1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice(s), this list of conditions and the following disclaimer as
12 * the first lines of this file unmodified other than the possible
13 * addition of one or more copyright notices.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice(s), this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 *
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
19 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
22 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
25 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
28 * DAMAGE.
29 */
30
31 #include "opt_witness.h"
32 #include "opt_hwpmc_hooks.h"
33
34 #include <sys/systm.h>
35 #include <sys/asan.h>
36 #include <sys/kernel.h>
37 #include <sys/lock.h>
38 #include <sys/msan.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/bitstring.h>
42 #include <sys/epoch.h>
43 #include <sys/rangelock.h>
44 #include <sys/resourcevar.h>
45 #include <sys/sdt.h>
46 #include <sys/smp.h>
47 #include <sys/sched.h>
48 #include <sys/sleepqueue.h>
49 #include <sys/selinfo.h>
50 #include <sys/syscallsubr.h>
51 #include <sys/dtrace_bsd.h>
52 #include <sys/sysent.h>
53 #include <sys/turnstile.h>
54 #include <sys/taskqueue.h>
55 #include <sys/ktr.h>
56 #include <sys/rwlock.h>
57 #include <sys/umtxvar.h>
58 #include <sys/vmmeter.h>
59 #include <sys/cpuset.h>
60 #ifdef HWPMC_HOOKS
61 #include <sys/pmckern.h>
62 #endif
63 #include <sys/priv.h>
64
65 #include <security/audit/audit.h>
66
67 #include <vm/pmap.h>
68 #include <vm/vm.h>
69 #include <vm/vm_extern.h>
70 #include <vm/uma.h>
71 #include <vm/vm_phys.h>
72 #include <sys/eventhandler.h>
73
74 /*
75 * Asserts below verify the stability of struct thread and struct proc
76 * layout, as exposed by KBI to modules. On head, the KBI is allowed
77 * to drift, change to the structures must be accompanied by the
78 * assert update.
79 *
80 * On the stable branches after KBI freeze, conditions must not be
81 * violated. Typically new fields are moved to the end of the
82 * structures.
83 */
84 #ifdef __amd64__
85 _Static_assert(offsetof(struct thread, td_flags) == 0x108,
86 "struct thread KBI td_flags");
87 _Static_assert(offsetof(struct thread, td_pflags) == 0x114,
88 "struct thread KBI td_pflags");
89 _Static_assert(offsetof(struct thread, td_frame) == 0x4b8,
90 "struct thread KBI td_frame");
91 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6c0,
92 "struct thread KBI td_emuldata");
93 _Static_assert(offsetof(struct proc, p_flag) == 0xb8,
94 "struct proc KBI p_flag");
95 _Static_assert(offsetof(struct proc, p_pid) == 0xc4,
96 "struct proc KBI p_pid");
97 _Static_assert(offsetof(struct proc, p_filemon) == 0x3c8,
98 "struct proc KBI p_filemon");
99 _Static_assert(offsetof(struct proc, p_comm) == 0x3e0,
100 "struct proc KBI p_comm");
101 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4d0,
102 "struct proc KBI p_emuldata");
103 #endif
104 #ifdef __i386__
105 _Static_assert(offsetof(struct thread, td_flags) == 0x9c,
106 "struct thread KBI td_flags");
107 _Static_assert(offsetof(struct thread, td_pflags) == 0xa8,
108 "struct thread KBI td_pflags");
109 _Static_assert(offsetof(struct thread, td_frame) == 0x318,
110 "struct thread KBI td_frame");
111 _Static_assert(offsetof(struct thread, td_emuldata) == 0x35c,
112 "struct thread KBI td_emuldata");
113 _Static_assert(offsetof(struct proc, p_flag) == 0x6c,
114 "struct proc KBI p_flag");
115 _Static_assert(offsetof(struct proc, p_pid) == 0x78,
116 "struct proc KBI p_pid");
117 _Static_assert(offsetof(struct proc, p_filemon) == 0x270,
118 "struct proc KBI p_filemon");
119 _Static_assert(offsetof(struct proc, p_comm) == 0x284,
120 "struct proc KBI p_comm");
121 _Static_assert(offsetof(struct proc, p_emuldata) == 0x318,
122 "struct proc KBI p_emuldata");
123 #endif
124
125 SDT_PROVIDER_DECLARE(proc);
126 SDT_PROBE_DEFINE(proc, , , lwp__exit);
127
128 /*
129 * thread related storage.
130 */
131 static uma_zone_t thread_zone;
132
133 struct thread_domain_data {
134 struct thread *tdd_zombies;
135 int tdd_reapticks;
136 } __aligned(CACHE_LINE_SIZE);
137
138 static struct thread_domain_data thread_domain_data[MAXMEMDOM];
139
140 static struct task thread_reap_task;
141 static struct callout thread_reap_callout;
142
143 static void thread_zombie(struct thread *);
144 static void thread_reap(void);
145 static void thread_reap_all(void);
146 static void thread_reap_task_cb(void *, int);
147 static void thread_reap_callout_cb(void *);
148 static int thread_unsuspend_one(struct thread *td, struct proc *p,
149 bool boundary);
150 static void thread_free_batched(struct thread *td);
151
152 static __exclusive_cache_line struct mtx tid_lock;
153 static bitstr_t *tid_bitmap;
154
155 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
156
157 static int maxthread;
158 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
159 &maxthread, 0, "Maximum number of threads");
160
161 static __exclusive_cache_line int nthreads;
162
163 static LIST_HEAD(tidhashhead, thread) *tidhashtbl;
164 static u_long tidhash;
165 static u_long tidhashlock;
166 static struct rwlock *tidhashtbl_lock;
167 #define TIDHASH(tid) (&tidhashtbl[(tid) & tidhash])
168 #define TIDHASHLOCK(tid) (&tidhashtbl_lock[(tid) & tidhashlock])
169
170 EVENTHANDLER_LIST_DEFINE(thread_ctor);
171 EVENTHANDLER_LIST_DEFINE(thread_dtor);
172 EVENTHANDLER_LIST_DEFINE(thread_init);
173 EVENTHANDLER_LIST_DEFINE(thread_fini);
174
175 static bool
thread_count_inc_try(void)176 thread_count_inc_try(void)
177 {
178 int nthreads_new;
179
180 nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1;
181 if (nthreads_new >= maxthread - 100) {
182 if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
183 nthreads_new >= maxthread) {
184 atomic_subtract_int(&nthreads, 1);
185 return (false);
186 }
187 }
188 return (true);
189 }
190
191 static bool
thread_count_inc(void)192 thread_count_inc(void)
193 {
194 static struct timeval lastfail;
195 static int curfail;
196
197 thread_reap();
198 if (thread_count_inc_try()) {
199 return (true);
200 }
201
202 thread_reap_all();
203 if (thread_count_inc_try()) {
204 return (true);
205 }
206
207 if (ppsratecheck(&lastfail, &curfail, 1)) {
208 printf("maxthread limit exceeded by uid %u "
209 "(pid %d); consider increasing kern.maxthread\n",
210 curthread->td_ucred->cr_ruid, curproc->p_pid);
211 }
212 return (false);
213 }
214
215 static void
thread_count_sub(int n)216 thread_count_sub(int n)
217 {
218
219 atomic_subtract_int(&nthreads, n);
220 }
221
222 static void
thread_count_dec(void)223 thread_count_dec(void)
224 {
225
226 thread_count_sub(1);
227 }
228
229 static lwpid_t
tid_alloc(void)230 tid_alloc(void)
231 {
232 static lwpid_t trytid;
233 lwpid_t tid;
234
235 mtx_lock(&tid_lock);
236 /*
237 * It is an invariant that the bitmap is big enough to hold maxthread
238 * IDs. If we got to this point there has to be at least one free.
