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: releng/6.1/sys/kern/kern_thread.c 158179 2006-04-30 16:44:43Z cvs2svn $");
31
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/kernel.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/proc.h>
38 #include <sys/resourcevar.h>
39 #include <sys/smp.h>
40 #include <sys/sysctl.h>
41 #include <sys/sched.h>
42 #include <sys/sleepqueue.h>
43 #include <sys/turnstile.h>
44 #include <sys/ktr.h>
45 #include <sys/umtx.h>
46
47 #include <vm/vm.h>
48 #include <vm/vm_extern.h>
49 #include <vm/uma.h>
50
51 /*
52 * KSEGRP related storage.
53 */
54 static uma_zone_t ksegrp_zone;
55 static uma_zone_t thread_zone;
56
57 /* DEBUG ONLY */
58 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
59 static int thread_debug = 0;
60 SYSCTL_INT(_kern_threads, OID_AUTO, debug, CTLFLAG_RW,
61 &thread_debug, 0, "thread debug");
62
63 int max_threads_per_proc = 1500;
64 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
65 &max_threads_per_proc, 0, "Limit on threads per proc");
66
67 int max_groups_per_proc = 1500;
68 SYSCTL_INT(_kern_threads, OID_AUTO, max_groups_per_proc, CTLFLAG_RW,
69 &max_groups_per_proc, 0, "Limit on thread groups per proc");
70
71 int max_threads_hits;
72 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
73 &max_threads_hits, 0, "");
74
75 int virtual_cpu;
76
77 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
78 TAILQ_HEAD(, ksegrp) zombie_ksegrps = TAILQ_HEAD_INITIALIZER(zombie_ksegrps);
79 struct mtx kse_zombie_lock;
80 MTX_SYSINIT(kse_zombie_lock, &kse_zombie_lock, "kse zombie lock", MTX_SPIN);
81
82 static int
83 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
84 {
85 int error, new_val;
86 int def_val;
87
88 def_val = mp_ncpus;
89 if (virtual_cpu == 0)
90 new_val = def_val;
91 else
92 new_val = virtual_cpu;
93 error = sysctl_handle_int(oidp, &new_val, 0, req);
94 if (error != 0 || req->newptr == NULL)
95 return (error);
96 if (new_val < 0)
97 return (EINVAL);
98 virtual_cpu = new_val;
99 return (0);
100 }
101
102 /* DEBUG ONLY */
103 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
104 0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
105 "debug virtual cpus");
106
107 struct mtx tid_lock;
108 static struct unrhdr *tid_unrhdr;
109
110 /*
111 * Prepare a thread for use.
112 */
113 static int
114 thread_ctor(void *mem, int size, void *arg, int flags)
115 {
116 struct thread *td;
117
118 td = (struct thread *)mem;
119 td->td_state = TDS_INACTIVE;
120 td->td_oncpu = NOCPU;
121
122 td->td_tid = alloc_unr(tid_unrhdr);
123
124 /*
125 * Note that td_critnest begins life as 1 because the thread is not
126 * running and is thereby implicitly waiting to be on the receiving
127 * end of a context switch. A context switch must occur inside a
128 * critical section, and in fact, includes hand-off of the sched_lock.
129 * After a context switch to a newly created thread, it will release
130 * sched_lock for the first time, and its td_critnest will hit 0 for
131 * the first time. This happens on the far end of a context switch,
132 * and when it context switches away from itself, it will in fact go
133 * back into a critical section, and hand off the sched lock to the
134 * next thread.
135 */
136 td->td_critnest = 1;
137 return (0);
138 }
139
140 /*
141 * Reclaim a thread after use.
142 */
143 static void
144 thread_dtor(void *mem, int size, void *arg)
145 {
146 struct thread *td;
147
148 td = (struct thread *)mem;
149
150 #ifdef INVARIANTS
151 /* Verify that this thread is in a safe state to free. */
152 switch (td->td_state) {
153 case TDS_INHIBITED:
154 case TDS_RUNNING:
155 case TDS_CAN_RUN:
156 case TDS_RUNQ:
157 /*
158 * We must never unlink a thread that is in one of
159 * these states, because it is currently active.
