupcalls == 0)
1730 upcalls++;
1731 else
1732 upcalls += kg2->kg_numupcalls;
1733 }
1734 if (upcalls >= max_threads_per_proc)
1735 break;
1736 mtx_unlock_spin(&sched_lock);
1737 if (msleep(&p->p_numthreads, &p->p_mtx, PPAUSE|PCATCH,
1738 "maxthreads", NULL)) {
1739 mtx_lock_spin(&sched_lock);
1740 break;
1741 } else {
1742 mtx_lock_spin(&sched_lock);
1743 }
1744 }
1745 p->p_maxthrwaits--;
1746 mtx_unlock_spin(&sched_lock);
1747 PROC_UNLOCK(p);
1748 }
1749
1750 if (td->td_pflags & TDP_UPCALLING) {
1751 uts_crit = 0;
1752 kg->kg_nextupcall = ticks+kg->kg_upquantum;
1753 /*
1754 * There is no more work to do and we are going to ride
1755 * this thread up to userland as an upcall.
1756 * Do the last parts of the setup needed for the upcall.
1757 */
1758 CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)",
1759 td, td->td_proc->p_pid, td->td_proc->p_comm);
1760
1761 td->td_pflags &= ~TDP_UPCALLING;
1762 if (ku->ku_flags & KUF_DOUPCALL) {
1763 mtx_lock_spin(&sched_lock);
1764 ku->ku_flags &= ~KUF_DOUPCALL;
1765 mtx_unlock_spin(&sched_lock);
1766 }
1767 /*
1768 * Set user context to the UTS
1769 */
1770 if (!(ku->ku_mflags & KMF_NOUPCALL)) {
1771 cpu_set_upcall_kse(td, ku);
1772 error = suword(&ku->ku_mailbox->km_curthread, 0);
1773 if (error)
1774 goto out;
1775 }
1776
1777 /*
1778 * Unhook the list of completed threads.
1779 * anything that completes after this gets to
1780 * come in next time.
1781 * Put the list of completed thread mailboxes on
1782 * this KSE's mailbox.
1783 */
1784 if (!(ku->ku_mflags & KMF_NOCOMPLETED) &&
1785 (error = thread_link_mboxes(kg, ku)) != 0)
1786 goto out;
1787 }
1788 if (!uts_crit) {
1789 nanotime(&ts);
1790 error = copyout(&ts, &ku->ku_mailbox->km_timeofday, sizeof(ts));
1791 }
1792
1793out:
1794 if (error) {
1795 /*
1796 * Things are going to be so screwed we should just kill
1797 * the process.
1798 * how do we do that?
1799 */
1800 PROC_LOCK(td->td_proc);
1801 psignal(td->td_proc, SIGSEGV);
1802 PROC_UNLOCK(td->td_proc);
1803 } else {
1804 /*
1805 * Optimisation:
1806 * Ensure that we have a spare thread available,
1807 * for when we re-enter the kernel.
1808 */
1809 if (td->td_standin == NULL)
1810 thread_alloc_spare(td, NULL);
1811 }
1812
1813 ku->ku_mflags = 0;
1814 /*
1815 * Clear thread mailbox first, then clear system tick count.
1816 * The order is important because thread_statclock() use
1817 * mailbox pointer to see if it is an userland thread or
1818 * an UTS kernel thread.
1819 */
1820 td->td_mailbox = NULL;
1821 td->td_usticks = 0;
1822 return (error); /* go sync */
1823}
1824
1825/*
1826 * Enforce single-threading.
1827 *
1828 * Returns 1 if the caller must abort (another thread is waiting to
1829 * exit the process or similar). Process is locked!
1830 * Returns 0 when you are successfully the only thread running.
1831 * A process has successfully single threaded in the suspend mode when
1832 * There are no threads in user mode. Threads in the kernel must be
1833 * allowed to continue until they get to the user boundary. They may even
1834 * copy out their return values and data before suspending. They may however be
1835 * accellerated in reaching the user boundary as we will wake up
1836 * any sleeping threads that are interruptable. (PCATCH).
1837 */
1838int
1839thread_single(int force_exit)
1840{
1841 struct thread *td;
1842 struct thread *td2;
1843 struct proc *p;
1844
1845 td = curthread;
1846 p = td->td_proc;
1847 mtx_assert(&Giant, MA_OWNED);
1848 PROC_LOCK_ASSERT(p, MA_OWNED);
1849 KASSERT((td != NULL), ("curthread is NULL"));
1850
1851 if ((p->p_flag & P_SA) == 0 && p->p_numthreads == 1)
1852 return (0);
1853
1854 /* Is someone already single threading? */
1855 if (p->p_singlethread)
1856 return (1);
1857
1858 if (force_exit == SINGLE_EXIT) {
1859 p->p_flag |= P_SINGLE_EXIT;
1860 } else
1861 p->p_flag &= ~P_SINGLE_EXIT;
1862 p->p_flag |= P_STOPPED_SINGLE;
1863 mtx_lock_spin(&sched_lock);
1864 p->p_singlethread = td;
1865 while ((p->p_numthreads - p->p_suspcount) != 1) {
1866 FOREACH_THREAD_IN_PROC(p, td2) {
1867 if (td2 == td)
1868 continue;
1869 td2->td_flags |= TDF_ASTPENDING;
1870 if (TD_IS_INHIBITED(td2)) {
1871 if (force_exit == SINGLE_EXIT) {
1872 if (TD_IS_SUSPENDED(td2)) {
1873 thread_unsuspend_one(td2);
1874 }
1875 if (TD_ON_SLEEPQ(td2) &&
1876 (td2->td_flags & TDF_SINTR)) {
1877 if (td2->td_flags & TDF_CVWAITQ)
1878 cv_abort(td2);
1879 else
1880 abortsleep(td2);
1881 }
1882 } else {
1883 if (TD_IS_SUSPENDED(td2))
1884 continue;
1885 /*
1886 * maybe other inhibitted states too?
1887 * XXXKSE Is it totally safe to
1888 * suspend a non-interruptable thread?
1889 */
1890 if (td2->td_inhibitors &
1891 (TDI_SLEEPING | TDI_SWAPPED))
1892 thread_suspend_one(td2);
1893 }
1894 }
1895 }
1896 /*
1897 * Maybe we suspended some threads.. was it enough?
1898 */
1899 if ((p->p_numthreads - p->p_suspcount) == 1)
1900 break;
1901
1902 /*
1903 * Wake us up when everyone else has suspended.
1904 * In the mean time we suspend as well.
1905 */
1906 thread_suspend_one(td);
1907 DROP_GIANT();
1908 PROC_UNLOCK(p);
1909 p->p_stats->p_ru.ru_nvcsw++;
1910 mi_switch();
1911 mtx_unlock_spin(&sched_lock);
1912 PICKUP_GIANT();
1913 PROC_LOCK(p);
1914 mtx_lock_spin(&sched_lock);
1915 }
1916 if (force_exit == SINGLE_EXIT) {
1917 if (td->td_upcall)
1918 upcall_remove(td);
1919 kse_purge(p, td);
1920 }
1921 mtx_unlock_spin(&sched_lock);
1922 return (0);
1923}
1924
1925/*
1926 * Called in from locations that can safely check to see
1927 * whether we have to suspend or at least throttle for a
1928 * single-thread event (e.g. fork).
1929 *
1930 * Such locations include userret().
1931 * If the "return_instead" argument is non zero, the thread must be able to
1932 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1933 *
1934 * The 'return_instead' argument tells the function if it may do a
1935 * thread_exit() or suspend, or whether the caller must abort and back
1936 * out instead.
1937 *
1938 * If the thread that set the single_threading request has set the
1939 * P_SINGLE_EXIT bit in the process flags then this call will never return
1940 * if 'return_instead' is false, but will exit.
1941 *
1942 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1943 *---------------+--------------------+---------------------
1944 * 0 | returns 0 | returns 0 or 1
1945 * | when ST ends | immediatly
1946 *---------------+--------------------+---------------------
1947 * 1 | thread exits | returns 1
1948 * | | immediatly
1949 * 0 = thread_exit() or suspension ok,
1950 * other = return error instead of stopping the thread.
1951 *
1952 * While a full suspension is under effect, even a single threading
1953 * thread would be suspended if it made this call (but it shouldn't).
1954 * This call should only be made from places where
1955 * thread_exit() would be safe as that may be the outcome unless
1956 * return_instead is set.
1957 */
1958int
1959thread_suspend_check(int return_instead)
1960{
1961 struct thread *td;
1962 struct proc *p;
1963
1964 td = curthread;
1965 p = td->td_proc;
1966 PROC_LOCK_ASSERT(p, MA_OWNED);
1967 while (P_SHOULDSTOP(p)) {
1968 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1969 KASSERT(p->p_singlethread != NULL,
1970 ("singlethread not set"));
1971 /*
1972 * The only suspension in action is a
1973 * single-threading. Single threader need not stop.
1974 * XXX Should be safe to access unlocked
1975 * as it can only be set to be true by us.
1976 */
1977 if (p->p_singlethread == td)
1978 return (0); /* Exempt from stopping. */
1979 }
1980 if (return_instead)
1981 return (1);
1982
1983 mtx_lock_spin(&sched_lock);
1984 thread_stopped(p);
1985 /*
1986 * If the process is waiting for us to exit,
1987 * this thread should just suicide.
1988 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1989 */
1990 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1991 while (mtx_owned(&Giant))
1992 mtx_unlock(&Giant);
1993 if (p->p_flag & P_SA)
1994 thread_exit();
1995 else
1996 thr_exit1();
1997 }
1998
1999 /*
2000 * When a thread suspends, it just
2001 * moves to the processes's suspend queue
2002 * and stays there.
2003 */
2004 thread_suspend_one(td);
2005 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
2006 if (p->p_numthreads == p->p_suspcount) {
2007 thread_unsuspend_one(p->p_singlethread);
2008 }
2009 }
2010 DROP_GIANT();
2011 PROC_UNLOCK(p);
2012 p->p_stats->p_ru.ru_nivcsw++;
2013 mi_switch();
2014 mtx_unlock_spin(&sched_lock);
2015 PICKUP_GIANT();
2016 PROC_LOCK(p);
2017 }
2018 return (0);
2019}
2020
2021void
2022thread_suspend_one(struct thread *td)
2023{
2024 struct proc *p = td->td_proc;
2025
2026 mtx_assert(&sched_lock, MA_OWNED);
2027 PROC_LOCK_ASSERT(p, MA_OWNED);
2028 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
2029 p->p_suspcount++;
2030 TD_SET_SUSPENDED(td);
2031 TAILQ_INSERT_TAIL(&p->p_suspended, td, td_runq);
2032 /*
2033 * Hack: If we are suspending but are on the sleep queue
2034 * then we are in msleep or the cv equivalent. We
2035 * want to look like we have two Inhibitors.