239 */
240 if (trytid >= maxthread)
241 trytid = 0;
242 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
243 if (tid == -1) {
244 KASSERT(trytid != 0, ("unexpectedly ran out of IDs"));
245 trytid = 0;
246 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
247 KASSERT(tid != -1, ("unexpectedly ran out of IDs"));
248 }
249 bit_set(tid_bitmap, tid);
250 trytid = tid + 1;
251 mtx_unlock(&tid_lock);
252 return (tid + NO_PID);
253 }
254
255 static void
tid_free_locked(lwpid_t rtid)256 tid_free_locked(lwpid_t rtid)
257 {
258 lwpid_t tid;
259
260 mtx_assert(&tid_lock, MA_OWNED);
261 KASSERT(rtid >= NO_PID,
262 ("%s: invalid tid %d\n", __func__, rtid));
263 tid = rtid - NO_PID;
264 KASSERT(bit_test(tid_bitmap, tid) != 0,
265 ("thread ID %d not allocated\n", rtid));
266 bit_clear(tid_bitmap, tid);
267 }
268
269 static void
tid_free(lwpid_t rtid)270 tid_free(lwpid_t rtid)
271 {
272
273 mtx_lock(&tid_lock);
274 tid_free_locked(rtid);
275 mtx_unlock(&tid_lock);
276 }
277
278 static void
tid_free_batch(lwpid_t * batch,int n)279 tid_free_batch(lwpid_t *batch, int n)
280 {
281 int i;
282
283 mtx_lock(&tid_lock);
284 for (i = 0; i < n; i++) {
285 tid_free_locked(batch[i]);
286 }
287 mtx_unlock(&tid_lock);
288 }
289
290 /*
291 * Batching for thread reapping.
292 */
293 struct tidbatch {
294 lwpid_t tab[16];
295 int n;
296 };
297
298 static void
tidbatch_prep(struct tidbatch * tb)299 tidbatch_prep(struct tidbatch *tb)
300 {
301
302 tb->n = 0;
303 }
304
305 static void
tidbatch_add(struct tidbatch * tb,struct thread * td)306 tidbatch_add(struct tidbatch *tb, struct thread *td)
307 {
308
309 KASSERT(tb->n < nitems(tb->tab),
310 ("%s: count too high %d", __func__, tb->n));
311 tb->tab[tb->n] = td->td_tid;
312 tb->n++;
313 }
314
315 static void
tidbatch_process(struct tidbatch * tb)316 tidbatch_process(struct tidbatch *tb)
317 {
318
319 KASSERT(tb->n <= nitems(tb->tab),
320 ("%s: count too high %d", __func__, tb->n));
321 if (tb->n == nitems(tb->tab)) {
322 tid_free_batch(tb->tab, tb->n);
323 tb->n = 0;
324 }
325 }
326
327 static void
tidbatch_final(struct tidbatch * tb)328 tidbatch_final(struct tidbatch *tb)
329 {
330
331 KASSERT(tb->n <= nitems(tb->tab),
332 ("%s: count too high %d", __func__, tb->n));
333 if (tb->n != 0) {
334 tid_free_batch(tb->tab, tb->n);
335 }
336 }
337
338 /*
339 * Batching thread count free, for consistency
340 */
341 struct tdcountbatch {
342 int n;
343 };
344
345 static void
tdcountbatch_prep(struct tdcountbatch * tb)346 tdcountbatch_prep(struct tdcountbatch *tb)
347 {
348
349 tb->n = 0;
350 }
351
352 static void
tdcountbatch_add(struct tdcountbatch * tb,struct thread * td __unused)353 tdcountbatch_add(struct tdcountbatch *tb, struct thread *td __unused)
354 {
355
356 tb->n++;
357 }
358
359 static void
tdcountbatch_process(struct tdcountbatch * tb)360 tdcountbatch_process(struct tdcountbatch *tb)
361 {
362
363 if (tb->n == 32) {
364 thread_count_sub(tb->n);
365 tb->n = 0;
366 }
367 }
368
369 static void
tdcountbatch_final(struct tdcountbatch * tb)370 tdcountbatch_final(struct tdcountbatch *tb)
371 {
372
373 if (tb->n != 0) {
374 thread_count_sub(tb->n);
375 }
376 }
377
378 /*
379 * Prepare a thread for use.
380 */
381 static int
thread_ctor(void * mem,int size,void * arg,int flags)382 thread_ctor(void *mem, int size, void *arg, int flags)
383 {
384 struct thread *td;
385
386 td = (struct thread *)mem;
387 TD_SET_STATE(td, TDS_INACTIVE);
388 td->td_lastcpu = td->td_oncpu = NOCPU;
389
390 /*
391 * Note that td_critnest begins life as 1 because the thread is not
392 * running and is thereby implicitly waiting to be on the receiving
393 * end of a context switch.
394 */
395 td->td_critnest = 1;
396 td->td_lend_user_pri = PRI_MAX;
397 #ifdef AUDIT
398 audit_thread_alloc(td);
399 #endif
400 #ifdef KDTRACE_HOOKS
401 kdtrace_thread_ctor(td);
402 #endif
403 umtx_thread_alloc(td);
404 MPASS(td->td_sel == NULL);
405 return (0);
406 }
407
408 /*
409 * Reclaim a thread after use.
410 */
411 static void
thread_dtor(void * mem,int size,void * arg)412 thread_dtor(void *mem, int size, void *arg)
413 {
414 struct thread *td;
415
416 td = (struct thread *)mem;
417
418 #ifdef INVARIANTS
419 /* Verify that this thread is in a safe state to free. */
420 switch (TD_GET_STATE(td)) {
421 case TDS_INHIBITED:
422 case TDS_RUNNING:
423 case TDS_CAN_RUN:
424 case TDS_RUNQ:
425 /*
426 * We must never unlink a thread that is in one of
427 * these states, because it is currently active.
428 */
429 panic("bad state for thread unlinking");
430 /* NOTREACHED */
431 case TDS_INACTIVE:
432 break;
433 default:
434 panic("bad thread state");
435 /* NOTREACHED */
436 }
437 #endif
438 #ifdef AUDIT
439 audit_thread_free(td);
440 #endif
441 #ifdef KDTRACE_HOOKS
442 kdtrace_thread_dtor(td);
443 #endif
444 /* Free all OSD associated to this thread. */
445 osd_thread_exit(td);
446 ast_kclear(td);
447 seltdfini(td);
448 }
449
450 /*
451 * Initialize type-stable parts of a thread (when newly created).
452 */
453 static int
thread_init(void * mem,int size,int flags)454 thread_init(void *mem, int size, int flags)
455 {
456 struct thread *td;
457
458 td = (struct thread *)mem;
459
460 td->td_allocdomain = vm_phys_domain(vtophys(td));
461 td->td_sleepqueue = sleepq_alloc();
462 td->td_turnstile = turnstile_alloc();
463 td->td_rlqe = NULL;
464 EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
465 umtx_thread_init(td);
466 td->td_kstack = 0;
467 td->td_sel = NULL;
468 return (0);
469 }
470
471 /*
472 * Tear down type-stable parts of a thread (just before being discarded).
473 */
474 static void
thread_fini(void * mem,int size)475 thread_fini(void *mem, int size)
476 {
477 struct thread *td;
478
479 td = (struct thread *)mem;
480 EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
481 rlqentry_free(td->td_rlqe);
482 turnstile_free(td->td_turnstile);
483 sleepq_free(td->td_sleepqueue);
484 umtx_thread_fini(td);
485 MPASS(td->td_sel == NULL);
486 }
487
488 /*
489 * For a newly created process,
490 * link up all the structures and its initial threads etc.