160 */
161 panic("bad state for thread unlinking");
162 /* NOTREACHED */
163 case TDS_INACTIVE:
164 break;
165 default:
166 panic("bad thread state");
167 /* NOTREACHED */
168 }
169 #endif
170
171 free_unr(tid_unrhdr, td->td_tid);
172 sched_newthread(td);
173 }
174
175 /*
176 * Initialize type-stable parts of a thread (when newly created).
177 */
178 static int
179 thread_init(void *mem, int size, int flags)
180 {
181 struct thread *td;
182
183 td = (struct thread *)mem;
184
185 vm_thread_new(td, 0);
186 cpu_thread_setup(td);
187 td->td_sleepqueue = sleepq_alloc();
188 td->td_turnstile = turnstile_alloc();
189 td->td_umtxq = umtxq_alloc();
190 td->td_sched = (struct td_sched *)&td[1];
191 sched_newthread(td);
192 return (0);
193 }
194
195 /*
196 * Tear down type-stable parts of a thread (just before being discarded).
197 */
198 static void
199 thread_fini(void *mem, int size)
200 {
201 struct thread *td;
202
203 td = (struct thread *)mem;
204 turnstile_free(td->td_turnstile);
205 sleepq_free(td->td_sleepqueue);
206 umtxq_free(td->td_umtxq);
207 vm_thread_dispose(td);
208 }
209
210 /*
211 * Initialize type-stable parts of a ksegrp (when newly created).
212 */
213 static int
214 ksegrp_ctor(void *mem, int size, void *arg, int flags)
215 {
216 struct ksegrp *kg;
217
218 kg = (struct ksegrp *)mem;
219 bzero(mem, size);
220 kg->kg_sched = (struct kg_sched *)&kg[1];
221 return (0);
222 }
223
224 void
225 ksegrp_link(struct ksegrp *kg, struct proc *p)
226 {
227
228 TAILQ_INIT(&kg->kg_threads);
229 TAILQ_INIT(&kg->kg_runq); /* links with td_runq */
230 TAILQ_INIT(&kg->kg_upcalls); /* all upcall structure in ksegrp */
231 kg->kg_proc = p;
232 /*
233 * the following counters are in the -zero- section
234 * and may not need clearing
235 */
236 kg->kg_numthreads = 0;
237 kg->kg_numupcalls = 0;
238 /* link it in now that it's consistent */
239 p->p_numksegrps++;
240 TAILQ_INSERT_HEAD(&p->p_ksegrps, kg, kg_ksegrp);
241 }
242
243 /*
244 * Called from:
245 * thread-exit()
246 */
247 void
248 ksegrp_unlink(struct ksegrp *kg)
249 {
250 struct proc *p;
251
252 mtx_assert(&sched_lock, MA_OWNED);
253 KASSERT((kg->kg_numthreads == 0), ("ksegrp_unlink: residual threads"));
254 KASSERT((kg->kg_numupcalls == 0), ("ksegrp_unlink: residual upcalls"));
255
256 p = kg->kg_proc;
257 TAILQ_REMOVE(&p->p_ksegrps, kg, kg_ksegrp);
258 p->p_numksegrps--;
259 /*
260 * Aggregate stats from the KSE
261 */
262 if (p->p_procscopegrp == kg)
263 p->p_procscopegrp = NULL;
264 }
265
266 /*
267 * For a newly created process,
268 * link up all the structures and its initial threads etc.
269 * called from:
270 * {arch}/{arch}/machdep.c ia64_init(), init386() etc.
271 * proc_dtor() (should go away)
272 * proc_init()
273 */
274 void
275 proc_linkup(struct proc *p, struct ksegrp *kg, struct thread *td)
276 {
277
278 TAILQ_INIT(&p->p_ksegrps); /* all ksegrps in proc */
279 TAILQ_INIT(&p->p_threads); /* all threads in proc */
280 TAILQ_INIT(&p->p_suspended); /* Threads suspended */
281 p->p_numksegrps = 0;
282 p->p_numthreads = 0;
283
284 ksegrp_link(kg, p);
285 thread_link(td, kg);
286 }
287
288 /*
289 * Initialize global thread allocation resources.