2036 * May already be set.. doesn't matter.
2037 */
2038 if (TD_ON_SLEEPQ(td))
2039 TD_SET_SLEEPING(td);
2040}
2041
2042void
2043thread_unsuspend_one(struct thread *td)
2044{
2045 struct proc *p = td->td_proc;
2046
2047 mtx_assert(&sched_lock, MA_OWNED);
2048 PROC_LOCK_ASSERT(p, MA_OWNED);
2049 TAILQ_REMOVE(&p->p_suspended, td, td_runq);
2050 TD_CLR_SUSPENDED(td);
2051 p->p_suspcount--;
2052 setrunnable(td);
2053}
2054
2055/*
2056 * Allow all threads blocked by single threading to continue running.
2057 */
2058void
2059thread_unsuspend(struct proc *p)
2060{
2061 struct thread *td;
2062
2063 mtx_assert(&sched_lock, MA_OWNED);
2064 PROC_LOCK_ASSERT(p, MA_OWNED);
2065 if (!P_SHOULDSTOP(p)) {
2066 while (( td = TAILQ_FIRST(&p->p_suspended))) {
2067 thread_unsuspend_one(td);
2068 }
2069 } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
2070 (p->p_numthreads == p->p_suspcount)) {
2071 /*
2072 * Stopping everything also did the job for the single
2073 * threading request. Now we've downgraded to single-threaded,
2074 * let it continue.
2075 */
2076 thread_unsuspend_one(p->p_singlethread);
2077 }
2078}
2079
2080void
2081thread_single_end(void)
2082{
2083 struct thread *td;
2084 struct proc *p;
2085
2086 td = curthread;
2087 p = td->td_proc;
2088 PROC_LOCK_ASSERT(p, MA_OWNED);
2089 p->p_flag &= ~P_STOPPED_SINGLE;
2090 mtx_lock_spin(&sched_lock);
2091 p->p_singlethread = NULL;
2092 /*
2093 * If there are other threads they mey now run,
2094 * unless of course there is a blanket 'stop order'
2095 * on the process. The single threader must be allowed
2096 * to continue however as this is a bad place to stop.
2097 */
2098 if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
2099 while (( td = TAILQ_FIRST(&p->p_suspended))) {
2100 thread_unsuspend_one(td);
2101 }
2102 }
2103 mtx_unlock_spin(&sched_lock);
2104}
2105
2106
2 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice(s), this list of conditions and the following disclaimer as
10 * the first lines of this file unmodified other than the possible
11 * addition of one or more copyright notices.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice(s), this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
17 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
20 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
21 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
26 * DAMAGE.
27 */
28
29#include <sys/cdefs.h>
30__FBSDID("$FreeBSD: head/sys/kern/kern_thread.c 118486 2003-08-05 12:00:55Z davidxu $");
31
32#include <sys/param.h>
33#include <sys/systm.h>
34#include <sys/kernel.h>
35#include <sys/lock.h>
36#include <sys/malloc.h>
37#include <sys/mutex.h>
38#include <sys/proc.h>
39#include <sys/smp.h>
40#include <sys/sysctl.h>
41#include <sys/sysproto.h>
42#include <sys/filedesc.h>
43#include <sys/sched.h>
44#include <sys/signalvar.h>
45#include <sys/sx.h>
46#include <sys/tty.h>
47#include <sys/user.h>
48#include <sys/jail.h>
49#include <sys/kse.h>
50#include <sys/ktr.h>
51#include <sys/ucontext.h>
52
53#include <vm/vm.h>
54#include <vm/vm_extern.h>
55#include <vm/vm_object.h>
56#include <vm/pmap.h>
57#include <vm/uma.h>
58#include <vm/vm_map.h>
59
60#include <machine/frame.h>
61
62/*
63 * KSEGRP related storage.
64 */
65static uma_zone_t ksegrp_zone;
66static uma_zone_t kse_zone;
67static uma_zone_t thread_zone;
68static uma_zone_t upcall_zone;
69
70/* DEBUG ONLY */
71SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
72static int thread_debug = 0;
73SYSCTL_INT(_kern_threads, OID_AUTO, debug, CTLFLAG_RW,
74 &thread_debug, 0, "thread debug");
75
76static int max_threads_per_proc = 150;
77SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
78 &max_threads_per_proc, 0, "Limit on threads per proc");
79
80static int max_groups_per_proc = 50;
81SYSCTL_INT(_kern_threads, OID_AUTO, max_groups_per_proc, CTLFLAG_RW,
82 &max_groups_per_proc, 0, "Limit on thread groups per proc");
83
84static int max_threads_hits;
85SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
86 &max_threads_hits, 0, "");
87
88static int virtual_cpu;
89
90#define RANGEOF(type, start, end) (offsetof(type, end) - offsetof(type, start))
91
92TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
93TAILQ_HEAD(, kse) zombie_kses = TAILQ_HEAD_INITIALIZER(zombie_kses);
94TAILQ_HEAD(, ksegrp) zombie_ksegrps = TAILQ_HEAD_INITIALIZER(zombie_ksegrps);
95TAILQ_HEAD(, kse_upcall) zombie_upcalls =
96 TAILQ_HEAD_INITIALIZER(zombie_upcalls);
97struct mtx kse_zombie_lock;
98MTX_SYSINIT(kse_zombie_lock, &kse_zombie_lock, "kse zombie lock", MTX_SPIN);
99
100static void kse_purge(struct proc *p, struct thread *td);
101static void kse_purge_group(struct thread *td);
102static int thread_update_usr_ticks(struct thread *td, int user);
103static void thread_alloc_spare(struct thread *td, struct thread *spare);
104
105static int
106sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
107{
108 int error, new_val;
109 int def_val;
110
111#ifdef SMP
112 def_val = mp_ncpus;
113#else
114 def_val = 1;
115#endif
116 if (virtual_cpu == 0)
117 new_val = def_val;
118 else
119 new_val = virtual_cpu;
120 error = sysctl_handle_int(oidp, &new_val, 0, req);
121 if (error != 0 || req->newptr == NULL)
122 return (error);
123 if (new_val < 0)
124 return (EINVAL);
125 virtual_cpu = new_val;
126 return (0);
127}
128
129/* DEBUG ONLY */
130SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
131 0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
132 "debug virtual cpus");
133
134/*
135 * Prepare a thread for use.
136 */
137static void
138thread_ctor(void *mem, int size, void *arg)
139{
140 struct thread *td;
141
142 td = (struct thread *)mem;
143 td->td_state = TDS_INACTIVE;
144 td->td_oncpu = NOCPU;
145 td->td_critnest = 1;
146}
147
148/*
149 * Reclaim a thread after use.
150 */
151static void
152thread_dtor(void *mem, int size, void *arg)
153{
154 struct thread *td;
155
156 td = (struct thread *)mem;
157
158#ifdef INVARIANTS
159 /* Verify that this thread is in a safe state to free. */
160 switch (td->td_state) {
161 case TDS_INHIBITED:
162 case TDS_RUNNING:
163 case TDS_CAN_RUN:
164 case TDS_RUNQ:
165 /*
166 * We must never unlink a thread that is in one of
167 * these states, because it is currently active.
168 */
169 panic("bad state for thread unlinking");
170 /* NOTREACHED */
171 case TDS_INACTIVE:
172 break;
173 default:
174 panic("bad thread state");
175 /* NOTREACHED */
176 }
177#endif
178}
179
180/*
181 * Initialize type-stable parts of a thread (when newly created).
182 */
183static void
184thread_init(void *mem, int size)
185{
186 struct thread *td;
187
188 td = (struct thread *)mem;
189 mtx_lock(&Giant);
190 vm_thread_new(td, 0);
191 mtx_unlock(&Giant);
192 cpu_thread_setup(td);
193 td->td_sched = (struct td_sched *)&td[1];
194}
195
196/*
197 * Tear down type-stable parts of a thread (just before being discarded).
198 */
199static void
200thread_fini(void *mem, int size)
201{
202 struct thread *td;
203
204 td = (struct thread *)mem;
205 vm_thread_dispose(td);
206}
207
208/*
209 * Initialize type-stable parts of a kse (when newly created).
210 */
211static void
212kse_init(void *mem, int size)
213{
214 struct kse *ke;
215
216 ke = (struct kse *)mem;
217 ke->ke_sched = (struct ke_sched *)&ke[1];
218}
219
220/*
221 * Initialize type-stable parts of a ksegrp (when newly created).
222 */
223static void
224ksegrp_init(void *mem, int size)
225{
226 struct ksegrp *kg;
227
228 kg = (struct ksegrp *)mem;
229 kg->kg_sched = (struct kg_sched *)&kg[1];
230}
231
232/*
233 * KSE is linked into kse group.