491 * called from:
492 * {arch}/{arch}/machdep.c {arch}_init(), init386() etc.
493 * proc_dtor() (should go away)
494 * proc_init()
495 */
496 void
proc_linkup0(struct proc * p,struct thread * td)497 proc_linkup0(struct proc *p, struct thread *td)
498 {
499 TAILQ_INIT(&p->p_threads); /* all threads in proc */
500 proc_linkup(p, td);
501 }
502
503 void
proc_linkup(struct proc * p,struct thread * td)504 proc_linkup(struct proc *p, struct thread *td)
505 {
506
507 sigqueue_init(&p->p_sigqueue, p);
508 p->p_ksi = ksiginfo_alloc(M_WAITOK);
509 if (p->p_ksi != NULL) {
510 /* XXX p_ksi may be null if ksiginfo zone is not ready */
511 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
512 }
513 LIST_INIT(&p->p_mqnotifier);
514 p->p_numthreads = 0;
515 thread_link(td, p);
516 }
517
518 static void
ast_suspend(struct thread * td,int tda __unused)519 ast_suspend(struct thread *td, int tda __unused)
520 {
521 struct proc *p;
522
523 p = td->td_proc;
524 /*
525 * We need to check to see if we have to exit or wait due to a
526 * single threading requirement or some other STOP condition.
527 */
528 PROC_LOCK(p);
529 thread_suspend_check(0);
530 PROC_UNLOCK(p);
531 }
532
533 extern int max_threads_per_proc;
534
535 /*
536 * Initialize global thread allocation resources.
537 */
538 void
threadinit(void)539 threadinit(void)
540 {
541 u_long i;
542 lwpid_t tid0;
543
544 /*
545 * Place an upper limit on threads which can be allocated.
546 *
547 * Note that other factors may make the de facto limit much lower.
548 *
549 * Platform limits are somewhat arbitrary but deemed "more than good
550 * enough" for the foreseable future.
551 */
552 if (maxthread == 0) {
553 #ifdef _LP64
554 maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
555 #else
556 maxthread = MIN(maxproc * max_threads_per_proc, 100000);
557 #endif
558 }
559
560 mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
561 tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
562 /*
563 * Handle thread0.
564 */
565 thread_count_inc();
566 tid0 = tid_alloc();
567 if (tid0 != THREAD0_TID)
568 panic("tid0 %d != %d\n", tid0, THREAD0_TID);
569
570 /*
571 * Thread structures are specially aligned so that (at least) the
572 * 5 lower bits of a pointer to 'struct thead' must be 0. These bits
573 * are used by synchronization primitives to store flags in pointers to
574 * such structures.
575 */
576 thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
577 thread_ctor, thread_dtor, thread_init, thread_fini,
578 UMA_ALIGN_CACHE_AND_MASK(32 - 1), UMA_ZONE_NOFREE);
579 tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
580 tidhashlock = (tidhash + 1) / 64;
581 if (tidhashlock > 0)
582 tidhashlock--;
583 tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
584 M_TIDHASH, M_WAITOK | M_ZERO);
585 for (i = 0; i < tidhashlock + 1; i++)
586 rw_init(&tidhashtbl_lock[i], "tidhash");
587
588 TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL);
589 callout_init(&thread_reap_callout, 1);
590 callout_reset(&thread_reap_callout, 5 * hz,
591 thread_reap_callout_cb, NULL);
592 ast_register(TDA_SUSPEND, ASTR_ASTF_REQUIRED, 0, ast_suspend);
593 }
594
595 /*
596 * Place an unused thread on the zombie list.
597 */
598 void
thread_zombie(struct thread * td)599 thread_zombie(struct thread *td)
600 {
601 struct thread_domain_data *tdd;
602 struct thread *ztd;
603
604 tdd = &thread_domain_data[td->td_allocdomain];
605 ztd = atomic_load_ptr(&tdd->tdd_zombies);
606 for (;;) {
607 td->td_zombie = ztd;
608 if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
609 (uintptr_t *)&ztd, (uintptr_t)td))
610 break;
611 continue;
612 }
613 }
614
615 /*
616 * Release a thread that has exited after cpu_throw().
617 */
618 void
thread_stash(struct thread * td)619 thread_stash(struct thread *td)
620 {
621 atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
622 thread_zombie(td);
623 }
624
625 /*
626 * Reap zombies from passed domain.
627 */
628 static void
thread_reap_domain(struct thread_domain_data * tdd)629 thread_reap_domain(struct thread_domain_data *tdd)
630 {
631 struct thread *itd, *ntd;
632 struct tidbatch tidbatch;
633 struct credbatch credbatch;
634 struct limbatch limbatch;
635 struct tdcountbatch tdcountbatch;
636
637 /*
638 * Reading upfront is pessimal if followed by concurrent atomic_swap,
639 * but most of the time the list is empty.
640 */
641 if (tdd->tdd_zombies == NULL)
642 return;
643
644 itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
645 (uintptr_t)NULL);
646 if (itd == NULL)
647 return;
648
649 /*
650 * Multiple CPUs can get here, the race is fine as ticks is only
651 * advisory.
652 */
653 tdd->tdd_reapticks = ticks;
654
655 tidbatch_prep(&tidbatch);
656 credbatch_prep(&credbatch);
657 limbatch_prep(&limbatch);
658 tdcountbatch_prep(&tdcountbatch);
659
660 while (itd != NULL) {
661 ntd = itd->td_zombie;
662 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd);
663
664 tidbatch_add(&tidbatch, itd);
665 credbatch_add(&credbatch, itd);
666 limbatch_add(&limbatch, itd);
667 tdcountbatch_add(&tdcountbatch, itd);
668
669 thread_free_batched(itd);
670
671 tidbatch_process(&tidbatch);
672 credbatch_process(&credbatch);
673 limbatch_process(&limbatch);
674 tdcountbatch_process(&tdcountbatch);
675
676 itd = ntd;
677 }
678
679 tidbatch_final(&tidbatch);
680 credbatch_final(&credbatch);
681 limbatch_final(&limbatch);
682 tdcountbatch_final(&tdcountbatch);
683 }
684
685 /*
686 * Reap zombies from all domains.
687 */
688 static void
thread_reap_all(void)689 thread_reap_all(void)
690 {
691 struct thread_domain_data *tdd;
692 int i, domain;
693
694 domain = PCPU_GET(domain);
695 for (i = 0; i < vm_ndomains; i++) {
696 tdd = &thread_domain_data[(i + domain) % vm_ndomains];
697 thread_reap_domain(tdd);
698 }
699 }
700
701 /*
702 * Reap zombies from local domain.
703 */
704 static void
thread_reap(void)705 thread_reap(void)
706 {
707 struct thread_domain_data *tdd;
708 int domain;
709
710 domain = PCPU_GET(domain);
711 tdd = &thread_domain_data[domain];
712
713 thread_reap_domain(tdd);
714 }
715
716 static void
thread_reap_task_cb(void * arg __unused,int pending __unused)717 thread_reap_task_cb(void *arg __unused, int pending __unused)
718 {
719
720 thread_reap_all();
721 }
722
723 static void
thread_reap_callout_cb(void * arg __unused)724 thread_reap_callout_cb(void *arg __unused)
725 {
726 struct thread_domain_data *tdd;
727 int i, cticks, lticks;
728 bool wantreap;
729
730 wantreap = false;
731 cticks = atomic_load_int(&ticks);
732 for (i = 0; i < vm_ndomains; i++) {
733 tdd = &thread_domain_data[i];
734 lticks = tdd->tdd_reapticks;
735 if (tdd->tdd_zombies != NULL &&
736 (u_int)(cticks - lticks) > 5 * hz) {
737 wantreap = true;
738 break;
739 }
740 }
741
742 if (wantreap)
743 taskqueue_enqueue(taskqueue_thread, &thread_reap_task);
744 callout_reset(&thread_reap_callout, 5 * hz,
745 thread_reap_callout_cb, NULL);
746 }
747
748 /*
749 * Calling this function guarantees that any thread that exited before
750 * the call is reaped when the function returns. By 'exited' we mean
751 * a thread removed from the process linkage with thread_unlink().