290 */
291 void
292 threadinit(void)
293 {
294
295 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
296 tid_unrhdr = new_unrhdr(PID_MAX + 1, INT_MAX, &tid_lock);
297
298 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
299 thread_ctor, thread_dtor, thread_init, thread_fini,
300 UMA_ALIGN_CACHE, 0);
301 ksegrp_zone = uma_zcreate("KSEGRP", sched_sizeof_ksegrp(),
302 ksegrp_ctor, NULL, NULL, NULL,
303 UMA_ALIGN_CACHE, 0);
304 kseinit(); /* set up kse specific stuff e.g. upcall zone*/
305 }
306
307 /*
308 * Stash an embarasingly extra thread into the zombie thread queue.
309 */
310 void
311 thread_stash(struct thread *td)
312 {
313 mtx_lock_spin(&kse_zombie_lock);
314 TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq);
315 mtx_unlock_spin(&kse_zombie_lock);
316 }
317
318 /*
319 * Stash an embarasingly extra ksegrp into the zombie ksegrp queue.
320 */
321 void
322 ksegrp_stash(struct ksegrp *kg)
323 {
324 mtx_lock_spin(&kse_zombie_lock);
325 TAILQ_INSERT_HEAD(&zombie_ksegrps, kg, kg_ksegrp);
326 mtx_unlock_spin(&kse_zombie_lock);
327 }
328
329 /*
330 * Reap zombie kse resource.
331 */
332 void
333 thread_reap(void)
334 {
335 struct thread *td_first, *td_next;
336 struct ksegrp *kg_first, * kg_next;
337
338 /*
339 * Don't even bother to lock if none at this instant,
340 * we really don't care about the next instant..
341 */
342 if ((!TAILQ_EMPTY(&zombie_threads))
343 || (!TAILQ_EMPTY(&zombie_ksegrps))) {
344 mtx_lock_spin(&kse_zombie_lock);
345 td_first = TAILQ_FIRST(&zombie_threads);
346 kg_first = TAILQ_FIRST(&zombie_ksegrps);
347 if (td_first)
348 TAILQ_INIT(&zombie_threads);
349 if (kg_first)
350 TAILQ_INIT(&zombie_ksegrps);
351 mtx_unlock_spin(&kse_zombie_lock);
352 while (td_first) {
353 td_next = TAILQ_NEXT(td_first, td_runq);
354 if (td_first->td_ucred)
355 crfree(td_first->td_ucred);
356 thread_free(td_first);
357 td_first = td_next;
358 }
359 while (kg_first) {
360 kg_next = TAILQ_NEXT(kg_first, kg_ksegrp);
361 ksegrp_free(kg_first);
362 kg_first = kg_next;
363 }
364 /*
365 * there will always be a thread on the list if one of these
366 * is there.
367 */
368 kse_GC();
369 }
370 }
371
372 /*
373 * Allocate a ksegrp.
374 */
375 struct ksegrp *
376 ksegrp_alloc(void)
377 {
378 return (uma_zalloc(ksegrp_zone, M_WAITOK));
379 }
380
381 /*
382 * Allocate a thread.
383 */
384 struct thread *
385 thread_alloc(void)
386 {
387 thread_reap(); /* check if any zombies to get */
388 return (uma_zalloc(thread_zone, M_WAITOK));
389 }
390
391 /*
392 * Deallocate a ksegrp.
393 */
394 void
395 ksegrp_free(struct ksegrp *td)
396 {
397 uma_zfree(ksegrp_zone, td);
398 }
399
400 /*
401 * Deallocate a thread.
402 */
403 void
404 thread_free(struct thread *td)
405 {
406
407 cpu_thread_clean(td);
408 uma_zfree(thread_zone, td);
409 }
410
411 /*
412 * Discard the current thread and exit from its context.
413 * Always called with scheduler locked.
414 *
415 * Because we can't free a thread while we're operating under its context,
416 * push the current thread into our CPU's deadthread holder. This means
417 * we needn't worry about someone else grabbing our context before we
418 * do a cpu_throw(). This may not be needed now as we are under schedlock.
419 * Maybe we can just do a thread_stash() as thr_exit1 does.