234 */
235void
236kse_link(struct kse *ke, struct ksegrp *kg)
237{
238 struct proc *p = kg->kg_proc;
239
240 TAILQ_INSERT_HEAD(&kg->kg_kseq, ke, ke_kglist);
241 kg->kg_kses++;
242 ke->ke_state = KES_UNQUEUED;
243 ke->ke_proc = p;
244 ke->ke_ksegrp = kg;
245 ke->ke_thread = NULL;
246 ke->ke_oncpu = NOCPU;
247 ke->ke_flags = 0;
248}
249
250void
251kse_unlink(struct kse *ke)
252{
253 struct ksegrp *kg;
254
255 mtx_assert(&sched_lock, MA_OWNED);
256 kg = ke->ke_ksegrp;
257 TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist);
258 if (ke->ke_state == KES_IDLE) {
259 TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist);
260 kg->kg_idle_kses--;
261 }
262 if (--kg->kg_kses == 0)
263 ksegrp_unlink(kg);
264 /*
265 * Aggregate stats from the KSE
266 */
267 kse_stash(ke);
268}
269
270void
271ksegrp_link(struct ksegrp *kg, struct proc *p)
272{
273
274 TAILQ_INIT(&kg->kg_threads);
275 TAILQ_INIT(&kg->kg_runq); /* links with td_runq */
276 TAILQ_INIT(&kg->kg_slpq); /* links with td_runq */
277 TAILQ_INIT(&kg->kg_kseq); /* all kses in ksegrp */
278 TAILQ_INIT(&kg->kg_iq); /* all idle kses in ksegrp */
279 TAILQ_INIT(&kg->kg_upcalls); /* all upcall structure in ksegrp */
280 kg->kg_proc = p;
281 /*
282 * the following counters are in the -zero- section
283 * and may not need clearing
284 */
285 kg->kg_numthreads = 0;
286 kg->kg_runnable = 0;
287 kg->kg_kses = 0;
288 kg->kg_runq_kses = 0; /* XXXKSE change name */
289 kg->kg_idle_kses = 0;
290 kg->kg_numupcalls = 0;
291 /* link it in now that it's consistent */
292 p->p_numksegrps++;
293 TAILQ_INSERT_HEAD(&p->p_ksegrps, kg, kg_ksegrp);
294}
295
296void
297ksegrp_unlink(struct ksegrp *kg)
298{
299 struct proc *p;
300
301 mtx_assert(&sched_lock, MA_OWNED);
302 KASSERT((kg->kg_numthreads == 0), ("ksegrp_unlink: residual threads"));
303 KASSERT((kg->kg_kses == 0), ("ksegrp_unlink: residual kses"));
304 KASSERT((kg->kg_numupcalls == 0), ("ksegrp_unlink: residual upcalls"));
305
306 p = kg->kg_proc;
307 TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
308 p->p_numksegrps--;
309 /*
310 * Aggregate stats from the KSE
311 */
312 ksegrp_stash(kg);
313}
314
315struct kse_upcall *
316upcall_alloc(void)
317{
318 struct kse_upcall *ku;
319
320 ku = uma_zalloc(upcall_zone, M_WAITOK);
321 bzero(ku, sizeof(*ku));
322 return (ku);
323}
324
325void
326upcall_free(struct kse_upcall *ku)
327{
328
329 uma_zfree(upcall_zone, ku);
330}
331
332void
333upcall_link(struct kse_upcall *ku, struct ksegrp *kg)
334{
335
336 mtx_assert(&sched_lock, MA_OWNED);
337 TAILQ_INSERT_TAIL(&kg->kg_upcalls, ku, ku_link);
338 ku->ku_ksegrp = kg;
339 kg->kg_numupcalls++;
340}
341
342void
343upcall_unlink(struct kse_upcall *ku)
344{
345 struct ksegrp *kg = ku->ku_ksegrp;
346
347 mtx_assert(&sched_lock, MA_OWNED);
348 KASSERT(ku->ku_owner == NULL, ("%s: have owner", __func__));
349 TAILQ_REMOVE(&kg->kg_upcalls, ku, ku_link);
350 kg->kg_numupcalls--;
351 upcall_stash(ku);
352}
353
354void
355upcall_remove(struct thread *td)
356{
357
358 if (td->td_upcall) {
359 td->td_upcall->ku_owner = NULL;
360 upcall_unlink(td->td_upcall);
361 td->td_upcall = 0;
362 }
363}
364
365/*
366 * For a newly created process,
367 * link up all the structures and its initial threads etc.
368 */
369void
370proc_linkup(struct proc *p, struct ksegrp *kg,
371 struct kse *ke, struct thread *td)
372{
373
374 TAILQ_INIT(&p->p_ksegrps); /* all ksegrps in proc */
375 TAILQ_INIT(&p->p_threads); /* all threads in proc */
376 TAILQ_INIT(&p->p_suspended); /* Threads suspended */
377 p->p_numksegrps = 0;
378 p->p_numthreads = 0;
379
380 ksegrp_link(kg, p);
381 kse_link(ke, kg);
382 thread_link(td, kg);
383}
384
385/*
386struct kse_thr_interrupt_args {
387 struct kse_thr_mailbox * tmbx;
388 int cmd;
389 long data;
390};
391*/
392int
393kse_thr_interrupt(struct thread *td, struct kse_thr_interrupt_args *uap)
394{
395 struct proc *p;
396 struct thread *td2;
397
398 p = td->td_proc;
399 if (!(p->p_flag & P_SA))
400 return (EINVAL);
401
402 switch (uap->cmd) {
403 case KSE_INTR_SENDSIG:
404 if (uap->data < 0 || uap->data > _SIG_MAXSIG)
405 return (EINVAL);
406 case KSE_INTR_INTERRUPT:
407 case KSE_INTR_RESTART:
408 PROC_LOCK(p);
409 mtx_lock_spin(&sched_lock);
410 FOREACH_THREAD_IN_PROC(p, td2) {
411 if (td2->td_mailbox == uap->tmbx)
412 break;
413 }
414 if (td2 == NULL) {
415 mtx_unlock_spin(&sched_lock);
416 PROC_UNLOCK(p);
417 return (ESRCH);
418 }
419 if (uap->cmd == KSE_INTR_SENDSIG) {
420 if (uap->data > 0) {
421 td2->td_flags &= ~TDF_INTERRUPT;
422 mtx_unlock_spin(&sched_lock);
423 tdsignal(td2, (int)uap->data, SIGTARGET_TD);
424 } else {
425 mtx_unlock_spin(&sched_lock);
426 }
427 } else {
428 td2->td_flags |= TDF_INTERRUPT | TDF_ASTPENDING;
429 if (TD_CAN_UNBIND(td2))
430 td2->td_upcall->ku_flags |= KUF_DOUPCALL;
431 if (uap->cmd == KSE_INTR_INTERRUPT)
432 td2->td_intrval = EINTR;
433 else
434 td2->td_intrval = ERESTART;
435 if (TD_ON_SLEEPQ(td2) && (td2->td_flags & TDF_SINTR)) {
436 if (td2->td_flags & TDF_CVWAITQ)
437 cv_abort(td2);
438 else
439 abortsleep(td2);
440 }
441 mtx_unlock_spin(&sched_lock);
442 }
443 PROC_UNLOCK(p);
444 break;
445 case KSE_INTR_SIGEXIT:
446 if (uap->data < 1 || uap->data > _SIG_MAXSIG)
447 return (EINVAL);
448 PROC_LOCK(p);
449 sigexit(td, (int)uap->data);
450 break;
451 default:
452 return (EINVAL);
453 }
454 return (0);
455}
456
457/*
458struct kse_exit_args {
459 register_t dummy;
460};
461*/
462int
463kse_exit(struct thread *td, struct kse_exit_args *uap)
464{
465 struct proc *p;
466 struct ksegrp *kg;
467 struct kse *ke;
468 struct kse_upcall *ku, *ku2;
469 int error, count;
470
471 p = td->td_proc;
472 if ((ku = td->td_upcall) == NULL || TD_CAN_UNBIND(td))
473 return (EINVAL);
474 kg = td->td_ksegrp;
475 count = 0;
476 PROC_LOCK(p);
477 mtx_lock_spin(&sched_lock);
478 FOREACH_UPCALL_IN_GROUP(kg, ku2) {
479 if (ku2->ku_flags & KUF_EXITING)
480 count++;
481 }
482 if ((kg->kg_numupcalls - count) == 1 &&
483 (kg->kg_numthreads > 1)) {
484 mtx_unlock_spin(&sched_lock);
485 PROC_UNLOCK(p);
486 return (EDEADLK);
487 }
488 ku->ku_flags |= KUF_EXITING;
489 mtx_unlock_spin(&sched_lock);
490 PROC_UNLOCK(p);
491 error = suword(&ku->ku_mailbox->km_flags, ku->ku_mflags|KMF_DONE);
492 PROC_LOCK(p);
493 if (error)
494 psignal(p, SIGSEGV);
495 mtx_lock_spin(&sched_lock);
496 upcall_remove(td);
497 ke = td->td_kse;
498 if (p->p_numthreads == 1) {
499 kse_purge(p, td);
500 p->p_flag &= ~P_SA;
501 mtx_unlock_spin(&sched_lock);
502 PROC_UNLOCK(p);
503 } else {
504 if (kg->kg_numthreads == 1) { /* Shutdown a group */
505 kse_purge_group(td);
506 ke->ke_flags |= KEF_EXIT;
507 }
508 thread_stopped(p);
509 thread_exit();
510 /* NOTREACHED */
511 }
512 return (0);
513}
514
515/*
516 * Either becomes an upcall or waits for an awakening event and
517 * then becomes an upcall. Only error cases return.
518 */
519/*
520struct kse_release_args {
521 struct timespec *timeout;
522};
523*/
524int
525kse_release(struct thread *td, struct kse_release_args *uap)
526{
527 struct proc *p;
528 struct ksegrp *kg;
529 struct kse_upcall *ku;
530 struct timespec timeout;
531 struct timeval tv;
532 sigset_t sigset;
533 int error;
534
535 p = td->td_proc;
536 kg = td->td_ksegrp;
537 if ((ku = td->td_upcall) == NULL || TD_CAN_UNBIND(td))
538 return (EINVAL);
539 if (uap->timeout != NULL) {
540 if ((error = copyin(uap->timeout, &timeout, sizeof(timeout))))
541 return (error);
542 TIMESPEC_TO_TIMEVAL(&tv, &timeout);
543 }
544 if (td->td_flags & TDF_SA)
545 td->td_pflags |= TDP_UPCALLING;
546 else {
547 ku->ku_mflags = fuword(&ku->ku_mailbox->km_flags);
548 if (ku->ku_mflags == -1) {
549 PROC_LOCK(p);
550 sigexit(td, SIGSEGV);
551 }
552 }
553 PROC_LOCK(p);
554 if (ku->ku_mflags & KMF_WAITSIGEVENT) {
555 /* UTS wants to wait for signal event */
556 if (!(p->p_flag & P_SIGEVENT) && !(ku->ku_flags & KUF_DOUPCALL))
557 error = msleep(&p->p_siglist, &p->p_mtx, PPAUSE|PCATCH,
558 "ksesigwait", (uap->timeout ? tvtohz(&tv) : 0));
559 p->p_flag &= ~P_SIGEVENT;
560 sigset = p->p_siglist;
561 PROC_UNLOCK(p);
562 error = copyout(&sigset, &ku->ku_mailbox->km_sigscaught,
563 sizeof(sigset));
564 } else {
565 if (! kg->kg_completed && !(ku->ku_flags & KUF_DOUPCALL)) {
566 kg->kg_upsleeps++;
567 error = msleep(&kg->kg_completed, &p->p_mtx,
568 PPAUSE|PCATCH, "kserel",
569 (uap->timeout ? tvtohz(&tv) : 0));
570 kg->kg_upsleeps--;
571 }
572 PROC_UNLOCK(p);
573 }
574 if (ku->ku_flags & KUF_DOUPCALL) {
575 mtx_lock_spin(&sched_lock);
576 ku->ku_flags &= ~KUF_DOUPCALL;
577 mtx_unlock_spin(&sched_lock);
578 }
579 return (0);
580}
581
582/* struct kse_wakeup_args {
583 struct kse_mailbox *mbx;
584}; */
585int
586kse_wakeup(struct thread *td, struct kse_wakeup_args *uap)
587{
588 struct proc *p;
589 struct ksegrp *kg;
590 struct kse_upcall *ku;
591 struct thread *td2;
592
593 p = td->td_proc;
594 td2 = NULL;
595 ku = NULL;
596 /* KSE-enabled processes only, please. */
597 if (!(p->p_flag & P_SA))
598 return (EINVAL);
599 PROC_LOCK(p);
600 mtx_lock_spin(&sched_lock);
601 if (uap->mbx) {
602 FOREACH_KSEGRP_IN_PROC(p, kg) {
603 FOREACH_UPCALL_IN_GROUP(kg, ku) {
604 if (ku->ku_mailbox == uap->mbx)
605 break;
606 }
607 if (ku)
608 break;
609 }
610 } else {
611 kg = td->td_ksegrp;
612 if (kg->kg_upsleeps) {
613 wakeup_one(&kg->kg_completed);
614 mtx_unlock_spin(&sched_lock);
615 PROC_UNLOCK(p);
616 return (0);
617 }
618 ku = TAILQ_FIRST(&kg->kg_upcalls);
619 }
620 if (ku) {
621 if ((td2 = ku->ku_owner) == NULL) {
622 panic("%s: no owner", __func__);
623 } else if (TD_ON_SLEEPQ(td2) &&
624 ((td2->td_wchan == &kg->kg_completed) ||
625 (td2->td_wchan == &p->p_siglist &&
626 (ku->ku_mflags & KMF_WAITSIGEVENT)))) {
627 abortsleep(td2);
628 } else {
629 ku->ku_flags |= KUF_DOUPCALL;
630 }
631 mtx_unlock_spin(&sched_lock);
632 PROC_UNLOCK(p);
633 return (0);
634 }
635 mtx_unlock_spin(&sched_lock);
636 PROC_UNLOCK(p);
637 return (ESRCH);
638}
639
640/*
641 * No new KSEG: first call: use current KSE, don't schedule an upcall
642 * All other situations, do allocate max new KSEs and schedule an upcall.