752 * Practically this means that caller must lock/unlock corresponding
753 * process lock before the call, to synchronize with thread_exit().
754 */
755 void
thread_reap_barrier(void)756 thread_reap_barrier(void)
757 {
758 struct task *t;
759
760 /*
761 * First do context switches to each CPU to ensure that all
762 * PCPU pc_deadthreads are moved to zombie list.
763 */
764 quiesce_all_cpus("", PDROP);
765
766 /*
767 * Second, fire the task in the same thread as normal
768 * thread_reap() is done, to serialize reaping.
769 */
770 t = malloc(sizeof(*t), M_TEMP, M_WAITOK);
771 TASK_INIT(t, 0, thread_reap_task_cb, t);
772 taskqueue_enqueue(taskqueue_thread, t);
773 taskqueue_drain(taskqueue_thread, t);
774 free(t, M_TEMP);
775 }
776
777 /*
778 * Allocate a thread.
779 */
780 struct thread *
thread_alloc(int pages)781 thread_alloc(int pages)
782 {
783 struct thread *td;
784 lwpid_t tid;
785
786 if (!thread_count_inc()) {
787 return (NULL);
788 }
789
790 tid = tid_alloc();
791 td = uma_zalloc(thread_zone, M_WAITOK);
792 KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
793 if (!vm_thread_new(td, pages)) {
794 uma_zfree(thread_zone, td);
795 tid_free(tid);
796 thread_count_dec();
797 return (NULL);
798 }
799 td->td_tid = tid;
800 bzero(&td->td_sa.args, sizeof(td->td_sa.args));
801 kasan_thread_alloc(td);
802 kmsan_thread_alloc(td);
803 cpu_thread_alloc(td);
804 EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
805 return (td);
806 }
807
808 int
thread_recycle(struct thread * td,int pages)809 thread_recycle(struct thread *td, int pages)
810 {
811 if (td->td_kstack == 0 || td->td_kstack_pages != pages) {
812 if (td->td_kstack != 0)
813 vm_thread_dispose(td);
814 if (!vm_thread_new(td, pages))
815 return (ENOMEM);
816 cpu_thread_alloc(td);
817 }
818 kasan_thread_alloc(td);
819 kmsan_thread_alloc(td);
820 return (0);
821 }
822
823 /*
824 * Deallocate a thread.
825 */
826 static void
thread_free_batched(struct thread * td)827 thread_free_batched(struct thread *td)
828 {
829
830 lock_profile_thread_exit(td);
831 if (td->td_cpuset)
832 cpuset_rel(td->td_cpuset);
833 td->td_cpuset = NULL;
834 cpu_thread_free(td);
835 if (td->td_kstack != 0)
836 vm_thread_dispose(td);
837 callout_drain(&td->td_slpcallout);
838 /*
839 * Freeing handled by the caller.
840 */
841 td->td_tid = -1;
842 kmsan_thread_free(td);
843 uma_zfree(thread_zone, td);
844 }
845
846 void
thread_free(struct thread * td)847 thread_free(struct thread *td)
848 {
849 lwpid_t tid;
850
851 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
852 tid = td->td_tid;
853 thread_free_batched(td);
854 tid_free(tid);
855 thread_count_dec();
856 }
857
858 void
thread_cow_get_proc(struct thread * newtd,struct proc * p)859 thread_cow_get_proc(struct thread *newtd, struct proc *p)
860 {
861
862 PROC_LOCK_ASSERT(p, MA_OWNED);
863 newtd->td_realucred = crcowget(p->p_ucred);
864 newtd->td_ucred = newtd->td_realucred;
865 newtd->td_limit = lim_hold(p->p_limit);
866 newtd->td_cowgen = p->p_cowgen;
867 }
868
869 void
thread_cow_get(struct thread * newtd,struct thread * td)870 thread_cow_get(struct thread *newtd, struct thread *td)
871 {
872
873 MPASS(td->td_realucred == td->td_ucred);
874 newtd->td_realucred = crcowget(td->td_realucred);
875 newtd->td_ucred = newtd->td_realucred;
876 newtd->td_limit = lim_hold(td->td_limit);
877 newtd->td_cowgen = td->td_cowgen;
878 }
879
880 void
thread_cow_free(struct thread * td)881 thread_cow_free(struct thread *td)
882 {
883
884 if (td->td_realucred != NULL)
885 crcowfree(td);
886 if (td->td_limit != NULL)
887 lim_free(td->td_limit);
888 }
889
890 void
thread_cow_update(struct thread * td)891 thread_cow_update(struct thread *td)
892 {
893 struct proc *p;
894 struct ucred *oldcred;
895 struct plimit *oldlimit;
896
897 p = td->td_proc;
898 PROC_LOCK(p);
899 oldcred = crcowsync();
900 oldlimit = lim_cowsync();
901 td->td_cowgen = p->p_cowgen;
902 PROC_UNLOCK(p);
903 if (oldcred != NULL)
904 crfree(oldcred);
905 if (oldlimit != NULL)
906 lim_free(oldlimit);
907 }
908
909 void
thread_cow_synced(struct thread * td)910 thread_cow_synced(struct thread *td)
911 {
912 struct proc *p;
913
914 p = td->td_proc;
915 PROC_LOCK_ASSERT(p, MA_OWNED);
916 MPASS(td->td_cowgen != p->p_cowgen);
917 MPASS(td->td_ucred == p->p_ucred);
918 MPASS(td->td_limit == p->p_limit);
919 td->td_cowgen = p->p_cowgen;
920 }
921
922 /*
923 * Discard the current thread and exit from its context.
924 * Always called with scheduler locked.
925 *
926 * Because we can't free a thread while we're operating under its context,
927 * push the current thread into our CPU's deadthread holder. This means
928 * we needn't worry about someone else grabbing our context before we
929 * do a cpu_throw().
930 */
931 void
thread_exit(void)932 thread_exit(void)
933 {
934 uint64_t runtime, new_switchtime;
935 struct thread *td;
936 struct thread *td2;
937 struct proc *p;
938 int wakeup_swapper;
939
940 td = curthread;
941 p = td->td_proc;
942
943 PROC_SLOCK_ASSERT(p, MA_OWNED);
944 mtx_assert(&Giant, MA_NOTOWNED);
945
946 PROC_LOCK_ASSERT(p, MA_OWNED);
947 KASSERT(p != NULL, ("thread exiting without a process"));
948 CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
949 (long)p->p_pid, td->td_name);
950 SDT_PROBE0(proc, , , lwp__exit);
951 KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
952 MPASS(td->td_realucred == td->td_ucred);
953
954 /*
955 * drop FPU & debug register state storage, or any other
956 * architecture specific resources that
957 * would not be on a new untouched process.
958 */
959 cpu_thread_exit(td);
960
961 /*
962 * The last thread is left attached to the process
963 * So that the whole bundle gets recycled. Skip
964 * all this stuff if we never had threads.