420 */
421 /* XXX
422 * libthr expects its thread exit to return for the last
423 * thread, meaning that the program is back to non-threaded
424 * mode I guess. Because we do this (cpu_throw) unconditionally
425 * here, they have their own version of it. (thr_exit1())
426 * that doesn't do it all if this was the last thread.
427 * It is also called from thread_suspend_check().
428 * Of course in the end, they end up coming here through exit1
429 * anyhow.. After fixing 'thr' to play by the rules we should be able
430 * to merge these two functions together.
431 *
432 * called from:
433 * exit1()
434 * kse_exit()
435 * thr_exit()
436 * thread_user_enter()
437 * thread_userret()
438 * thread_suspend_check()
439 */
440 void
441 thread_exit(void)
442 {
443 struct bintime new_switchtime;
444 struct thread *td;
445 struct proc *p;
446 struct ksegrp *kg;
447
448 td = curthread;
449 kg = td->td_ksegrp;
450 p = td->td_proc;
451
452 mtx_assert(&sched_lock, MA_OWNED);
453 mtx_assert(&Giant, MA_NOTOWNED);
454 PROC_LOCK_ASSERT(p, MA_OWNED);
455 KASSERT(p != NULL, ("thread exiting without a process"));
456 KASSERT(kg != NULL, ("thread exiting without a kse group"));
457 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
458 (long)p->p_pid, p->p_comm);
459
460 if (td->td_standin != NULL) {
461 /*
462 * Note that we don't need to free the cred here as it
463 * is done in thread_reap().
464 */
465 thread_stash(td->td_standin);
466 td->td_standin = NULL;
467 }
468
469 /*
470 * drop FPU & debug register state storage, or any other
471 * architecture specific resources that
472 * would not be on a new untouched process.
473 */
474 cpu_thread_exit(td); /* XXXSMP */
475
476 /*
477 * The thread is exiting. scheduler can release its stuff
478 * and collect stats etc.
479 */
480 sched_thread_exit(td);
481
482 /* Do the same timestamp bookkeeping that mi_switch() would do. */
483 binuptime(&new_switchtime);
484 bintime_add(&p->p_rux.rux_runtime, &new_switchtime);
485 bintime_sub(&p->p_rux.rux_runtime, PCPU_PTR(switchtime));
486 PCPU_SET(switchtime, new_switchtime);
487 PCPU_SET(switchticks, ticks);
488 cnt.v_swtch++;
489
490 /* Add our usage into the usage of all our children. */
491 if (p->p_numthreads == 1)
492 ruadd(p->p_ru, &p->p_rux, &p->p_stats->p_cru, &p->p_crux);
493
494 /*
495 * The last thread is left attached to the process
496 * So that the whole bundle gets recycled. Skip
497 * all this stuff if we never had threads.
498 * EXIT clears all sign of other threads when
499 * it goes to single threading, so the last thread always
500 * takes the short path.
501 */
502 if (p->p_flag & P_HADTHREADS) {
503 if (p->p_numthreads > 1) {
504 thread_unlink(td);
505
506 /* XXX first arg not used in 4BSD or ULE */
507 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
508
509 /*
510 * The test below is NOT true if we are the
511 * sole exiting thread. P_STOPPED_SNGL is unset
512 * in exit1() after it is the only survivor.
513 */
514 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
515 if (p->p_numthreads == p->p_suspcount) {
516 thread_unsuspend_one(p->p_singlethread);
517 }
518 }
519
520 /*
521 * Because each upcall structure has an owner thread,
522 * owner thread exits only when process is in exiting
523 * state, so upcall to userland is no longer needed,
524 * deleting upcall structure is safe here.
525 * So when all threads in a group is exited, all upcalls
526 * in the group should be automatically freed.
527 * XXXKSE This is a KSE thing and should be exported
528 * there somehow.
529 */
530 upcall_remove(td);
531
532 /*
533 * If the thread we unlinked above was the last one,
534 * then this ksegrp should go away too.
535 */
536 if (kg->kg_numthreads == 0) {
537 /*
538 * let the scheduler know about this in case
539 * it needs to recover stats or resources.