643 */
644/* struct kse_create_args {
645 struct kse_mailbox *mbx;
646 int newgroup;
647}; */
648int
649kse_create(struct thread *td, struct kse_create_args *uap)
650{
651 struct kse *newke;
652 struct ksegrp *newkg;
653 struct ksegrp *kg;
654 struct proc *p;
655 struct kse_mailbox mbx;
656 struct kse_upcall *newku;
657 int err, ncpus, sa = 0, first = 0;
658 struct thread *newtd;
659
660 p = td->td_proc;
661 if ((err = copyin(uap->mbx, &mbx, sizeof(mbx))))
662 return (err);
663
664 /* Too bad, why hasn't kernel always a cpu counter !? */
665#ifdef SMP
666 ncpus = mp_ncpus;
667#else
668 ncpus = 1;
669#endif
670 if (virtual_cpu != 0)
671 ncpus = virtual_cpu;
672 if (!(mbx.km_flags & KMF_BOUND))
673 sa = TDF_SA;
674 else
675 ncpus = 1;
676 PROC_LOCK(p);
677 if (!(p->p_flag & P_SA)) {
678 first = 1;
679 p->p_flag |= P_SA;
680 }
681 PROC_UNLOCK(p);
682 if (!sa && !uap->newgroup && !first)
683 return (EINVAL);
684 kg = td->td_ksegrp;
685 if (uap->newgroup) {
686 /* Have race condition but it is cheap */
687 if (p->p_numksegrps >= max_groups_per_proc)
688 return (EPROCLIM);
689 /*
690 * If we want a new KSEGRP it doesn't matter whether
691 * we have already fired up KSE mode before or not.
692 * We put the process in KSE mode and create a new KSEGRP.
693 */
694 newkg = ksegrp_alloc();
695 bzero(&newkg->kg_startzero, RANGEOF(struct ksegrp,
696 kg_startzero, kg_endzero));
697 bcopy(&kg->kg_startcopy, &newkg->kg_startcopy,
698 RANGEOF(struct ksegrp, kg_startcopy, kg_endcopy));
699 mtx_lock_spin(&sched_lock);
700 if (p->p_numksegrps >= max_groups_per_proc) {
701 mtx_unlock_spin(&sched_lock);
702 ksegrp_free(newkg);
703 return (EPROCLIM);
704 }
705 ksegrp_link(newkg, p);
706 mtx_unlock_spin(&sched_lock);
707 } else {
708 if (!first && ((td->td_flags & TDF_SA) ^ sa) != 0)
709 return (EINVAL);
710 newkg = kg;
711 }
712
713 /*
714 * Creating upcalls more than number of physical cpu does
715 * not help performance.
716 */
717 if (newkg->kg_numupcalls >= ncpus)
718 return (EPROCLIM);
719
720 if (newkg->kg_numupcalls == 0) {
721 /*
722 * Initialize KSE group
723 *
724 * For multiplxed group, create KSEs as many as physical
725 * cpus. This increases concurrent even if userland
726 * is not MP safe and can only run on single CPU.
727 * In ideal world, every physical cpu should execute a thread.
728 * If there is enough KSEs, threads in kernel can be
729 * executed parallel on different cpus with full speed,
730 * Concurrent in kernel shouldn't be restricted by number of
731 * upcalls userland provides. Adding more upcall structures
732 * only increases concurrent in userland.
733 *
734 * For bound thread group, because there is only thread in the
735 * group, we only create one KSE for the group. Thread in this
736 * kind of group will never schedule an upcall when blocked,
737 * this intends to simulate pthread system scope thread.
738 */
739 while (newkg->kg_kses < ncpus) {
740 newke = kse_alloc();
741 bzero(&newke->ke_startzero, RANGEOF(struct kse,
742 ke_startzero, ke_endzero));
743#if 0
744 mtx_lock_spin(&sched_lock);
745 bcopy(&ke->ke_startcopy, &newke->ke_startcopy,
746 RANGEOF(struct kse, ke_startcopy, ke_endcopy));
747 mtx_unlock_spin(&sched_lock);
748#endif
749 mtx_lock_spin(&sched_lock);
750 kse_link(newke, newkg);
751 /* Add engine */
752 kse_reassign(newke);
753 mtx_unlock_spin(&sched_lock);
754 }
755 }
756 newku = upcall_alloc();
757 newku->ku_mailbox = uap->mbx;
758 newku->ku_func = mbx.km_func;
759 bcopy(&mbx.km_stack, &newku->ku_stack, sizeof(stack_t));
760
761 /* For the first call this may not have been set */
762 if (td->td_standin == NULL)
763 thread_alloc_spare(td, NULL);
764
765 PROC_LOCK(p);
766 if (newkg->kg_numupcalls >= ncpus) {
767 PROC_UNLOCK(p);
768 upcall_free(newku);
769 return (EPROCLIM);
770 }
771 if (first && sa) {
772 SIGSETOR(p->p_siglist, td->td_siglist);
773 SIGEMPTYSET(td->td_siglist);
774 SIGFILLSET(td->td_sigmask);
775 SIG_CANTMASK(td->td_sigmask);
776 }
777 mtx_lock_spin(&sched_lock);
778 PROC_UNLOCK(p);
779 upcall_link(newku, newkg);
780 if (mbx.km_quantum)
781 newkg->kg_upquantum = max(1, mbx.km_quantum/tick);
782
783 /*
784 * Each upcall structure has an owner thread, find which
785 * one owns it.
786 */
787 if (uap->newgroup) {
788 /*
789 * Because new ksegrp hasn't thread,
790 * create an initial upcall thread to own it.
791 */
792 newtd = thread_schedule_upcall(td, newku);
793 } else {
794 /*
795 * If current thread hasn't an upcall structure,
796 * just assign the upcall to it.
797 */
798 if (td->td_upcall == NULL) {
799 newku->ku_owner = td;
800 td->td_upcall = newku;
801 newtd = td;
802 } else {
803 /*
804 * Create a new upcall thread to own it.
805 */
806 newtd = thread_schedule_upcall(td, newku);
807 }
808 }
809 if (!sa) {
810 newtd->td_mailbox = mbx.km_curthread;
811 newtd->td_flags &= ~TDF_SA;
812 if (newtd != td) {
813 mtx_unlock_spin(&sched_lock);
814 cpu_set_upcall_kse(newtd, newku);
815 mtx_lock_spin(&sched_lock);
816 }
817 } else {
818 newtd->td_flags |= TDF_SA;
819 }
820 if (newtd != td)
821 setrunqueue(newtd);
822 mtx_unlock_spin(&sched_lock);
823 return (0);
824}
825
826/*
827 * Initialize global thread allocation resources.
828 */
829void
830threadinit(void)
831{
832
833 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
834 thread_ctor, thread_dtor, thread_init, thread_fini,
835 UMA_ALIGN_CACHE, 0);
836 ksegrp_zone = uma_zcreate("KSEGRP", sched_sizeof_ksegrp(),
837 NULL, NULL, ksegrp_init, NULL,
838 UMA_ALIGN_CACHE, 0);
839 kse_zone = uma_zcreate("KSE", sched_sizeof_kse(),
840 NULL, NULL, kse_init, NULL,
841 UMA_ALIGN_CACHE, 0);
842 upcall_zone = uma_zcreate("UPCALL", sizeof(struct kse_upcall),
843 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
844}
845
846/*
847 * Stash an embarasingly extra thread into the zombie thread queue.
848 */
849void
850thread_stash(struct thread *td)
851{
852 mtx_lock_spin(&kse_zombie_lock);
853 TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq);
854 mtx_unlock_spin(&kse_zombie_lock);
855}
856
857/*
858 * Stash an embarasingly extra kse into the zombie kse queue.
859 */
860void
861kse_stash(struct kse *ke)
862{
863 mtx_lock_spin(&kse_zombie_lock);
864 TAILQ_INSERT_HEAD(&zombie_kses, ke, ke_procq);
865 mtx_unlock_spin(&kse_zombie_lock);
866}
867
868/*
869 * Stash an embarasingly extra upcall into the zombie upcall queue.
870 */
871
872void
873upcall_stash(struct kse_upcall *ku)
874{
875 mtx_lock_spin(&kse_zombie_lock);
876 TAILQ_INSERT_HEAD(&zombie_upcalls, ku, ku_link);
877 mtx_unlock_spin(&kse_zombie_lock);
878}
879
880/*
881 * Stash an embarasingly extra ksegrp into the zombie ksegrp queue.
882 */
883void
884ksegrp_stash(struct ksegrp *kg)
885{
886 mtx_lock_spin(&kse_zombie_lock);
887 TAILQ_INSERT_HEAD(&zombie_ksegrps, kg, kg_ksegrp);
888 mtx_unlock_spin(&kse_zombie_lock);
889}
890
891/*
892 * Reap zombie kse resource.
893 */
894void
895thread_reap(void)
896{
897 struct thread *td_first, *td_next;
898 struct kse *ke_first, *ke_next;
899 struct ksegrp *kg_first, * kg_next;
900 struct kse_upcall *ku_first, *ku_next;
901
902 /*
903 * Don't even bother to lock if none at this instant,
904 * we really don't care about the next instant..