965 * EXIT clears all sign of other threads when
966 * it goes to single threading, so the last thread always
967 * takes the short path.
968 */
969 if (p->p_flag & P_HADTHREADS) {
970 if (p->p_numthreads > 1) {
971 atomic_add_int(&td->td_proc->p_exitthreads, 1);
972 thread_unlink(td);
973 td2 = FIRST_THREAD_IN_PROC(p);
974 sched_exit_thread(td2, td);
975
976 /*
977 * The test below is NOT true if we are the
978 * sole exiting thread. P_STOPPED_SINGLE is unset
979 * in exit1() after it is the only survivor.
980 */
981 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
982 if (p->p_numthreads == p->p_suspcount) {
983 thread_lock(p->p_singlethread);
984 wakeup_swapper = thread_unsuspend_one(
985 p->p_singlethread, p, false);
986 if (wakeup_swapper)
987 kick_proc0();
988 }
989 }
990
991 PCPU_SET(deadthread, td);
992 } else {
993 /*
994 * The last thread is exiting.. but not through exit()
995 */
996 panic ("thread_exit: Last thread exiting on its own");
997 }
998 }
999 #ifdef HWPMC_HOOKS
1000 /*
1001 * If this thread is part of a process that is being tracked by hwpmc(4),
1002 * inform the module of the thread's impending exit.
1003 */
1004 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
1005 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1006 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
1007 } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
1008 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
1009 #endif
1010 PROC_UNLOCK(p);
1011 PROC_STATLOCK(p);
1012 thread_lock(td);
1013 PROC_SUNLOCK(p);
1014
1015 /* Do the same timestamp bookkeeping that mi_switch() would do. */
1016 new_switchtime = cpu_ticks();
1017 runtime = new_switchtime - PCPU_GET(switchtime);
1018 td->td_runtime += runtime;
1019 td->td_incruntime += runtime;
1020 PCPU_SET(switchtime, new_switchtime);
1021 PCPU_SET(switchticks, ticks);
1022 VM_CNT_INC(v_swtch);
1023
1024 /* Save our resource usage in our process. */
1025 td->td_ru.ru_nvcsw++;
1026 ruxagg_locked(p, td);
1027 rucollect(&p->p_ru, &td->td_ru);
1028 PROC_STATUNLOCK(p);
1029
1030 TD_SET_STATE(td, TDS_INACTIVE);
1031 #ifdef WITNESS
1032 witness_thread_exit(td);
1033 #endif
1034 CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
1035 sched_throw(td);
1036 panic("I'm a teapot!");
1037 /* NOTREACHED */
1038 }
1039
1040 /*
1041 * Do any thread specific cleanups that may be needed in wait()
1042 * called with Giant, proc and schedlock not held.
1043 */
1044 void
thread_wait(struct proc * p)1045 thread_wait(struct proc *p)
1046 {
1047 struct thread *td;
1048
1049 mtx_assert(&Giant, MA_NOTOWNED);
1050 KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
1051 KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
1052 td = FIRST_THREAD_IN_PROC(p);
1053 /* Lock the last thread so we spin until it exits cpu_throw(). */
1054 thread_lock(td);
1055 thread_unlock(td);
1056 lock_profile_thread_exit(td);
1057 cpuset_rel(td->td_cpuset);
1058 td->td_cpuset = NULL;
1059 cpu_thread_clean(td);
1060 thread_cow_free(td);
1061 callout_drain(&td->td_slpcallout);
1062 thread_reap(); /* check for zombie threads etc. */
1063 }
1064
1065 /*
1066 * Link a thread to a process.
1067 * set up anything that needs to be initialized for it to
1068 * be used by the process.
1069 */
1070 void
thread_link(struct thread * td,struct proc * p)1071 thread_link(struct thread *td, struct proc *p)
1072 {
1073
1074 /*
1075 * XXX This can't be enabled because it's called for proc0 before
1076 * its lock has been created.
1077 * PROC_LOCK_ASSERT(p, MA_OWNED);
1078 */
1079 TD_SET_STATE(td, TDS_INACTIVE);
1080 td->td_proc = p;
1081 td->td_flags = TDF_INMEM;
1082
1083 LIST_INIT(&td->td_contested);
1084 LIST_INIT(&td->td_lprof[0]);
1085 LIST_INIT(&td->td_lprof[1]);
1086 #ifdef EPOCH_TRACE
1087 SLIST_INIT(&td->td_epochs);
1088 #endif
1089 sigqueue_init(&td->td_sigqueue, p);
1090 callout_init(&td->td_slpcallout, 1);
1091 TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
1092 p->p_numthreads++;
1093 }
1094
1095 /*
1096 * Called from:
1097 * thread_exit()
1098 */
1099 void
thread_unlink(struct thread * td)1100 thread_unlink(struct thread *td)
1101 {
1102 struct proc *p = td->td_proc;
1103
1104 PROC_LOCK_ASSERT(p, MA_OWNED);
1105 #ifdef EPOCH_TRACE
1106 MPASS(SLIST_EMPTY(&td->td_epochs));
1107 #endif
1108
1109 TAILQ_REMOVE(&p->p_threads, td, td_plist);
1110 p->p_numthreads--;
1111 /* could clear a few other things here */
1112 /* Must NOT clear links to proc! */
1113 }
1114
1115 static int
calc_remaining(struct proc * p,int mode)1116 calc_remaining(struct proc *p, int mode)
1117 {
1118 int remaining;
1119
1120 PROC_LOCK_ASSERT(p, MA_OWNED);
1121 PROC_SLOCK_ASSERT(p, MA_OWNED);
1122 if (mode == SINGLE_EXIT)
1123 remaining = p->p_numthreads;
1124 else if (mode == SINGLE_BOUNDARY)
1125 remaining = p->p_numthreads - p->p_boundary_count;
1126 else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
1127 remaining = p->p_numthreads - p->p_suspcount;
1128 else
1129 panic("calc_remaining: wrong mode %d", mode);
1130 return (remaining);
1131 }
1132
1133 static int
remain_for_mode(int mode)1134 remain_for_mode(int mode)
1135 {
1136
1137 return (mode == SINGLE_ALLPROC ? 0 : 1);
1138 }
1139
1140 static int
weed_inhib(int mode,struct thread * td2,struct proc * p)1141 weed_inhib(int mode, struct thread *td2, struct proc *p)
1142 {
1143 int wakeup_swapper;
1144
1145 PROC_LOCK_ASSERT(p, MA_OWNED);
1146 PROC_SLOCK_ASSERT(p, MA_OWNED);
1147 THREAD_LOCK_ASSERT(td2, MA_OWNED);
1148
1149 wakeup_swapper = 0;
1150
1151 /*
1152 * Since the thread lock is dropped by the scheduler we have
1153 * to retry to check for races.
1154 */
1155 restart:
1156 switch (mode) {
1157 case SINGLE_EXIT:
1158 if (TD_IS_SUSPENDED(td2)) {
1159 wakeup_swapper |= thread_unsuspend_one(td2, p, true);
1160 thread_lock(td2);
1161 goto restart;
1162 }
1163 if (TD_CAN_ABORT(td2)) {
1164 wakeup_swapper |= sleepq_abort(td2, EINTR);
1165 return (wakeup_swapper);
1166 }
1167 break;
1168 case SINGLE_BOUNDARY:
1169 case SINGLE_NO_EXIT:
1170 if (TD_IS_SUSPENDED(td2) &&
1171 (td2->td_flags & TDF_BOUNDARY) == 0) {
1172 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1173 thread_lock(td2);
1174 goto restart;
1175 }
1176 if (TD_CAN_ABORT(td2)) {
1177 wakeup_swapper |= sleepq_abort(td2, ERESTART);
1178 return (wakeup_swapper);
1179 }
1180 break;
1181 case SINGLE_ALLPROC:
1182 /*
1183 * ALLPROC suspend tries to avoid spurious EINTR for
1184 * threads sleeping interruptable, by suspending the
1185 * thread directly, similarly to sig_suspend_threads().