540 * Theoretically we could let
541 * sched_exit_ksegrp() do the equivalent of
542 * setting the concurrency to 0
543 * but don't do it yet to avoid changing
544 * the existing scheduler code until we
545 * are ready.
546 * We supply a random other ksegrp
547 * as the recipient of any built up
548 * cpu usage etc. (If the scheduler wants it).
549 * XXXKSE
550 * This is probably not fair so think of
551 * a better answer.
552 */
553 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), td);
554 sched_set_concurrency(kg, 0); /* XXX TEMP */
555 ksegrp_unlink(kg);
556 ksegrp_stash(kg);
557 }
558 PROC_UNLOCK(p);
559 td->td_ksegrp = NULL;
560 PCPU_SET(deadthread, td);
561 } else {
562 /*
563 * The last thread is exiting.. but not through exit()
564 * what should we do?
565 * Theoretically this can't happen
566 * exit1() - clears threading flags before coming here
567 * kse_exit() - treats last thread specially
568 * thr_exit() - treats last thread specially
569 * thread_user_enter() - only if more exist
570 * thread_userret() - only if more exist
571 * thread_suspend_check() - only if more exist
572 */
573 panic ("thread_exit: Last thread exiting on its own");
574 }
575 } else {
576 /*
577 * non threaded process comes here.
578 * This includes an EX threaded process that is coming
579 * here via exit1(). (exit1 dethreads the proc first).
580 */
581 PROC_UNLOCK(p);
582 }
583 td->td_state = TDS_INACTIVE;
584 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
585 cpu_throw(td, choosethread());
586 panic("I'm a teapot!");
587 /* NOTREACHED */
588 }
589
590 /*
591 * Do any thread specific cleanups that may be needed in wait()
592 * called with Giant, proc and schedlock not held.
593 */
594 void
595 thread_wait(struct proc *p)
596 {
597 struct thread *td;
598
599 mtx_assert(&Giant, MA_NOTOWNED);
600 KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
601 KASSERT((p->p_numksegrps == 1), ("Multiple ksegrps in wait1()"));
602 FOREACH_THREAD_IN_PROC(p, td) {
603 if (td->td_standin != NULL) {
604 if (td->td_standin->td_ucred != NULL) {
605 crfree(td->td_standin->td_ucred);
606 td->td_standin->td_ucred = NULL;
607 }
608 thread_free(td->td_standin);
609 td->td_standin = NULL;
610 }
611 cpu_thread_clean(td);
612 crfree(td->td_ucred);
613 }
614 thread_reap(); /* check for zombie threads etc. */
615 }
616
617 /*
618 * Link a thread to a process.
619 * set up anything that needs to be initialized for it to
620 * be used by the process.
621 *
622 * Note that we do not link to the proc's ucred here.
623 * The thread is linked as if running but no KSE assigned.
624 * Called from:
625 * proc_linkup()
626 * thread_schedule_upcall()
627 * thr_create()
628 */
629 void
630 thread_link(struct thread *td, struct ksegrp *kg)
631 {
632 struct proc *p;
633
634 p = kg->kg_proc;
635 td->td_state = TDS_INACTIVE;
636 td->td_proc = p;
637 td->td_ksegrp = kg;
638 td->td_flags = 0;
639 td->td_kflags = 0;
640
641 LIST_INIT(&td->td_contested);
642 callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
643 TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
644 TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist);
645 p->p_numthreads++;
646 kg->kg_numthreads++;
647 }
648
649 /*
650 * Convert a process with one thread to an unthreaded process.
651 * Called from:
652 * thread_single(exit) (called from execve and exit)
653 * kse_exit() XXX may need cleaning up wrt KSE stuff
654 */
655 void
656 thread_unthread(struct thread *td)
657 {
658 struct proc *p = td->td_proc;
659
660 KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
661 upcall_remove(td);
662 p->p_flag &= ~(P_SA|P_HADTHREADS);
663 td->td_mailbox = NULL;
664 td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
665 if (td->td_standin != NULL) {
666 thread_stash(td->td_standin);
667 td->td_standin = NULL;
668 }
669 sched_set_concurrency(td->td_ksegrp, 1);
670 }
671
672 /*
673 * Called from:
674 * thread_exit()
675 */
676 void
677 thread_unlink(struct thread *td)
678 {
679 struct proc *p = td->td_proc;
680 struct ksegrp *kg = td->td_ksegrp;
681
682 mtx_assert(&sched_lock, MA_OWNED);
683 TAILQ_REMOVE(&p->p_threads, td, td_plist);
684 p->p_numthreads--;
685 TAILQ_REMOVE(&kg->kg_threads, td, td_kglist);
686 kg->kg_numthreads--;
687 /* could clear a few other things here */
688 /* Must NOT clear links to proc and ksegrp! */
689 }
690
691 /*
692 * Enforce single-threading.