905 */
906 if ((!TAILQ_EMPTY(&zombie_threads))
907 || (!TAILQ_EMPTY(&zombie_kses))
908 || (!TAILQ_EMPTY(&zombie_ksegrps))
909 || (!TAILQ_EMPTY(&zombie_upcalls))) {
910 mtx_lock_spin(&kse_zombie_lock);
911 td_first = TAILQ_FIRST(&zombie_threads);
912 ke_first = TAILQ_FIRST(&zombie_kses);
913 kg_first = TAILQ_FIRST(&zombie_ksegrps);
914 ku_first = TAILQ_FIRST(&zombie_upcalls);
915 if (td_first)
916 TAILQ_INIT(&zombie_threads);
917 if (ke_first)
918 TAILQ_INIT(&zombie_kses);
919 if (kg_first)
920 TAILQ_INIT(&zombie_ksegrps);
921 if (ku_first)
922 TAILQ_INIT(&zombie_upcalls);
923 mtx_unlock_spin(&kse_zombie_lock);
924 while (td_first) {
925 td_next = TAILQ_NEXT(td_first, td_runq);
926 if (td_first->td_ucred)
927 crfree(td_first->td_ucred);
928 thread_free(td_first);
929 td_first = td_next;
930 }
931 while (ke_first) {
932 ke_next = TAILQ_NEXT(ke_first, ke_procq);
933 kse_free(ke_first);
934 ke_first = ke_next;
935 }
936 while (kg_first) {
937 kg_next = TAILQ_NEXT(kg_first, kg_ksegrp);
938 ksegrp_free(kg_first);
939 kg_first = kg_next;
940 }
941 while (ku_first) {
942 ku_next = TAILQ_NEXT(ku_first, ku_link);
943 upcall_free(ku_first);
944 ku_first = ku_next;
945 }
946 }
947}
948
949/*
950 * Allocate a ksegrp.
951 */
952struct ksegrp *
953ksegrp_alloc(void)
954{
955 return (uma_zalloc(ksegrp_zone, M_WAITOK));
956}
957
958/*
959 * Allocate a kse.
960 */
961struct kse *
962kse_alloc(void)
963{
964 return (uma_zalloc(kse_zone, M_WAITOK));
965}
966
967/*
968 * Allocate a thread.
969 */
970struct thread *
971thread_alloc(void)
972{
973 thread_reap(); /* check if any zombies to get */
974 return (uma_zalloc(thread_zone, M_WAITOK));
975}
976
977/*
978 * Deallocate a ksegrp.
979 */
980void
981ksegrp_free(struct ksegrp *td)
982{
983 uma_zfree(ksegrp_zone, td);
984}
985
986/*
987 * Deallocate a kse.
988 */
989void
990kse_free(struct kse *td)
991{
992 uma_zfree(kse_zone, td);
993}
994
995/*
996 * Deallocate a thread.
997 */
998void
999thread_free(struct thread *td)
1000{
1001
1002 cpu_thread_clean(td);
1003 uma_zfree(thread_zone, td);
1004}
1005
1006/*
1007 * Store the thread context in the UTS's mailbox.
1008 * then add the mailbox at the head of a list we are building in user space.
1009 * The list is anchored in the ksegrp structure.
1010 */
1011int
1012thread_export_context(struct thread *td, int willexit)
1013{
1014 struct proc *p;
1015 struct ksegrp *kg;
1016 uintptr_t mbx;
1017 void *addr;
1018 int error = 0, temp, sig;
1019 mcontext_t mc;
1020
1021 p = td->td_proc;
1022 kg = td->td_ksegrp;
1023
1024 /* Export the user/machine context. */
1025 get_mcontext(td, &mc, 0);
1026 addr = (void *)(&td->td_mailbox->tm_context.uc_mcontext);
1027 error = copyout(&mc, addr, sizeof(mcontext_t));
1028 if (error)
1029 goto bad;
1030
1031 /* Exports clock ticks in kernel mode */
1032 addr = (caddr_t)(&td->td_mailbox->tm_sticks);
1033 temp = fuword32(addr) + td->td_usticks;
1034 if (suword32(addr, temp)) {
1035 error = EFAULT;
1036 goto bad;
1037 }
1038
1039 /*
1040 * Post sync signal, or process SIGKILL and SIGSTOP.
1041 * For sync signal, it is only possible when the signal is not
1042 * caught by userland or process is being debugged.
1043 */
1044 PROC_LOCK(p);
1045 if (td->td_flags & TDF_NEEDSIGCHK) {
1046 mtx_lock_spin(&sched_lock);
1047 td->td_flags &= ~TDF_NEEDSIGCHK;
1048 mtx_unlock_spin(&sched_lock);
1049 mtx_lock(&p->p_sigacts->ps_mtx);
1050 while ((sig = cursig(td)) != 0)
1051 postsig(sig);
1052 mtx_unlock(&p->p_sigacts->ps_mtx);
1053 }
1054 if (willexit)
1055 SIGFILLSET(td->td_sigmask);
1056 PROC_UNLOCK(p);
1057
1058 /* Get address in latest mbox of list pointer */
1059 addr = (void *)(&td->td_mailbox->tm_next);
1060 /*
1061 * Put the saved address of the previous first
1062 * entry into this one
1063 */
1064 for (;;) {
1065 mbx = (uintptr_t)kg->kg_completed;
1066 if (suword(addr, mbx)) {
1067 error = EFAULT;
1068 goto bad;
1069 }
1070 PROC_LOCK(p);
1071 if (mbx == (uintptr_t)kg->kg_completed) {
1072 kg->kg_completed = td->td_mailbox;
1073 /*
1074 * The thread context may be taken away by
1075 * other upcall threads when we unlock
1076 * process lock. it's no longer valid to
1077 * use it again in any other places.
1078 */
1079 td->td_mailbox = NULL;
1080 PROC_UNLOCK(p);
1081 break;
1082 }
1083 PROC_UNLOCK(p);
1084 }
1085 td->td_usticks = 0;
1086 return (0);
1087
1088bad:
1089 PROC_LOCK(p);
1090 sigexit(td, SIGILL);
1091 return (error);
1092}
1093
1094/*
1095 * Take the list of completed mailboxes for this KSEGRP and put them on this
1096 * upcall's mailbox as it's the next one going up.
1097 */
1098static int
1099thread_link_mboxes(struct ksegrp *kg, struct kse_upcall *ku)
1100{
1101 struct proc *p = kg->kg_proc;
1102 void *addr;
1103 uintptr_t mbx;
1104
1105 addr = (void *)(&ku->ku_mailbox->km_completed);
1106 for (;;) {
1107 mbx = (uintptr_t)kg->kg_completed;
1108 if (suword(addr, mbx)) {
1109 PROC_LOCK(p);
1110 psignal(p, SIGSEGV);
1111 PROC_UNLOCK(p);
1112 return (EFAULT);
1113 }
1114 PROC_LOCK(p);
1115 if (mbx == (uintptr_t)kg->kg_completed) {
1116 kg->kg_completed = NULL;
1117 PROC_UNLOCK(p);
1118 break;
1119 }
1120 PROC_UNLOCK(p);
1121 }
1122 return (0);
1123}
1124
1125/*
1126 * This function should be called at statclock interrupt time
1127 */
1128int
1129thread_statclock(int user)
1130{
1131 struct thread *td = curthread;
1132 struct ksegrp *kg = td->td_ksegrp;
1133
1134 if (kg->kg_numupcalls == 0 || !(td->td_flags & TDF_SA))
1135 return (0);
1136 if (user) {
1137 /* Current always do via ast() */
1138 mtx_lock_spin(&sched_lock);
1139 td->td_flags |= (TDF_USTATCLOCK|TDF_ASTPENDING);
1140 mtx_unlock_spin(&sched_lock);
1141 td->td_uuticks++;
1142 } else {
1143 if (td->td_mailbox != NULL)
1144 td->td_usticks++;
1145 else {
1146 /* XXXKSE
1147 * We will call thread_user_enter() for every
1148 * kernel entry in future, so if the thread mailbox
1149 * is NULL, it must be a UTS kernel, don't account
1150 * clock ticks for it.
1151 */
1152 }
1153 }
1154 return (0);
1155}
1156
1157/*
1158 * Export state clock ticks for userland
1159 */
1160static int
1161thread_update_usr_ticks(struct thread *td, int user)
1162{
1163 struct proc *p = td->td_proc;
1164 struct kse_thr_mailbox *tmbx;
1165 struct kse_upcall *ku;
1166 struct ksegrp *kg;
1167 caddr_t addr;
1168 uint uticks;
1169
1170 if ((ku = td->td_upcall) == NULL)
1171 return (-1);
1172
1173 tmbx = (void *)fuword((void *)&ku->ku_mailbox->km_curthread);
1174 if ((tmbx == NULL) || (tmbx == (void *)-1))
1175 return (-1);
1176 if (user) {
1177 uticks = td->td_uuticks;
1178 td->td_uuticks = 0;
1179 addr = (caddr_t)&tmbx->tm_uticks;
1180 } else {
1181 uticks = td->td_usticks;
1182 td->td_usticks = 0;
1183 addr = (caddr_t)&tmbx->tm_sticks;
1184 }
1185 if (uticks) {
1186 if (suword32(addr, uticks+fuword32(addr))) {
1187 PROC_LOCK(p);
1188 psignal(p, SIGSEGV);
1189 PROC_UNLOCK(p);
1190 return (-2);
1191 }
1192 }
1193 kg = td->td_ksegrp;
1194 if (kg->kg_upquantum && ticks >= kg->kg_nextupcall) {
1195 mtx_lock_spin(&sched_lock);
1196 td->td_upcall->ku_flags |= KUF_DOUPCALL;
1197 mtx_unlock_spin(&sched_lock);
1198 }
1199 return (0);
1200}
1201
1202/*
1203 * Discard the current thread and exit from its context.
1204 *
1205 * Because we can't free a thread while we're operating under its context,
1206 * push the current thread into our CPU's deadthread holder. This means
1207 * we needn't worry about someone else grabbing our context before we
1208 * do a cpu_throw().
1209 */
1210void
1211thread_exit(void)
1212{
1213 struct thread *td;
1214 struct kse *ke;
1215 struct proc *p;
1216 struct ksegrp *kg;
1217
1218 td = curthread;
1219 kg = td->td_ksegrp;
1220 p = td->td_proc;
1221 ke = td->td_kse;
1222
1223 mtx_assert(&sched_lock, MA_OWNED);
1224 KASSERT(p != NULL, ("thread exiting without a process"));
1225 KASSERT(ke != NULL, ("thread exiting without a kse"));
1226 KASSERT(kg != NULL, ("thread exiting without a kse group"));
1227 PROC_LOCK_ASSERT(p, MA_OWNED);
1228 CTR1(KTR_PROC, "thread_exit: thread %p", td);
1229 KASSERT(!mtx_owned(&Giant), ("dying thread owns giant"));
1230
1231 if (td->td_standin != NULL) {
1232 thread_stash(td->td_standin);
1233 td->td_standin = NULL;
1234 }
1235
1236 cpu_thread_exit(td); /* XXXSMP */
1237
1238 /*
1239 * The last thread is left attached to the process
1240 * So that the whole bundle gets recycled. Skip
1241 * all this stuff.