1186 * Since such sleep is not neccessary performed at the user
1187 * boundary, TDF_ALLPROCSUSP is used to avoid immediate
1188 * un-suspend.
1189 */
1190 if (TD_IS_SUSPENDED(td2) &&
1191 (td2->td_flags & TDF_ALLPROCSUSP) == 0) {
1192 wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1193 thread_lock(td2);
1194 goto restart;
1195 }
1196 if (TD_CAN_ABORT(td2)) {
1197 td2->td_flags |= TDF_ALLPROCSUSP;
1198 wakeup_swapper |= sleepq_abort(td2, ERESTART);
1199 return (wakeup_swapper);
1200 }
1201 break;
1202 default:
1203 break;
1204 }
1205 thread_unlock(td2);
1206 return (wakeup_swapper);
1207 }
1208
1209 /*
1210 * Enforce single-threading.
1211 *
1212 * Returns 1 if the caller must abort (another thread is waiting to
1213 * exit the process or similar). Process is locked!
1214 * Returns 0 when you are successfully the only thread running.
1215 * A process has successfully single threaded in the suspend mode when
1216 * There are no threads in user mode. Threads in the kernel must be
1217 * allowed to continue until they get to the user boundary. They may even
1218 * copy out their return values and data before suspending. They may however be
1219 * accelerated in reaching the user boundary as we will wake up
1220 * any sleeping threads that are interruptable. (PCATCH).
1221 */
1222 int
thread_single(struct proc * p,int mode)1223 thread_single(struct proc *p, int mode)
1224 {
1225 struct thread *td;
1226 struct thread *td2;
1227 int remaining, wakeup_swapper;
1228
1229 td = curthread;
1230 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1231 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1232 ("invalid mode %d", mode));
1233 /*
1234 * If allowing non-ALLPROC singlethreading for non-curproc
1235 * callers, calc_remaining() and remain_for_mode() should be
1236 * adjusted to also account for td->td_proc != p. For now
1237 * this is not implemented because it is not used.
1238 */
1239 KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
1240 (mode != SINGLE_ALLPROC && td->td_proc == p),
1241 ("mode %d proc %p curproc %p", mode, p, td->td_proc));
1242 mtx_assert(&Giant, MA_NOTOWNED);
1243 PROC_LOCK_ASSERT(p, MA_OWNED);
1244
1245 /*
1246 * Is someone already single threading?
1247 * Or may be singlethreading is not needed at all.
1248 */
1249 if (mode == SINGLE_ALLPROC) {
1250 while ((p->p_flag & P_STOPPED_SINGLE) != 0) {
1251 if ((p->p_flag2 & P2_WEXIT) != 0)
1252 return (1);
1253 msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0);
1254 }
1255 if ((p->p_flag & (P_STOPPED_SIG | P_TRACED)) != 0 ||
1256 (p->p_flag2 & P2_WEXIT) != 0)
1257 return (1);
1258 } else if ((p->p_flag & P_HADTHREADS) == 0)
1259 return (0);
1260 if (p->p_singlethread != NULL && p->p_singlethread != td)
1261 return (1);
1262
1263 if (mode == SINGLE_EXIT) {
1264 p->p_flag |= P_SINGLE_EXIT;
1265 p->p_flag &= ~P_SINGLE_BOUNDARY;
1266 } else {
1267 p->p_flag &= ~P_SINGLE_EXIT;
1268 if (mode == SINGLE_BOUNDARY)
1269 p->p_flag |= P_SINGLE_BOUNDARY;
1270 else
1271 p->p_flag &= ~P_SINGLE_BOUNDARY;
1272 }
1273 if (mode == SINGLE_ALLPROC)
1274 p->p_flag |= P_TOTAL_STOP;
1275 p->p_flag |= P_STOPPED_SINGLE;
1276 PROC_SLOCK(p);
1277 p->p_singlethread = td;
1278 remaining = calc_remaining(p, mode);
1279 while (remaining != remain_for_mode(mode)) {
1280 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
1281 goto stopme;
1282 wakeup_swapper = 0;
1283 FOREACH_THREAD_IN_PROC(p, td2) {
1284 if (td2 == td)
1285 continue;
1286 thread_lock(td2);
1287 ast_sched_locked(td2, TDA_SUSPEND);
1288 if (TD_IS_INHIBITED(td2)) {
1289 wakeup_swapper |= weed_inhib(mode, td2, p);
1290 #ifdef SMP
1291 } else if (TD_IS_RUNNING(td2)) {
1292 forward_signal(td2);
1293 thread_unlock(td2);
1294 #endif
1295 } else
1296 thread_unlock(td2);
1297 }
1298 if (wakeup_swapper)
1299 kick_proc0();
1300 remaining = calc_remaining(p, mode);
1301
1302 /*
1303 * Maybe we suspended some threads.. was it enough?
1304 */
1305 if (remaining == remain_for_mode(mode))
1306 break;
1307
1308 stopme:
1309 /*
1310 * Wake us up when everyone else has suspended.
1311 * In the mean time we suspend as well.
1312 */
1313 thread_suspend_switch(td, p);
1314 remaining = calc_remaining(p, mode);
1315 }
1316 if (mode == SINGLE_EXIT) {
1317 /*
1318 * Convert the process to an unthreaded process. The
1319 * SINGLE_EXIT is called by exit1() or execve(), in
1320 * both cases other threads must be retired.
1321 */
1322 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1323 p->p_singlethread = NULL;
1324 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1325
1326 /*
1327 * Wait for any remaining threads to exit cpu_throw().
1328 */
1329 while (p->p_exitthreads != 0) {
1330 PROC_SUNLOCK(p);
1331 PROC_UNLOCK(p);
1332 sched_relinquish(td);
1333 PROC_LOCK(p);
1334 PROC_SLOCK(p);
1335 }
1336 } else if (mode == SINGLE_BOUNDARY) {
1337 /*
1338 * Wait until all suspended threads are removed from
1339 * the processors. The thread_suspend_check()
1340 * increments p_boundary_count while it is still
1341 * running, which makes it possible for the execve()
1342 * to destroy vmspace while our other threads are
1343 * still using the address space.
1344 *
1345 * We lock the thread, which is only allowed to
1346 * succeed after context switch code finished using
1347 * the address space.
1348 */
1349 FOREACH_THREAD_IN_PROC(p, td2) {
1350 if (td2 == td)
1351 continue;
1352 thread_lock(td2);
1353 KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1354 ("td %p not on boundary", td2));
1355 KASSERT(TD_IS_SUSPENDED(td2),
1356 ("td %p is not suspended", td2));
1357 thread_unlock(td2);
1358 }
1359 }
1360 PROC_SUNLOCK(p);
1361 return (0);
1362 }
1363
1364 bool
thread_suspend_check_needed(void)1365 thread_suspend_check_needed(void)
1366 {
1367 struct proc *p;
1368 struct thread *td;
1369
1370 td = curthread;
1371 p = td->td_proc;
1372 PROC_LOCK_ASSERT(p, MA_OWNED);
1373 return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1374 (td->td_dbgflags & TDB_SUSPEND) != 0));
1375 }
1376
1377 /*
1378 * Called in from locations that can safely check to see
1379 * whether we have to suspend or at least throttle for a
1380 * single-thread event (e.g. fork).
