sys/kern/kern_thread.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   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/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/kernel.h>
  40 #include <sys/lock.h>
  41 #include <sys/mutex.h>
  42 #include <sys/proc.h>
  43 #include <sys/bitstring.h>
  44 #include <sys/epoch.h>
  45 #include <sys/rangelock.h>
  46 #include <sys/resourcevar.h>
  47 #include <sys/sdt.h>
  48 #include <sys/smp.h>
  49 #include <sys/sched.h>
  50 #include <sys/sleepqueue.h>
  51 #include <sys/selinfo.h>
  52 #include <sys/syscallsubr.h>
  53 #include <sys/dtrace_bsd.h>
  54 #include <sys/sysent.h>
  55 #include <sys/turnstile.h>
  56 #include <sys/taskqueue.h>
  57 #include <sys/ktr.h>
  58 #include <sys/rwlock.h>
  59 #include <sys/umtxvar.h>
  60 #include <sys/vmmeter.h>
  61 #include <sys/cpuset.h>
  62 #ifdef  HWPMC_HOOKS
  63 #include <sys/pmckern.h>
  64 #endif
  65 #include <sys/priv.h>
  66
  67 #include <security/audit/audit.h>
  68
  69 #include <vm/pmap.h>
  70 #include <vm/vm.h>
  71 #include <vm/vm_extern.h>
  72 #include <vm/uma.h>
  73 #include <vm/vm_phys.h>
  74 #include <sys/eventhandler.h>
  75
  76 /*
  77  * Asserts below verify the stability of struct thread and struct proc
  78  * layout, as exposed by KBI to modules.  On head, the KBI is allowed
  79  * to drift, change to the structures must be accompanied by the
  80  * assert update.
  81  *
  82  * On the stable branches after KBI freeze, conditions must not be
  83  * violated.  Typically new fields are moved to the end of the
  84  * structures.
  85  */
  86 #ifdef __amd64__
  87 _Static_assert(offsetof(struct thread, td_flags) == 0xfc,
  88     "struct thread KBI td_flags");
  89 _Static_assert(offsetof(struct thread, td_pflags) == 0x104,
  90     "struct thread KBI td_pflags");
  91 _Static_assert(offsetof(struct thread, td_frame) == 0x4a0,
  92     "struct thread KBI td_frame");
  93 _Static_assert(offsetof(struct thread, td_emuldata) == 0x6b0,
  94     "struct thread KBI td_emuldata");
  95 _Static_assert(offsetof(struct proc, p_flag) == 0xb8,
  96     "struct proc KBI p_flag");
  97 _Static_assert(offsetof(struct proc, p_pid) == 0xc4,
  98     "struct proc KBI p_pid");
  99 _Static_assert(offsetof(struct proc, p_filemon) == 0x3c0,
 100     "struct proc KBI p_filemon");
 101 _Static_assert(offsetof(struct proc, p_comm) == 0x3d8,
 102     "struct proc KBI p_comm");
 103 _Static_assert(offsetof(struct proc, p_emuldata) == 0x4b8,
 104     "struct proc KBI p_emuldata");
 105 #endif
 106 #ifdef __i386__
 107 _Static_assert(offsetof(struct thread, td_flags) == 0x98,
 108     "struct thread KBI td_flags");
 109 _Static_assert(offsetof(struct thread, td_pflags) == 0xa0,
 110     "struct thread KBI td_pflags");
 111 _Static_assert(offsetof(struct thread, td_frame) == 0x300,
 112     "struct thread KBI td_frame");
 113 _Static_assert(offsetof(struct thread, td_emuldata) == 0x344,
 114     "struct thread KBI td_emuldata");
 115 _Static_assert(offsetof(struct proc, p_flag) == 0x6c,
 116     "struct proc KBI p_flag");
 117 _Static_assert(offsetof(struct proc, p_pid) == 0x78,
 118     "struct proc KBI p_pid");
 119 _Static_assert(offsetof(struct proc, p_filemon) == 0x26c,
 120     "struct proc KBI p_filemon");
 121 _Static_assert(offsetof(struct proc, p_comm) == 0x280,
 122     "struct proc KBI p_comm");
 123 _Static_assert(offsetof(struct proc, p_emuldata) == 0x30c,
 124     "struct proc KBI p_emuldata");
 125 #endif
 126
 127 SDT_PROVIDER_DECLARE(proc);
 128 SDT_PROBE_DEFINE(proc, , , lwp__exit);
 129
 130 /*
 131  * thread related storage.