1242 */
1243 if (p->p_numthreads > 1) {
1244 thread_unlink(td);
1245 if (p->p_maxthrwaits)
1246 wakeup(&p->p_numthreads);
1247 /*
1248 * The test below is NOT true if we are the
1249 * sole exiting thread. P_STOPPED_SNGL is unset
1250 * in exit1() after it is the only survivor.
1251 */
1252 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1253 if (p->p_numthreads == p->p_suspcount) {
1254 thread_unsuspend_one(p->p_singlethread);
1255 }
1256 }
1257
1258 /*
1259 * Because each upcall structure has an owner thread,
1260 * owner thread exits only when process is in exiting
1261 * state, so upcall to userland is no longer needed,
1262 * deleting upcall structure is safe here.
1263 * So when all threads in a group is exited, all upcalls
1264 * in the group should be automatically freed.
1265 */
1266 if (td->td_upcall)
1267 upcall_remove(td);
1268
1269 ke->ke_state = KES_UNQUEUED;
1270 ke->ke_thread = NULL;
1271 /*
1272 * Decide what to do with the KSE attached to this thread.
1273 */
1274 if (ke->ke_flags & KEF_EXIT)
1275 kse_unlink(ke);
1276 else
1277 kse_reassign(ke);
1278 PROC_UNLOCK(p);
1279 td->td_kse = NULL;
1280 td->td_state = TDS_INACTIVE;
1281#if 0
1282 td->td_proc = NULL;
1283#endif
1284 td->td_ksegrp = NULL;
1285 td->td_last_kse = NULL;
1286 PCPU_SET(deadthread, td);
1287 } else {
1288 PROC_UNLOCK(p);
1289 }
1290 /* XXX Shouldn't cpu_throw() here. */
1291 mtx_assert(&sched_lock, MA_OWNED);
1292#if !defined(__alpha__) && !defined(__powerpc__)
1293 cpu_throw(td, choosethread());
1294#else
1295 cpu_throw();
1296#endif
1297 panic("I'm a teapot!");
1298 /* NOTREACHED */
1299}
1300
1301/*
1302 * Do any thread specific cleanups that may be needed in wait()
1303 * called with Giant held, proc and schedlock not held.
1304 */
1305void
1306thread_wait(struct proc *p)
1307{
1308 struct thread *td;
1309
1310 KASSERT((p->p_numthreads == 1), ("Muliple threads in wait1()"));
1311 KASSERT((p->p_numksegrps == 1), ("Muliple ksegrps in wait1()"));
1312 FOREACH_THREAD_IN_PROC(p, td) {
1313 if (td->td_standin != NULL) {
1314 thread_free(td->td_standin);
1315 td->td_standin = NULL;
1316 }
1317 cpu_thread_clean(td);
1318 }
1319 thread_reap(); /* check for zombie threads etc. */
1320}
1321
1322/*
1323 * Link a thread to a process.
1324 * set up anything that needs to be initialized for it to
1325 * be used by the process.
1326 *
1327 * Note that we do not link to the proc's ucred here.
1328 * The thread is linked as if running but no KSE assigned.
1329 */
1330void
1331thread_link(struct thread *td, struct ksegrp *kg)
1332{
1333 struct proc *p;
1334
1335 p = kg->kg_proc;
1336 td->td_state = TDS_INACTIVE;
1337 td->td_proc = p;
1338 td->td_ksegrp = kg;
1339 td->td_last_kse = NULL;
1340 td->td_flags = 0;
1341 td->td_kse = NULL;
1342
1343 LIST_INIT(&td->td_contested);
1344 callout_init(&td->td_slpcallout, 1);
1345 TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
1346 TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist);
1347 p->p_numthreads++;
1348 kg->kg_numthreads++;
1349}
1350
1351void
1352thread_unlink(struct thread *td)
1353{
1354 struct proc *p = td->td_proc;
1355 struct ksegrp *kg = td->td_ksegrp;
1356
1357 mtx_assert(&sched_lock, MA_OWNED);
1358 TAILQ_REMOVE(&p->p_threads, td, td_plist);
1359 p->p_numthreads--;
1360 TAILQ_REMOVE(&kg->kg_threads, td, td_kglist);
1361 kg->kg_numthreads--;
1362 /* could clear a few other things here */
1363}
1364
1365/*
1366 * Purge a ksegrp resource. When a ksegrp is preparing to
1367 * exit, it calls this function.
1368 */
1369static void
1370kse_purge_group(struct thread *td)
1371{
1372 struct ksegrp *kg;
1373 struct kse *ke;
1374
1375 kg = td->td_ksegrp;
1376 KASSERT(kg->kg_numthreads == 1, ("%s: bad thread number", __func__));
1377 while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) {
1378 KASSERT(ke->ke_state == KES_IDLE,
1379 ("%s: wrong idle KSE state", __func__));
1380 kse_unlink(ke);
1381 }
1382 KASSERT((kg->kg_kses == 1),
1383 ("%s: ksegrp still has %d KSEs", __func__, kg->kg_kses));
1384 KASSERT((kg->kg_numupcalls == 0),
1385 ("%s: ksegrp still has %d upcall datas",
1386 __func__, kg->kg_numupcalls));
1387}
1388
1389/*
1390 * Purge a process's KSE resource. When a process is preparing to
1391 * exit, it calls kse_purge to release any extra KSE resources in
1392 * the process.
1393 */
1394static void
1395kse_purge(struct proc *p, struct thread *td)
1396{
1397 struct ksegrp *kg;
1398 struct kse *ke;
1399
1400 KASSERT(p->p_numthreads == 1, ("bad thread number"));
1401 while ((kg = TAILQ_FIRST(&p->p_ksegrps)) != NULL) {
1402 TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
1403 p->p_numksegrps--;
1404 /*
1405 * There is no ownership for KSE, after all threads
1406 * in the group exited, it is possible that some KSEs
1407 * were left in idle queue, gc them now.
1408 */
1409 while ((ke = TAILQ_FIRST(&kg->kg_iq)) != NULL) {
1410 KASSERT(ke->ke_state == KES_IDLE,
1411 ("%s: wrong idle KSE state", __func__));
1412 TAILQ_REMOVE(&kg->kg_iq, ke, ke_kgrlist);
1413 kg->kg_idle_kses--;
1414 TAILQ_REMOVE(&kg->kg_kseq, ke, ke_kglist);
1415 kg->kg_kses--;
1416 kse_stash(ke);
1417 }
1418 KASSERT(((kg->kg_kses == 0) && (kg != td->td_ksegrp)) ||
1419 ((kg->kg_kses == 1) && (kg == td->td_ksegrp)),
1420 ("ksegrp has wrong kg_kses: %d", kg->kg_kses));
1421 KASSERT((kg->kg_numupcalls == 0),
1422 ("%s: ksegrp still has %d upcall datas",
1423 __func__, kg->kg_numupcalls));
1424
1425 if (kg != td->td_ksegrp)
1426 ksegrp_stash(kg);
1427 }
1428 TAILQ_INSERT_HEAD(&p->p_ksegrps, td->td_ksegrp, kg_ksegrp);
1429 p->p_numksegrps++;
1430}
1431
1432/*
1433 * This function is intended to be used to initialize a spare thread
1434 * for upcall. Initialize thread's large data area outside sched_lock
1435 * for thread_schedule_upcall().
1436 */
1437void
1438thread_alloc_spare(struct thread *td, struct thread *spare)
1439{
1440 if (td->td_standin)
1441 return;
1442 if (spare == NULL)
1443 spare = thread_alloc();
1444 td->td_standin = spare;
1445 bzero(&spare->td_startzero,
1446 (unsigned)RANGEOF(struct thread, td_startzero, td_endzero));
1447 spare->td_proc = td->td_proc;
1448 spare->td_ucred = crhold(td->td_ucred);
1449}
1450
1451/*
1452 * Create a thread and schedule it for upcall on the KSE given.
1453 * Use our thread's standin so that we don't have to allocate one.
1454 */
1455struct thread *
1456thread_schedule_upcall(struct thread *td, struct kse_upcall *ku)
1457{
1458 struct thread *td2;
1459
1460 mtx_assert(&sched_lock, MA_OWNED);
1461
1462 /*
1463 * Schedule an upcall thread on specified kse_upcall,
1464 * the kse_upcall must be free.
1465 * td must have a spare thread.
1466 */
1467 KASSERT(ku->ku_owner == NULL, ("%s: upcall has owner", __func__));
1468 if ((td2 = td->td_standin) != NULL) {
1469 td->td_standin = NULL;
1470 } else {
1471 panic("no reserve thread when scheduling an upcall");
1472 return (NULL);
1473 }
1474 CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)",
1475 td2, td->td_proc->p_pid, td->td_proc->p_comm);
1476 bcopy(&td->td_startcopy, &td2->td_startcopy,
1477 (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy));
1478 thread_link(td2, ku->ku_ksegrp);
1479 /* inherit blocked thread's context */
1480 cpu_set_upcall(td2, td);
1481 /* Let the new thread become owner of the upcall */
1482 ku->ku_owner = td2;
1483 td2->td_upcall = ku;
1484 td2->td_flags = TDF_SA;
1485 td2->td_pflags = TDP_UPCALLING;
1486 td2->td_kse = NULL;
1487 td2->td_state = TDS_CAN_RUN;
1488 td2->td_inhibitors = 0;
1489 SIGFILLSET(td2->td_sigmask);
1490 SIG_CANTMASK(td2->td_sigmask);
1491 return (td2); /* bogus.. should be a void function */
1492}
1493
1494/*
1495 * It is only used when thread generated a trap and process is being
1496 * debugged.
1497 */
1498void
1499thread_signal_add(struct thread *td, int sig)
1500{
1501 struct proc *p;
1502 siginfo_t siginfo;
1503 struct sigacts *ps;
1504 int error;
1505
1506 p = td->td_proc;
1507 PROC_LOCK_ASSERT(p, MA_OWNED);
1508 ps = p->p_sigacts;
1509 mtx_assert(&ps->ps_mtx, MA_OWNED);
1510
1511 cpu_thread_siginfo(sig, 0, &siginfo);
1512 mtx_unlock(&ps->ps_mtx);
1513 PROC_UNLOCK(p);
1514 error = copyout(&siginfo, &td->td_mailbox->tm_syncsig, sizeof(siginfo));
1515 if (error) {
1516 PROC_LOCK(p);
1517 sigexit(td, SIGILL);
1518 }
1519 PROC_LOCK(p);
1520 SIGADDSET(td->td_sigmask, sig);
1521 mtx_lock(&ps->ps_mtx);
1522}
1523
1524void
1525thread_switchout(struct thread *td)
1526{
1527 struct kse_upcall *ku;
1528 struct thread *td2;
1529
1530 mtx_assert(&sched_lock, MA_OWNED);
1531
1532 /*
1533 * If the outgoing thread is in threaded group and has never
1534 * scheduled an upcall, decide whether this is a short
1535 * or long term event and thus whether or not to schedule
1536 * an upcall.