1381 *
1382 * Such locations include userret().
1383 * If the "return_instead" argument is non zero, the thread must be able to
1384 * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1385 *
1386 * The 'return_instead' argument tells the function if it may do a
1387 * thread_exit() or suspend, or whether the caller must abort and back
1388 * out instead.
1389 *
1390 * If the thread that set the single_threading request has set the
1391 * P_SINGLE_EXIT bit in the process flags then this call will never return
1392 * if 'return_instead' is false, but will exit.
1393 *
1394 * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1395 *---------------+--------------------+---------------------
1396 * 0 | returns 0 | returns 0 or 1
1397 * | when ST ends | immediately
1398 *---------------+--------------------+---------------------
1399 * 1 | thread exits | returns 1
1400 * | | immediately
1401 * 0 = thread_exit() or suspension ok,
1402 * other = return error instead of stopping the thread.
1403 *
1404 * While a full suspension is under effect, even a single threading
1405 * thread would be suspended if it made this call (but it shouldn't).
1406 * This call should only be made from places where
1407 * thread_exit() would be safe as that may be the outcome unless
1408 * return_instead is set.
1409 */
1410 int
thread_suspend_check(int return_instead)1411 thread_suspend_check(int return_instead)
1412 {
1413 struct thread *td;
1414 struct proc *p;
1415 int wakeup_swapper;
1416
1417 td = curthread;
1418 p = td->td_proc;
1419 mtx_assert(&Giant, MA_NOTOWNED);
1420 PROC_LOCK_ASSERT(p, MA_OWNED);
1421 while (thread_suspend_check_needed()) {
1422 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1423 KASSERT(p->p_singlethread != NULL,
1424 ("singlethread not set"));
1425 /*
1426 * The only suspension in action is a
1427 * single-threading. Single threader need not stop.
1428 * It is safe to access p->p_singlethread unlocked
1429 * because it can only be set to our address by us.
1430 */
1431 if (p->p_singlethread == td)
1432 return (0); /* Exempt from stopping. */
1433 }
1434 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1435 return (EINTR);
1436
1437 /* Should we goto user boundary if we didn't come from there? */
1438 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1439 (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1440 return (ERESTART);
1441
1442 /*
1443 * Ignore suspend requests if they are deferred.
1444 */
1445 if ((td->td_flags & TDF_SBDRY) != 0) {
1446 KASSERT(return_instead,
1447 ("TDF_SBDRY set for unsafe thread_suspend_check"));
1448 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1449 (TDF_SEINTR | TDF_SERESTART),
1450 ("both TDF_SEINTR and TDF_SERESTART"));
1451 return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1452 }
1453
1454 /*
1455 * If the process is waiting for us to exit,
1456 * this thread should just suicide.
1457 * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1458 */
1459 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1460 PROC_UNLOCK(p);
1461
1462 /*
1463 * Allow Linux emulation layer to do some work
1464 * before thread suicide.
1465 */
1466 if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1467 (p->p_sysent->sv_thread_detach)(td);
1468 umtx_thread_exit(td);
1469 kern_thr_exit(td);
1470 panic("stopped thread did not exit");
1471 }
1472
1473 PROC_SLOCK(p);
1474 thread_stopped(p);
1475 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1476 if (p->p_numthreads == p->p_suspcount + 1) {
1477 thread_lock(p->p_singlethread);
1478 wakeup_swapper = thread_unsuspend_one(
1479 p->p_singlethread, p, false);
1480 if (wakeup_swapper)
1481 kick_proc0();
1482 }
1483 }
1484 PROC_UNLOCK(p);
1485 thread_lock(td);
1486 /*
1487 * When a thread suspends, it just
1488 * gets taken off all queues.
1489 */
1490 thread_suspend_one(td);
1491 if (return_instead == 0) {
1492 p->p_boundary_count++;
1493 td->td_flags |= TDF_BOUNDARY;
1494 }
1495 PROC_SUNLOCK(p);
1496 mi_switch(SW_INVOL | SWT_SUSPEND);
1497 PROC_LOCK(p);
1498 }
1499 return (0);
1500 }
1501
1502 /*
1503 * Check for possible stops and suspensions while executing a
1504 * casueword or similar transiently failing operation.
1505 *
1506 * The sleep argument controls whether the function can handle a stop
1507 * request itself or it should return ERESTART and the request is
1508 * proceed at the kernel/user boundary in ast.
1509 *
1510 * Typically, when retrying due to casueword(9) failure (rv == 1), we
1511 * should handle the stop requests there, with exception of cases when
1512 * the thread owns a kernel resource, for instance busied the umtx
1513 * key, or when functions return immediately if thread_check_susp()
1514 * returned non-zero. On the other hand, retrying the whole lock
1515 * operation, we better not stop there but delegate the handling to
1516 * ast.
1517 *
1518 * If the request is for thread termination P_SINGLE_EXIT, we cannot
1519 * handle it at all, and simply return EINTR.
1520 */
1521 int
thread_check_susp(struct thread * td,bool sleep)1522 thread_check_susp(struct thread *td, bool sleep)
1523 {
1524 struct proc *p;
1525 int error;
1526
1527 /*
1528 * The check for TDA_SUSPEND is racy, but it is enough to
1529 * eventually break the lockstep loop.
1530 */
1531 if (!td_ast_pending(td, TDA_SUSPEND))
1532 return (0);
1533 error = 0;
1534 p = td->td_proc;
1535 PROC_LOCK(p);
1536 if (p->p_flag & P_SINGLE_EXIT)
1537 error = EINTR;
1538 else if (P_SHOULDSTOP(p) ||
1539 ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1540 error = sleep ? thread_suspend_check(0) : ERESTART;
1541 PROC_UNLOCK(p);
1542 return (error);
1543 }
1544
1545 void
thread_suspend_switch(struct thread * td,struct proc * p)1546 thread_suspend_switch(struct thread *td, struct proc *p)
1547 {
1548
1549 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1550 PROC_LOCK_ASSERT(p, MA_OWNED);
1551 PROC_SLOCK_ASSERT(p, MA_OWNED);
1552 /*
1553 * We implement thread_suspend_one in stages here to avoid
1554 * dropping the proc lock while the thread lock is owned.