 132  */
 133 static uma_zone_t thread_zone;
 134
 135 struct thread_domain_data {
 136         struct thread   *tdd_zombies;
 137         int             tdd_reapticks;
 138 } __aligned(CACHE_LINE_SIZE);
 139
 140 static struct thread_domain_data thread_domain_data[MAXMEMDOM];
 141
 142 static struct task      thread_reap_task;
 143 static struct callout   thread_reap_callout;
 144
 145 static void thread_zombie(struct thread *);
 146 static void thread_reap(void);
 147 static void thread_reap_all(void);
 148 static void thread_reap_task_cb(void *, int);
 149 static void thread_reap_callout_cb(void *);
 150 static int thread_unsuspend_one(struct thread *td, struct proc *p,
 151     bool boundary);
 152 static void thread_free_batched(struct thread *td);
 153
 154 static __exclusive_cache_line struct mtx tid_lock;
 155 static bitstr_t *tid_bitmap;
 156
 157 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
 158
 159 static int maxthread;
 160 SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN,
 161     &maxthread, 0, "Maximum number of threads");
 162
 163 static __exclusive_cache_line int nthreads;
 164
 165 static LIST_HEAD(tidhashhead, thread) *tidhashtbl;
 166 static u_long   tidhash;
 167 static u_long   tidhashlock;
 168 static struct   rwlock *tidhashtbl_lock;
 169 #define TIDHASH(tid)            (&tidhashtbl[(tid) & tidhash])
 170 #define TIDHASHLOCK(tid)        (&tidhashtbl_lock[(tid) & tidhashlock])
 171
 172 EVENTHANDLER_LIST_DEFINE(thread_ctor);
 173 EVENTHANDLER_LIST_DEFINE(thread_dtor);
 174 EVENTHANDLER_LIST_DEFINE(thread_init);
 175 EVENTHANDLER_LIST_DEFINE(thread_fini);
 176
 177 static bool
 178 thread_count_inc_try(void)
 179 {
 180         int nthreads_new;
 181
 182         nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1;
 183         if (nthreads_new >= maxthread - 100) {
 184                 if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 ||
 185                     nthreads_new >= maxthread) {
 186                         atomic_subtract_int(&nthreads, 1);
 187                         return (false);
 188                 }
 189         }
 190         return (true);
 191 }
 192
 193 static bool
 194 thread_count_inc(void)
 195 {
 196         static struct timeval lastfail;
 197         static int curfail;
 198
 199         thread_reap();
 200         if (thread_count_inc_try()) {
 201                 return (true);
 202         }
 203
 204         thread_reap_all();
 205         if (thread_count_inc_try()) {
 206                 return (true);
 207         }
 208
 209         if (ppsratecheck(&lastfail, &curfail, 1)) {
 210                 printf("maxthread limit exceeded by uid %u "
 211                     "(pid %d); consider increasing kern.maxthread\n",
 212                     curthread->td_ucred->cr_ruid, curproc->p_pid);
 213         }
 214         return (false);
 215 }
 216
 217 static void
 218 thread_count_sub(int n)
 219 {
 220
 221         atomic_subtract_int(&nthreads, n);
 222 }
 223
 224 static void
 225 thread_count_dec(void)
 226 {
 227
 228         thread_count_sub(1);
 229 }
 230
 231 static lwpid_t
 232 tid_alloc(void)
 233 {
 234         static lwpid_t trytid;
 235         lwpid_t tid;
 236
 237         mtx_lock(&tid_lock);
 238         /*
 239          * It is an invariant that the bitmap is big enough to hold maxthread
 240          * IDs. If we got to this point there has to be at least one free.