1537 * If it is a short term event, just suspend it in
1538 * a way that takes its KSE with it.
1539 * Select the events for which we want to schedule upcalls.
1540 * For now it's just sleep.
1541 * XXXKSE eventually almost any inhibition could do.
1542 */
1543 if (TD_CAN_UNBIND(td) && (td->td_standin) && TD_ON_SLEEPQ(td)) {
1544 /*
1545 * Release ownership of upcall, and schedule an upcall
1546 * thread, this new upcall thread becomes the owner of
1547 * the upcall structure.
1548 */
1549 ku = td->td_upcall;
1550 ku->ku_owner = NULL;
1551 td->td_upcall = NULL;
1552 td->td_flags &= ~TDF_CAN_UNBIND;
1553 td2 = thread_schedule_upcall(td, ku);
1554 setrunqueue(td2);
1555 }
1556}
1557
1558/*
1559 * Setup done on the thread when it enters the kernel.
1560 * XXXKSE Presently only for syscalls but eventually all kernel entries.
1561 */
1562void
1563thread_user_enter(struct proc *p, struct thread *td)
1564{
1565 struct ksegrp *kg;
1566 struct kse_upcall *ku;
1567 struct kse_thr_mailbox *tmbx;
1568 uint32_t tflags;
1569
1570 kg = td->td_ksegrp;
1571
1572 /*
1573 * First check that we shouldn't just abort.
1574 * But check if we are the single thread first!
1575 */
1576 if (p->p_flag & P_SINGLE_EXIT) {
1577 PROC_LOCK(p);
1578 mtx_lock_spin(&sched_lock);
1579 thread_stopped(p);
1580 thread_exit();
1581 /* NOTREACHED */
1582 }
1583
1584 /*
1585 * If we are doing a syscall in a KSE environment,
1586 * note where our mailbox is. There is always the
1587 * possibility that we could do this lazily (in kse_reassign()),
1588 * but for now do it every time.
1589 */
1590 kg = td->td_ksegrp;
1591 if (td->td_flags & TDF_SA) {
1592 ku = td->td_upcall;
1593 KASSERT(ku, ("%s: no upcall owned", __func__));
1594 KASSERT((ku->ku_owner == td), ("%s: wrong owner", __func__));
1595 KASSERT(!TD_CAN_UNBIND(td), ("%s: can unbind", __func__));
1596 ku->ku_mflags = fuword32((void *)&ku->ku_mailbox->km_flags);
1597 tmbx = (void *)fuword((void *)&ku->ku_mailbox->km_curthread);
1598 if ((tmbx == NULL) || (tmbx == (void *)-1L) ||
1599 (ku->ku_mflags & KMF_NOUPCALL)) {
1600 td->td_mailbox = NULL;
1601 } else {
1602 if (td->td_standin == NULL)
1603 thread_alloc_spare(td, NULL);
1604 tflags = fuword32(tmbx);
1605 /*
1606 * On some architectures, TP register points to thread
1607 * mailbox but not points to kse mailbox, and userland
1608 * can not atomically clear km_curthread, but can
1609 * use TP register, and set TMF_NOUPCALL in thread
1610 * flag to indicate a critical region.
1611 */
1612 if (tflags & TMF_NOUPCALL) {
1613 td->td_mailbox = NULL;
1614 } else {
1615 td->td_mailbox = tmbx;
1616 mtx_lock_spin(&sched_lock);
1617 td->td_flags |= TDF_CAN_UNBIND;
1618 mtx_unlock_spin(&sched_lock);
1619 }
1620 }
1621 }
1622}
1623
1624/*
1625 * The extra work we go through if we are a threaded process when we
1626 * return to userland.
1627 *
1628 * If we are a KSE process and returning to user mode, check for
1629 * extra work to do before we return (e.g. for more syscalls
1630 * to complete first). If we were in a critical section, we should
1631 * just return to let it finish. Same if we were in the UTS (in
1632 * which case the mailbox's context's busy indicator will be set).
1633 * The only traps we suport will have set the mailbox.
1634 * We will clear it here.
1635 */
1636int
1637thread_userret(struct thread *td, struct trapframe *frame)
1638{
1639 int error = 0, upcalls, uts_crit;
1640 struct kse_upcall *ku;
1641 struct ksegrp *kg, *kg2;
1642 struct proc *p;
1643 struct timespec ts;
1644
1645 p = td->td_proc;
1646 kg = td->td_ksegrp;
1647 ku = td->td_upcall;
1648
1649 /* Nothing to do with bound thread */
1650 if (!(td->td_flags & TDF_SA))
1651 return (0);
1652
1653 /*
1654 * Stat clock interrupt hit in userland, it
1655 * is returning from interrupt, charge thread's
1656 * userland time for UTS.
1657 */
1658 if (td->td_flags & TDF_USTATCLOCK) {
1659 thread_update_usr_ticks(td, 1);
1660 mtx_lock_spin(&sched_lock);
1661 td->td_flags &= ~TDF_USTATCLOCK;
1662 mtx_unlock_spin(&sched_lock);
1663 if (kg->kg_completed ||
1664 (td->td_upcall->ku_flags & KUF_DOUPCALL))
1665 thread_user_enter(p, td);
1666 }
1667
1668 uts_crit = (td->td_mailbox == NULL);
1669 /*
1670 * Optimisation:
1671 * This thread has not started any upcall.
1672 * If there is no work to report other than ourself,
1673 * then it can return direct to userland.
1674 */
1675 if (TD_CAN_UNBIND(td)) {
1676 mtx_lock_spin(&sched_lock);
1677 td->td_flags &= ~TDF_CAN_UNBIND;
1678 if ((td->td_flags & TDF_NEEDSIGCHK) == 0 &&
1679 (kg->kg_completed == NULL) &&
1680 (ku->ku_flags & KUF_DOUPCALL) == 0 &&
1681 (kg->kg_upquantum && ticks < kg->kg_nextupcall)) {
1682 mtx_unlock_spin(&sched_lock);
1683 thread_update_usr_ticks(td, 0);
1684 nanotime(&ts);
1685 error = copyout(&ts,
1686 (caddr_t)&ku->ku_mailbox->km_timeofday,
1687 sizeof(ts));
1688 td->td_mailbox = 0;
1689 ku->ku_mflags = 0;
1690 if (error)
1691 goto out;
1692 return (0);
1693 }
1694 mtx_unlock_spin(&sched_lock);
1695 thread_export_context(td, 0);
1696 /*
1697 * There is something to report, and we own an upcall
1698 * strucuture, we can go to userland.
1699 * Turn ourself into an upcall thread.
1700 */
1701 td->td_pflags |= TDP_UPCALLING;
1702 } else if (td->td_mailbox && (ku == NULL)) {
1703 thread_export_context(td, 1);
1704 PROC_LOCK(p);
1705 /*
1706 * There are upcall threads waiting for
1707 * work to do, wake one of them up.
1708 * XXXKSE Maybe wake all of them up.
1709 */
1710 if (kg->kg_upsleeps)
1711 wakeup_one(&kg->kg_completed);
1712 mtx_lock_spin(&sched_lock);
1713 thread_stopped(p);
1714 thread_exit();
1715 /* NOTREACHED */
1716 }
1717
1718 KASSERT(ku != NULL, ("upcall is NULL\n"));
1719 KASSERT(TD_CAN_UNBIND(td) == 0, ("can unbind"));
1720
1721 if (p->p_numthreads > max_threads_per_proc) {
1722 max_threads_hits++;
1723 PROC_LOCK(p);
1724 mtx_lock_spin(&sched_lock);
1725 p->p_maxthrwaits++;
1726 while (p->p_numthreads > max_threads_per_proc) {
1727 upcalls = 0;
1728 FOREACH_KSEGRP_IN_PROC(p, kg2) {
1729 if (kg2->kg_numupcalls == 0)
1730 upcalls++;
1731 else
1732 upcalls += kg2->kg_numupcalls;
1733 }
1734 if (upcalls >= max_threads_per_proc)
1735 break;
1736 mtx_unlock_spin(&sched_lock);
1737 if (msleep(&p->p_numthreads, &p->p_mtx, PPAUSE|PCATCH,
1738 "maxthreads", NULL)) {
1739 mtx_lock_spin(&sched_lock);
1740 break;
1741 } else {
1742 mtx_lock_spin(&sched_lock);
1743 }
1744 }
1745 p->p_maxthrwaits--;
1746 mtx_unlock_spin(&sched_lock);
1747 PROC_UNLOCK(p);
1748 }
1749
1750 if (td->td_pflags & TDP_UPCALLING) {
1751 uts_crit = 0;
1752 kg->kg_nextupcall = ticks+kg->kg_upquantum;
1753 /*
1754 * There is no more work to do and we are going to ride
1755 * this thread up to userland as an upcall.
1756 * Do the last parts of the setup needed for the upcall.
1757 */
1758 CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)",
1759 td, td->td_proc->p_pid, td->td_proc->p_comm);
1760
1761 td->td_pflags &= ~TDP_UPCALLING;
1762 if (ku->ku_flags & KUF_DOUPCALL) {
1763 mtx_lock_spin(&sched_lock);
1764 ku->ku_flags &= ~KUF_DOUPCALL;
1765 mtx_unlock_spin(&sched_lock);
1766 }
1767 /*
1768 * Set user context to the UTS
1769 */
1770 if (!(ku->ku_mflags & KMF_NOUPCALL)) {
1771 cpu_set_upcall_kse(td, ku);
1772 error = suword(&ku->ku_mailbox->km_curthread, 0);
1773 if (error)
1774 goto out;
1775 }
1776
1777 /*
1778 * Unhook the list of completed threads.
1779 * anything that completes after this gets to
1780 * come in next time.
1781 * Put the list of completed thread mailboxes on
1782 * this KSE's mailbox.
1783 */
1784 if (!(ku->ku_mflags & KMF_NOCOMPLETED) &&
1785 (error = thread_link_mboxes(kg, ku)) != 0)
1786 goto out;
1787 }
1788 if (!uts_crit) {
1789 nanotime(&ts);
1790 error = copyout(&ts, &ku->ku_mailbox->km_timeofday, sizeof(ts));
1791 }
1792
1793out:
1794 if (error) {
1795 /*
1796 * Things are going to be so screwed we should just kill
1797 * the process.