693 *
694 * Returns 1 if the caller must abort (another thread is waiting to
695 * exit the process or similar). Process is locked!
696 * Returns 0 when you are successfully the only thread running.
697 * A process has successfully single threaded in the suspend mode when
698 * There are no threads in user mode. Threads in the kernel must be
699 * allowed to continue until they get to the user boundary. They may even
700 * copy out their return values and data before suspending. They may however be
701 * accellerated in reaching the user boundary as we will wake up
702 * any sleeping threads that are interruptable. (PCATCH).
703 */
704 int
705 thread_single(int mode)
706 {
707 struct thread *td;
708 struct thread *td2;
709 struct proc *p;
710 int remaining;
711
712 td = curthread;
713 p = td->td_proc;
714 mtx_assert(&Giant, MA_NOTOWNED);
715 PROC_LOCK_ASSERT(p, MA_OWNED);
716 KASSERT((td != NULL), ("curthread is NULL"));
717
718 if ((p->p_flag & P_HADTHREADS) == 0)
719 return (0);
720
721 /* Is someone already single threading? */
722 if (p->p_singlethread != NULL && p->p_singlethread != td)
723 return (1);
724
725 if (mode == SINGLE_EXIT) {
726 p->p_flag |= P_SINGLE_EXIT;
727 p->p_flag &= ~P_SINGLE_BOUNDARY;
728 } else {
729 p->p_flag &= ~P_SINGLE_EXIT;
730 if (mode == SINGLE_BOUNDARY)
731 p->p_flag |= P_SINGLE_BOUNDARY;
732 else
733 p->p_flag &= ~P_SINGLE_BOUNDARY;
734 }
735 p->p_flag |= P_STOPPED_SINGLE;
736 mtx_lock_spin(&sched_lock);
737 p->p_singlethread = td;
738 if (mode == SINGLE_EXIT)
739 remaining = p->p_numthreads;
740 else if (mode == SINGLE_BOUNDARY)
741 remaining = p->p_numthreads - p->p_boundary_count;
742 else
743 remaining = p->p_numthreads - p->p_suspcount;
744 while (remaining != 1) {
745 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
746 goto stopme;
747 FOREACH_THREAD_IN_PROC(p, td2) {
748 if (td2 == td)
749 continue;
750 td2->td_flags |= TDF_ASTPENDING;
751 if (TD_IS_INHIBITED(td2)) {
752 switch (mode) {
753 case SINGLE_EXIT:
754 if (td->td_flags & TDF_DBSUSPEND)
755 td->td_flags &= ~TDF_DBSUSPEND;
756 if (TD_IS_SUSPENDED(td2))
757 thread_unsuspend_one(td2);
758 if (TD_ON_SLEEPQ(td2) &&
759 (td2->td_flags & TDF_SINTR))
760 sleepq_abort(td2, EINTR);
761 break;
762 case SINGLE_BOUNDARY:
763 if (TD_IS_SUSPENDED(td2) &&
764 !(td2->td_flags & TDF_BOUNDARY))
765 thread_unsuspend_one(td2);
766 if (TD_ON_SLEEPQ(td2) &&
767 (td2->td_flags & TDF_SINTR))
768 sleepq_abort(td2, ERESTART);
769 break;
770 default:
771 if (TD_IS_SUSPENDED(td2))
772 continue;
773 /*
774 * maybe other inhibitted states too?