 241          */
 242         if (trytid >= maxthread)
 243                 trytid = 0;
 244         bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
 245         if (tid == -1) {
 246                 KASSERT(trytid != 0, ("unexpectedly ran out of IDs"));
 247                 trytid = 0;
 248                 bit_ffc_at(tid_bitmap, trytid, maxthread, &tid);
 249                 KASSERT(tid != -1, ("unexpectedly ran out of IDs"));
 250         }
 251         bit_set(tid_bitmap, tid);
 252         trytid = tid + 1;
 253         mtx_unlock(&tid_lock);
 254         return (tid + NO_PID);
 255 }
 256
 257 static void
 258 tid_free_locked(lwpid_t rtid)
 259 {
 260         lwpid_t tid;
 261
 262         mtx_assert(&tid_lock, MA_OWNED);
 263         KASSERT(rtid >= NO_PID,
 264             ("%s: invalid tid %d\n", __func__, rtid));
 265         tid = rtid - NO_PID;
 266         KASSERT(bit_test(tid_bitmap, tid) != 0,
 267             ("thread ID %d not allocated\n", rtid));
 268         bit_clear(tid_bitmap, tid);
 269 }
 270
 271 static void
 272 tid_free(lwpid_t rtid)
 273 {
 274
 275         mtx_lock(&tid_lock);
 276         tid_free_locked(rtid);
 277         mtx_unlock(&tid_lock);
 278 }
 279
 280 static void
 281 tid_free_batch(lwpid_t *batch, int n)
 282 {
 283         int i;
 284
 285         mtx_lock(&tid_lock);
 286         for (i = 0; i < n; i++) {
 287                 tid_free_locked(batch[i]);
 288         }
 289         mtx_unlock(&tid_lock);
 290 }
 291
 292 /*
 293  * Batching for thread reapping.
 294  */
 295 struct tidbatch {
 296         lwpid_t tab[16];
 297         int n;
 298 };
 299
 300 static void
 301 tidbatch_prep(struct tidbatch *tb)
 302 {
 303
 304         tb->n = 0;
 305 }
 306
 307 static void
 308 tidbatch_add(struct tidbatch *tb, struct thread *td)
 309 {
 310
 311         KASSERT(tb->n < nitems(tb->tab),
 312             ("%s: count too high %d", __func__, tb->n));
 313         tb->tab[tb->n] = td->td_tid;
 314         tb->n++;
 315 }
 316
 317 static void
 318 tidbatch_process(struct tidbatch *tb)
 319 {
 320
 321         KASSERT(tb->n <= nitems(tb->tab),
 322             ("%s: count too high %d", __func__, tb->n));
 323         if (tb->n == nitems(tb->tab)) {
 324                 tid_free_batch(tb->tab, tb->n);
 325                 tb->n = 0;
 326         }
 327 }
 328
 329 static void
 330 tidbatch_final(struct tidbatch *tb)
 331 {
 332
 333         KASSERT(tb->n <= nitems(tb->tab),
 334             ("%s: count too high %d", __func__, tb->n));
 335         if (tb->n != 0) {
 336                 tid_free_batch(tb->tab, tb->n);
 337         }
 338 }
 339
 340 /*
 341  * Prepare a thread for use.
 342  */
 343 static int
 344 thread_ctor(void *mem, int size, void *arg, int flags)
 345 {
 346         struct thread   *td;
 347
 348         td = (struct thread *)mem;
 349         td->td_state = TDS_INACTIVE;
 350         td->td_lastcpu = td->td_oncpu = NOCPU;
 351
 352         /*
 353          * Note that td_critnest begins life as 1 because the thread is not
 354          * running and is thereby implicitly waiting to be on the receiving
 355          * end of a context switch.
 356          */
 357         td->td_critnest = 1;
 358         td->td_lend_user_pri = PRI_MAX;
 359 #ifdef AUDIT
 360         audit_thread_alloc(td);
 361 #endif
 362 #ifdef KDTRACE_HOOKS
 363         kdtrace_thread_ctor(td);
 364 #endif
 365         umtx_thread_alloc(td);
 366         MPASS(td->td_sel == NULL);
 367         return (0);
 368 }
 369
 370 /*
 371  * Reclaim a thread after use.
 372  */
 373 static void
 374 thread_dtor(void *mem, int size, void *arg)
 375 {
 376         struct thread *td;
 377
 378         td = (struct thread *)mem;
 379
 380 #ifdef INVARIANTS
 381         /* Verify that this thread is in a safe state to free. */
 382         switch (td->td_state) {
 383         case TDS_INHIBITED:
 384         case TDS_RUNNING:
 385         case TDS_CAN_RUN:
 386         case TDS_RUNQ:
 387                 /*
 388                  * We must never unlink a thread that is in one of
 389                  * these states, because it is currently active.