1798 * how do we do that?
1799 */
1800 PROC_LOCK(td->td_proc);
1801 psignal(td->td_proc, SIGSEGV);
1802 PROC_UNLOCK(td->td_proc);
1803 } else {
1804 /*
1805 * Optimisation:
1806 * Ensure that we have a spare thread available,
1807 * for when we re-enter the kernel.
1808 */
1809 if (td->td_standin == NULL)
1810 thread_alloc_spare(td, NULL);
1811 }
1812
1813 ku->ku_mflags = 0;
1814 /*
1815 * Clear thread mailbox first, then clear system tick count.
1816 * The order is important because thread_statclock() use
1817 * mailbox pointer to see if it is an userland thread or
1818 * an UTS kernel thread.
1819 */
1820 td->td_mailbox = NULL;
1821 td->td_usticks = 0;
1822 return (error); /* go sync */
1823}
1824
1825/*
1826 * Enforce single-threading.
1827 *
1828 * Returns 1 if the caller must abort (another thread is waiting to
1829 * exit the process or similar). Process is locked!
1830 * Returns 0 when you are successfully the only thread running.
1831 * A process has successfully single threaded in the suspend mode when
1832 * There are no threads in user mode. Threads in the kernel must be
1833 * allowed to continue until they get to the user boundary. They may even
1834 * copy out their return values and data before suspending. They may however be
1835 * accellerated in reaching the user boundary as we will wake up
1836 * any sleeping threads that are interruptable. (PCATCH).
1837 */
1838int
1839thread_single(int force_exit)
1840{
1841 struct thread *td;
1842 struct thread *td2;
1843 struct proc *p;
1844
1845 td = curthread;
1846 p = td->td_proc;
1847 mtx_assert(&Giant, MA_OWNED);
1848 PROC_LOCK_ASSERT(p, MA_OWNED);
1849 KASSERT((td != NULL), ("curthread is NULL"));
1850
1851 if ((p->p_flag & P_SA) == 0 && p->p_numthreads == 1)
1852 return (0);
1853
1854 /* Is someone already single threading? */
1855 if (p->p_singlethread)
1856 return (1);
1857
1858 if (force_exit == SINGLE_EXIT) {
1859 p->p_flag |= P_SINGLE_EXIT;
1860 } else
1861 p->p_flag &= ~P_SINGLE_EXIT;
1862 p->p_flag |= P_STOPPED_SINGLE;
1863 mtx_lock_spin(&sched_lock);
1864 p->p_singlethread = td;
1865 while ((p->p_numthreads - p->p_suspcount) != 1) {
1866 FOREACH_THREAD_IN_PROC(p, td2) {
1867 if (td2 == td)
1868 continue;
1869 td2->td_flags |= TDF_ASTPENDING;
1870 if (TD_IS_INHIBITED(td2)) {
1871 if (force_exit == SINGLE_EXIT) {
1872 if (TD_IS_SUSPENDED(td2)) {
1873 thread_unsuspend_one(td2);
1874 }
1875 if (TD_ON_SLEEPQ(td2) &&
1876 (td2->td_flags & TDF_SINTR)) {
1877 if (td2->td_flags & TDF_CVWAITQ)
1878 cv_abort(td2);
1879 else
1880 abortsleep(td2);
1881 }
1882 } else {
1883 if (TD_IS_SUSPENDED(td2))
1884 continue;
1885 /*
1886 * maybe other inhibitted states too?
1887 * XXXKSE Is it totally safe to
1888 * suspend a non-interruptable thread?
1889 */
1890 if (td2->td_inhibitors &
1891 (TDI_SLEEPING | TDI_SWAPPED))
1892 thread_suspend_one(td2);
1893 }
1894 }
1895 }
1896 /*
1897 * Maybe we suspended some threads.. was it enough?
1898 */
1899 if ((p->p_numthreads - p->p_suspcount) == 1)
1900 break;
1901
1902 /*
1903 * Wake us up when everyone else has suspended.
1904 * In the mean time we suspend as well.
1905 */
1906 thread_suspend_one(td);
1907 DROP_GIANT();
1908 PROC_UNLOCK(p);
1909 p->p_stats->p_ru.ru_nvcsw++;
1910 mi_switch();
1911 mtx_unlock_spin(&sched_lock);
1912 PICKUP_GIANT();
1913 PROC_LOCK(p);
1914 mtx_lock_spin(&sched_lock);
1915 }
1916 if (force_exit == SINGLE_EXIT) {
1917 if (td->td_upcall)
1918 upcall_remove(td);
1919 kse_purge(p, td);
1920 }
1921 mtx_unlock_spin(&sched_lock);
1922 return (0);
1923}
1924
1925/*
1926 * Called in from locations that can safely check to see
1927 * whether we have to suspend or at least throttle for a
1928 * single-thread event (e.g. fork).
1929 *
1930 * Such locations include userret().
1931 * If the "return_instead" argument is non zero, the thread must be able to
1932 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1933 *
1934 * The 'return_instead' argument tells the function if it may do a
1935 * thread_exit() or suspend, or whether the caller must abort and back
1936 * out instead.
1937 *
1938 * If the thread that set the single_threading request has set the
1939 * P_SINGLE_EXIT bit in the process flags then this call will never return
1940 * if 'return_instead' is false, but will exit.
1941 *
1942 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1943 *---------------+--------------------+---------------------
1944 * 0 | returns 0 | returns 0 or 1
1945 * | when ST ends | immediatly
1946 *---------------+--------------------+---------------------
1947 * 1 | thread exits | returns 1
1948 * | | immediatly
1949 * 0 = thread_exit() or suspension ok,
1950 * other = return error instead of stopping the thread.
1951 *
1952 * While a full suspension is under effect, even a single threading
1953 * thread would be suspended if it made this call (but it shouldn't).
1954 * This call should only be made from places where
1955 * thread_exit() would be safe as that may be the outcome unless
1956 * return_instead is set.
1957 */
1958int
1959thread_suspend_check(int return_instead)
1960{
1961 struct thread *td;
1962 struct proc *p;
1963
1964 td = curthread;
1965 p = td->td_proc;
1966 PROC_LOCK_ASSERT(p, MA_OWNED);
1967 while (P_SHOULDSTOP(p)) {
1968 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1969 KASSERT(p->p_singlethread != NULL,
1970 ("singlethread not set"));
1971 /*
1972 * The only suspension in action is a
1973 * single-threading. Single threader need not stop.
1974 * XXX Should be safe to access unlocked
1975 * as it can only be set to be true by us.
1976 */
1977 if (p->p_singlethread == td)
1978 return (0); /* Exempt from stopping. */
1979 }
1980 if (return_instead)
1981 return (1);
1982
1983 mtx_lock_spin(&sched_lock);
1984 thread_stopped(p);
1985 /*
1986 * If the process is waiting for us to exit,
1987 * this thread should just suicide.
1988 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1989 */
1990 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1991 while (mtx_owned(&Giant))
1992 mtx_unlock(&Giant);
1993 if (p->p_flag & P_SA)
1994 thread_exit();
1995 else
1996 thr_exit1();
1997 }
1998
1999 /*
2000 * When a thread suspends, it just
2001 * moves to the processes's suspend queue
2002 * and stays there.
2003 */
2004 thread_suspend_one(td);
2005 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
2006 if (p->p_numthreads == p->p_suspcount) {
2007 thread_unsuspend_one(p->p_singlethread);
2008 }
2009 }
2010 DROP_GIANT();
2011 PROC_UNLOCK(p);
2012 p->p_stats->p_ru.ru_nivcsw++;
2013 mi_switch();
2014 mtx_unlock_spin(&sched_lock);
2015 PICKUP_GIANT();
2016 PROC_LOCK(p);
2017 }
2018 return (0);
2019}
2020
2021void
2022thread_suspend_one(struct thread *td)
2023{
2024 struct proc *p = td->td_proc;
2025
2026 mtx_assert(&sched_lock, MA_OWNED);
2027 PROC_LOCK_ASSERT(p, MA_OWNED);
2028 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
2029 p->p_suspcount++;
2030 TD_SET_SUSPENDED(td);
2031 TAILQ_INSERT_TAIL(&p->p_suspended, td, td_runq);
2032 /*
2033 * Hack: If we are suspending but are on the sleep queue
2034 * then we are in msleep or the cv equivalent. We
2035 * want to look like we have two Inhibitors.
2036 * May already be set.. doesn't matter.
2037 */
2038 if (TD_ON_SLEEPQ(td))
2039 TD_SET_SLEEPING(td);
2040}
2041
2042void
2043thread_unsuspend_one(struct thread *td)
2044{
2045 struct proc *p = td->td_proc;
2046
2047 mtx_assert(&sched_lock, MA_OWNED);
2048 PROC_LOCK_ASSERT(p, MA_OWNED);
2049 TAILQ_REMOVE(&p->p_suspended, td, td_runq);
2050 TD_CLR_SUSPENDED(td);
2051 p->p_suspcount--;
2052 setrunnable(td);
2053}
2054
2055/*
2056 * Allow all threads blocked by single threading to continue running.
2057 */
2058void
2059thread_unsuspend(struct proc *p)
2060{
2061 struct thread *td;
2062
2063 mtx_assert(&sched_lock, MA_OWNED);
2064 PROC_LOCK_ASSERT(p, MA_OWNED);
2065 if (!P_SHOULDSTOP(p)) {
2066 while (( td = TAILQ_FIRST(&p->p_suspended))) {
2067 thread_unsuspend_one(td);
2068 }
2069 } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
2070 (p->p_numthreads == p->p_suspcount)) {
2071 /*
2072 * Stopping everything also did the job for the single
2073 * threading request. Now we've downgraded to single-threaded,
2074 * let it continue.
2075 */
2076 thread_unsuspend_one(p->p_singlethread);
2077 }
2078}
2079
2080void
2081thread_single_end(void)
2082{
2083 struct thread *td;
2084 struct proc *p;
2085
2086 td = curthread;
2087 p = td->td_proc;
2088 PROC_LOCK_ASSERT(p, MA_OWNED);
2089 p->p_flag &= ~P_STOPPED_SINGLE;
2090 mtx_lock_spin(&sched_lock);
2091 p->p_singlethread = NULL;
2092 /*
2093 * If there are other threads they mey now run,
2094 * unless of course there is a blanket 'stop order'
2095 * on the process. The single threader must be allowed
2096 * to continue however as this is a bad place to stop.
2097 */
2098 if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
2099 while (( td = TAILQ_FIRST(&p->p_suspended))) {
2100 thread_unsuspend_one(td);
2101 }
2102 }
2103 mtx_unlock_spin(&sched_lock);
2104}
2105
2106