775 */
776 if ((td2->td_flags & TDF_SINTR) &&
777 (td2->td_inhibitors &
778 (TDI_SLEEPING | TDI_SWAPPED)))
779 thread_suspend_one(td2);
780 break;
781 }
782 }
783 #ifdef SMP
784 else if (TD_IS_RUNNING(td2) && td != td2) {
785 forward_signal(td2);
786 }
787 #endif
788 }
789 if (mode == SINGLE_EXIT)
790 remaining = p->p_numthreads;
791 else if (mode == SINGLE_BOUNDARY)
792 remaining = p->p_numthreads - p->p_boundary_count;
793 else
794 remaining = p->p_numthreads - p->p_suspcount;
795
796 /*
797 * Maybe we suspended some threads.. was it enough?
798 */
799 if (remaining == 1)
800 break;
801
802 stopme:
803 /*
804 * Wake us up when everyone else has suspended.
805 * In the mean time we suspend as well.
806 */
807 thread_stopped(p);
808 thread_suspend_one(td);
809 PROC_UNLOCK(p);
810 mi_switch(SW_VOL, NULL);
811 mtx_unlock_spin(&sched_lock);
812 PROC_LOCK(p);
813 mtx_lock_spin(&sched_lock);
814 if (mode == SINGLE_EXIT)
815 remaining = p->p_numthreads;
816 else if (mode == SINGLE_BOUNDARY)
817 remaining = p->p_numthreads - p->p_boundary_count;
818 else
819 remaining = p->p_numthreads - p->p_suspcount;
820 }
821 if (mode == SINGLE_EXIT) {
822 /*
823 * We have gotten rid of all the other threads and we
824 * are about to either exit or exec. In either case,
825 * we try our utmost to revert to being a non-threaded
826 * process.
827 */
828 p->p_singlethread = NULL;
829 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
830 thread_unthread(td);
831 }
832 mtx_unlock_spin(&sched_lock);
833 return (0);
834 }
835
836 /*
837 * Called in from locations that can safely check to see
838 * whether we have to suspend or at least throttle for a
839 * single-thread event (e.g. fork).
840 *
841 * Such locations include userret().
842 * If the "return_instead" argument is non zero, the thread must be able to
843 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
844 *
845 * The 'return_instead' argument tells the function if it may do a
846 * thread_exit() or suspend, or whether the caller must abort and back
847 * out instead.
848 *
849 * If the thread that set the single_threading request has set the
850 * P_SINGLE_EXIT bit in the process flags then this call will never return
851 * if 'return_instead' is false, but will exit.
852 *
853 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
854 *---------------+--------------------+---------------------
855 * 0 | returns 0 | returns 0 or 1
856 * | when ST ends | immediatly
857 *---------------+--------------------+---------------------
858 * 1 | thread exits | returns 1
859 * | | immediatly
860 * 0 = thread_exit() or suspension ok,
861 * other = return error instead of stopping the thread.
862 *
863 * While a full suspension is under effect, even a single threading
864 * thread would be suspended if it made this call (but it shouldn't).
865 * This call should only be made from places where
866 * thread_exit() would be safe as that may be the outcome unless
867 * return_instead is set.
868 */
869 int
870 thread_suspend_check(int return_instead)
871 {
872 struct thread *td;
873 struct proc *p;
874
875 td = curthread;
876 p = td->td_proc;
877 mtx_assert(&Giant, MA_NOTOWNED);
878 PROC_LOCK_ASSERT(p, MA_OWNED);
879 while (P_SHOULDSTOP(p) ||
880 ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
881 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
882 KASSERT(p->p_singlethread != NULL,
883 ("singlethread not set"));
884 /*
885 * The only suspension in action is a
886 * single-threading. Single threader need not stop.
887 * XXX Should be safe to access unlocked
888 * as it can only be set to be true by us.
889 */
890 if (p->p_singlethread == td)
891 return (0); /* Exempt from stopping. */
892 }
893 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
894 return (EINTR);
895
896 /* Should we goto user boundary if we didn't come from there? */
897 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
898 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
899 return (ERESTART);
900
901 mtx_lock_spin(&sched_lock);
902 thread_stopped(p);
903 /*
904 * If the process is waiting for us to exit,
905 * this thread should just suicide.
906 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
907 */
908 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
909 thread_exit();
910
911 /*
912 * When a thread suspends, it just
913 * moves to the processes's suspend queue
914 * and stays there.