 390                  */
 391                 panic("bad state for thread unlinking");
 392                 /* NOTREACHED */
 393         case TDS_INACTIVE:
 394                 break;
 395         default:
 396                 panic("bad thread state");
 397                 /* NOTREACHED */
 398         }
 399 #endif
 400 #ifdef AUDIT
 401         audit_thread_free(td);
 402 #endif
 403 #ifdef KDTRACE_HOOKS
 404         kdtrace_thread_dtor(td);
 405 #endif
 406         /* Free all OSD associated to this thread. */
 407         osd_thread_exit(td);
 408         td_softdep_cleanup(td);
 409         MPASS(td->td_su == NULL);
 410         seltdfini(td);
 411 }
 412
 413 /*
 414  * Initialize type-stable parts of a thread (when newly created).
 415  */
 416 static int
 417 thread_init(void *mem, int size, int flags)
 418 {
 419         struct thread *td;
 420
 421         td = (struct thread *)mem;
 422
 423         td->td_allocdomain = vm_phys_domain(vtophys(td));
 424         td->td_sleepqueue = sleepq_alloc();
 425         td->td_turnstile = turnstile_alloc();
 426         td->td_rlqe = NULL;
 427         EVENTHANDLER_DIRECT_INVOKE(thread_init, td);
 428         umtx_thread_init(td);
 429         td->td_kstack = 0;
 430         td->td_sel = NULL;
 431         return (0);
 432 }
 433
 434 /*
 435  * Tear down type-stable parts of a thread (just before being discarded).
 436  */
 437 static void
 438 thread_fini(void *mem, int size)
 439 {
 440         struct thread *td;
 441
 442         td = (struct thread *)mem;
 443         EVENTHANDLER_DIRECT_INVOKE(thread_fini, td);
 444         rlqentry_free(td->td_rlqe);
 445         turnstile_free(td->td_turnstile);
 446         sleepq_free(td->td_sleepqueue);
 447         umtx_thread_fini(td);
 448         MPASS(td->td_sel == NULL);
 449 }
 450
 451 /*
 452  * For a newly created process,
 453  * link up all the structures and its initial threads etc.
 454  * called from:
 455  * {arch}/{arch}/machdep.c   {arch}_init(), init386() etc.
 456  * proc_dtor() (should go away)
 457  * proc_init()
 458  */
 459 void
 460 proc_linkup0(struct proc *p, struct thread *td)
 461 {
 462         TAILQ_INIT(&p->p_threads);           /* all threads in proc */
 463         proc_linkup(p, td);
 464 }
 465
 466 void
 467 proc_linkup(struct proc *p, struct thread *td)
 468 {
 469
 470         sigqueue_init(&p->p_sigqueue, p);
 471         p->p_ksi = ksiginfo_alloc(1);
 472         if (p->p_ksi != NULL) {
 473                 /* XXX p_ksi may be null if ksiginfo zone is not ready */
 474                 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
 475         }
 476         LIST_INIT(&p->p_mqnotifier);
 477         p->p_numthreads = 0;
 478         thread_link(td, p);
 479 }
 480
 481 extern int max_threads_per_proc;
 482
 483 /*
 484  * Initialize global thread allocation resources.
 485  */
 486 void
 487 threadinit(void)
 488 {
 489         u_long i;
 490         lwpid_t tid0;
 491         uint32_t flags;
 492
 493         /*
 494          * Place an upper limit on threads which can be allocated.
 495          *
 496          * Note that other factors may make the de facto limit much lower.
 497          *
 498          * Platform limits are somewhat arbitrary but deemed "more than good
 499          * enough" for the foreseable future.
 500          */
 501         if (maxthread == 0) {
 502 #ifdef _LP64
 503                 maxthread = MIN(maxproc * max_threads_per_proc, 1000000);
 504 #else
 505                 maxthread = MIN(maxproc * max_threads_per_proc, 100000);
 506 #endif
 507         }
 508
 509         mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
 510         tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK);
 511         /*
 512          * Handle thread0.