1555 */
1556 if (p == td->td_proc) {
1557 thread_stopped(p);
1558 p->p_suspcount++;
1559 }
1560 PROC_UNLOCK(p);
1561 thread_lock(td);
1562 ast_unsched_locked(td, TDA_SUSPEND);
1563 TD_SET_SUSPENDED(td);
1564 sched_sleep(td, 0);
1565 PROC_SUNLOCK(p);
1566 DROP_GIANT();
1567 mi_switch(SW_VOL | SWT_SUSPEND);
1568 PICKUP_GIANT();
1569 PROC_LOCK(p);
1570 PROC_SLOCK(p);
1571 }
1572
1573 void
thread_suspend_one(struct thread * td)1574 thread_suspend_one(struct thread *td)
1575 {
1576 struct proc *p;
1577
1578 p = td->td_proc;
1579 PROC_SLOCK_ASSERT(p, MA_OWNED);
1580 THREAD_LOCK_ASSERT(td, MA_OWNED);
1581 KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1582 p->p_suspcount++;
1583 ast_unsched_locked(td, TDA_SUSPEND);
1584 TD_SET_SUSPENDED(td);
1585 sched_sleep(td, 0);
1586 }
1587
1588 static int
thread_unsuspend_one(struct thread * td,struct proc * p,bool boundary)1589 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1590 {
1591
1592 THREAD_LOCK_ASSERT(td, MA_OWNED);
1593 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1594 TD_CLR_SUSPENDED(td);
1595 td->td_flags &= ~TDF_ALLPROCSUSP;
1596 if (td->td_proc == p) {
1597 PROC_SLOCK_ASSERT(p, MA_OWNED);
1598 p->p_suspcount--;
1599 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1600 td->td_flags &= ~TDF_BOUNDARY;
1601 p->p_boundary_count--;
1602 }
1603 }
1604 return (setrunnable(td, 0));
1605 }
1606
1607 void
thread_run_flash(struct thread * td)1608 thread_run_flash(struct thread *td)
1609 {
1610 struct proc *p;
1611
1612 p = td->td_proc;
1613 PROC_LOCK_ASSERT(p, MA_OWNED);
1614
1615 if (TD_ON_SLEEPQ(td))
1616 sleepq_remove_nested(td);
1617 else
1618 thread_lock(td);
1619
1620 THREAD_LOCK_ASSERT(td, MA_OWNED);
1621 KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1622
1623 TD_CLR_SUSPENDED(td);
1624 PROC_SLOCK(p);
1625 MPASS(p->p_suspcount > 0);
1626 p->p_suspcount--;
1627 PROC_SUNLOCK(p);
1628 if (setrunnable(td, 0))
1629 kick_proc0();
1630 }
1631
1632 /*
1633 * Allow all threads blocked by single threading to continue running.
1634 */
1635 void
thread_unsuspend(struct proc * p)1636 thread_unsuspend(struct proc *p)
1637 {
1638 struct thread *td;
1639 int wakeup_swapper;
1640
1641 PROC_LOCK_ASSERT(p, MA_OWNED);
1642 PROC_SLOCK_ASSERT(p, MA_OWNED);
1643 wakeup_swapper = 0;
1644 if (!P_SHOULDSTOP(p)) {
1645 FOREACH_THREAD_IN_PROC(p, td) {
1646 thread_lock(td);
1647 if (TD_IS_SUSPENDED(td))
1648 wakeup_swapper |= thread_unsuspend_one(td, p,
1649 true);
1650 else
1651 thread_unlock(td);
1652 }
1653 } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1654 p->p_numthreads == p->p_suspcount) {
1655 /*
1656 * Stopping everything also did the job for the single
1657 * threading request. Now we've downgraded to single-threaded,
1658 * let it continue.
1659 */
1660 if (p->p_singlethread->td_proc == p) {
1661 thread_lock(p->p_singlethread);
1662 wakeup_swapper = thread_unsuspend_one(
1663 p->p_singlethread, p, false);
1664 }
1665 }
1666 if (wakeup_swapper)
1667 kick_proc0();
1668 }
1669
1670 /*
1671 * End the single threading mode..
1672 */
1673 void
thread_single_end(struct proc * p,int mode)1674 thread_single_end(struct proc *p, int mode)
1675 {
1676 struct thread *td;
1677 int wakeup_swapper;
1678
1679 KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1680 mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1681 ("invalid mode %d", mode));
1682 PROC_LOCK_ASSERT(p, MA_OWNED);
1683 KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1684 (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1685 ("mode %d does not match P_TOTAL_STOP", mode));
1686 KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1687 ("thread_single_end from other thread %p %p",
1688 curthread, p->p_singlethread));
1689 KASSERT(mode != SINGLE_BOUNDARY ||
1690 (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1691 ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1692 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1693 P_TOTAL_STOP);
1694 PROC_SLOCK(p);
1695 p->p_singlethread = NULL;
1696 wakeup_swapper = 0;
1697 /*
1698 * If there are other threads they may now run,
1699 * unless of course there is a blanket 'stop order'
1700 * on the process. The single threader must be allowed
1701 * to continue however as this is a bad place to stop.
1702 */
1703 if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1704 FOREACH_THREAD_IN_PROC(p, td) {
1705 thread_lock(td);
1706 if (TD_IS_SUSPENDED(td)) {
1707 wakeup_swapper |= thread_unsuspend_one(td, p,
1708 true);
1709 } else
1710 thread_unlock(td);
1711 }
1712 }
1713 KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1714 ("inconsistent boundary count %d", p->p_boundary_count));
1715 PROC_SUNLOCK(p);
1716 if (wakeup_swapper)
1717 kick_proc0();
1718 wakeup(&p->p_flag);
1719 }
1720
1721 /*
1722 * Locate a thread by number and return with proc lock held.
1723 *
1724 * thread exit establishes proc -> tidhash lock ordering, but lookup
1725 * takes tidhash first and needs to return locked proc.
1726 *
1727 * The problem is worked around by relying on type-safety of both
1728 * structures and doing the work in 2 steps:
1729 * - tidhash-locked lookup which saves both thread and proc pointers
1730 * - proc-locked verification that the found thread still matches
1731 */
1732 static bool
tdfind_hash(lwpid_t tid,pid_t pid,struct proc ** pp,struct thread ** tdp)1733 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1734 {
1735 #define RUN_THRESH 16
1736 struct proc *p;
1737 struct thread *td;
1738 int run;
1739 bool locked;
1740
1741 run = 0;
1742 rw_rlock(TIDHASHLOCK(tid));
1743 locked = true;
1744 LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1745 if (td->td_tid != tid) {
1746 run++;
1747 continue;
1748 }
1749 p = td->td_proc;
1750 if (pid != -1 && p->p_pid != pid) {
1751 td = NULL;
1752 break;
1753 }
1754 if (run > RUN_THRESH) {
1755 if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1756 LIST_REMOVE(td, td_hash);
1757 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1758 td, td_hash);
1759 rw_wunlock(TIDHASHLOCK(tid));
1760 locked = false;
1761 break;
1762 }
1763 }
1764 break;
1765 }
1766 if (locked)
1767 rw_runlock(TIDHASHLOCK(tid));
1768 if (td == NULL)
1769 return (false);
1770 *pp = p;
1771 *tdp = td;
1772 return (true);
1773 }
1774
1775 struct thread *
tdfind(lwpid_t tid,pid_t pid)1776 tdfind(lwpid_t tid, pid_t pid)
1777 {
1778 struct proc *p;
1779 struct thread *td;
1780
1781 td = curthread;
1782 if (td->td_tid == tid) {
1783 if (pid != -1 && td->td_proc->p_pid != pid)
1784 return (NULL);
1785 PROC_LOCK(td->td_proc);
1786 return (td);
1787 }
1788
1789 for (;;) {
1790 if (!tdfind_hash(tid, pid, &p, &td))
1791 return (NULL);
1792 PROC_LOCK(p);
1793 if (td->td_tid != tid) {
1794 PROC_UNLOCK(p);
1795 continue;
1796 }
1797 if (td->td_proc != p) {
1798 PROC_UNLOCK(p);
1799 continue;
1800 }
1801 if (p->p_state == PRS_NEW) {
1802 PROC_UNLOCK(p);
1803 return (NULL);
1804 }
1805 return (td);
1806 }
1807 }
1808
1809 void
tidhash_add(struct thread * td)1810 tidhash_add(struct thread *td)
1811 {
1812 rw_wlock(TIDHASHLOCK(td->td_tid));
1813 LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1814 rw_wunlock(TIDHASHLOCK(td->td_tid));
1815 }
1816
1817 void
tidhash_remove(struct thread * td)1818 tidhash_remove(struct thread *td)
1819 {
1820
1821 rw_wlock(TIDHASHLOCK(td->td_tid));
1822 LIST_REMOVE(td, td_hash);
1823 rw_wunlock(TIDHASHLOCK(td->td_tid));
1824 }
1825