915 */
916 thread_suspend_one(td);
917 if (return_instead == 0) {
918 p->p_boundary_count++;
919 td->td_flags |= TDF_BOUNDARY;
920 }
921 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
922 if (p->p_numthreads == p->p_suspcount)
923 thread_unsuspend_one(p->p_singlethread);
924 }
925 PROC_UNLOCK(p);
926 mi_switch(SW_INVOL, NULL);
927 if (return_instead == 0) {
928 p->p_boundary_count--;
929 td->td_flags &= ~TDF_BOUNDARY;
930 }
931 mtx_unlock_spin(&sched_lock);
932 PROC_LOCK(p);
933 }
934 return (0);
935 }
936
937 void
938 thread_suspend_one(struct thread *td)
939 {
940 struct proc *p = td->td_proc;
941
942 mtx_assert(&sched_lock, MA_OWNED);
943 PROC_LOCK_ASSERT(p, MA_OWNED);
944 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
945 p->p_suspcount++;
946 TD_SET_SUSPENDED(td);
947 TAILQ_INSERT_TAIL(&p->p_suspended, td, td_runq);
948 }
949
950 void
951 thread_unsuspend_one(struct thread *td)
952 {
953 struct proc *p = td->td_proc;
954
955 mtx_assert(&sched_lock, MA_OWNED);
956 PROC_LOCK_ASSERT(p, MA_OWNED);
957 TAILQ_REMOVE(&p->p_suspended, td, td_runq);
958 TD_CLR_SUSPENDED(td);
959 p->p_suspcount--;
960 setrunnable(td);
961 }
962
963 /*
964 * Allow all threads blocked by single threading to continue running.
965 */
966 void
967 thread_unsuspend(struct proc *p)
968 {
969 struct thread *td;
970
971 mtx_assert(&sched_lock, MA_OWNED);
972 PROC_LOCK_ASSERT(p, MA_OWNED);
973 if (!P_SHOULDSTOP(p)) {
974 while ((td = TAILQ_FIRST(&p->p_suspended))) {
975 thread_unsuspend_one(td);
976 }
977 } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
978 (p->p_numthreads == p->p_suspcount)) {
979 /*
980 * Stopping everything also did the job for the single
981 * threading request. Now we've downgraded to single-threaded,
982 * let it continue.
983 */
984 thread_unsuspend_one(p->p_singlethread);
985 }
986 }
987
988 /*
989 * End the single threading mode..
990 */
991 void
992 thread_single_end(void)
993 {
994 struct thread *td;
995 struct proc *p;
996
997 td = curthread;
998 p = td->td_proc;
999 PROC_LOCK_ASSERT(p, MA_OWNED);
1000 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
1001 mtx_lock_spin(&sched_lock);
1002 p->p_singlethread = NULL;
1003 p->p_procscopegrp = NULL;
1004 /*
1005 * If there are other threads they mey now run,
1006 * unless of course there is a blanket 'stop order'
1007 * on the process. The single threader must be allowed
1008 * to continue however as this is a bad place to stop.
1009 */
1010 if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
1011 while ((td = TAILQ_FIRST(&p->p_suspended))) {
1012 thread_unsuspend_one(td);
1013 }
1014 }
1015 mtx_unlock_spin(&sched_lock);
1016 }
1017
1018 /*
1019 * Called before going into an interruptible sleep to see if we have been
1020 * interrupted or requested to exit.
1021 */
1022 int
1023 thread_sleep_check(struct thread *td)
1024 {
1025 struct proc *p;
1026
1027 p = td->td_proc;
1028 mtx_assert(&sched_lock, MA_OWNED);
1029 if (p->p_flag & P_HADTHREADS) {
1030 if (p->p_singlethread != td) {
1031 if (p->p_flag & P_SINGLE_EXIT)
1032 return (EINTR);
1033 if (p->p_flag & P_SINGLE_BOUNDARY)
1034 return (ERESTART);
1035 }
1036 if (td->td_flags & TDF_INTERRUPT)
1037 return (td->td_intrval);
1038 }
1039 return (0);
1040 }
Cache object: ee8c0ad8cad57fdde39c9335bb7b8817
|