 513          */
 514         thread_count_inc();
 515         tid0 = tid_alloc();
 516         if (tid0 != THREAD0_TID)
 517                 panic("tid0 %d != %d\n", tid0, THREAD0_TID);
 518
 519         flags = UMA_ZONE_NOFREE;
 520 #ifdef __aarch64__
 521         /*
 522          * Force thread structures to be allocated from the direct map.
 523          * Otherwise, superpage promotions and demotions may temporarily
 524          * invalidate thread structure mappings.  For most dynamically allocated
 525          * structures this is not a problem, but translation faults cannot be
 526          * handled without accessing curthread.
 527          */
 528         flags |= UMA_ZONE_CONTIG;
 529 #endif
 530         thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
 531             thread_ctor, thread_dtor, thread_init, thread_fini,
 532             32 - 1, flags);
 533         tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
 534         tidhashlock = (tidhash + 1) / 64;
 535         if (tidhashlock > 0)
 536                 tidhashlock--;
 537         tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1),
 538             M_TIDHASH, M_WAITOK | M_ZERO);
 539         for (i = 0; i < tidhashlock + 1; i++)
 540                 rw_init(&tidhashtbl_lock[i], "tidhash");
 541
 542         TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL);
 543         callout_init(&thread_reap_callout, 1);
 544         callout_reset(&thread_reap_callout, 5 * hz,
 545             thread_reap_callout_cb, NULL);
 546 }
 547
 548 /*
 549  * Place an unused thread on the zombie list.
 550  */
 551 void
 552 thread_zombie(struct thread *td)
 553 {
 554         struct thread_domain_data *tdd;
 555         struct thread *ztd;
 556
 557         tdd = &thread_domain_data[td->td_allocdomain];
 558         ztd = atomic_load_ptr(&tdd->tdd_zombies);
 559         for (;;) {
 560                 td->td_zombie = ztd;
 561                 if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies,
 562                     (uintptr_t *)&ztd, (uintptr_t)td))
 563                         break;
 564                 continue;
 565         }
 566 }
 567
 568 /*
 569  * Release a thread that has exited after cpu_throw().
 570  */
 571 void
 572 thread_stash(struct thread *td)
 573 {
 574         atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
 575         thread_zombie(td);
 576 }
 577
 578 /*
 579  * Reap zombies from passed domain.
 580  */
 581 static void
 582 thread_reap_domain(struct thread_domain_data *tdd)
 583 {
 584         struct thread *itd, *ntd;
 585         struct tidbatch tidbatch;
 586         struct credbatch credbatch;
 587         int tdcount;
 588         struct plimit *lim;
 589         int limcount;
 590
 591         /*
 592          * Reading upfront is pessimal if followed by concurrent atomic_swap,
 593          * but most of the time the list is empty.
 594          */
 595         if (tdd->tdd_zombies == NULL)
 596                 return;
 597
 598         itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies,
 599             (uintptr_t)NULL);
 600         if (itd == NULL)
 601                 return;
 602
 603         /*
 604          * Multiple CPUs can get here, the race is fine as ticks is only
 605          * advisory.
 606          */
 607         tdd->tdd_reapticks = ticks;
 608
 609         tidbatch_prep(&tidbatch);
 610         credbatch_prep(&credbatch);
 611         tdcount = 0;
 612         lim = NULL;
 613         limcount = 0;
 614
 615         while (itd != NULL) {
 616                 ntd = itd->td_zombie;
 617                 EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd);
 618                 tidbatch_add(&tidbatch, itd);
 619                 credbatch_add(&credbatch, itd);
 620                 MPASS(itd->td_limit != NULL);
 621                 if (lim != itd->td_limit) {
 622                         if (limcount != 0) {
 623                                 lim_freen(lim, limcount);
 624                                 limcount = 0;
 625                         }
 626                 }
 627                 lim = itd->td_limit;
 628                 limcount++;
 629                 thread_free_batched(itd);
 630                 tidbatch_process(&tidbatch);
 631                 credbatch_process(&credbatch);
 632                 tdcount++;
 633                 if (tdcount == 32) {
 634                         thread_count_sub(tdcount);
 635                         tdcount = 0;
 636                 }
 637                 itd = ntd;
 638         }
 639
 640         tidbatch_final(&tidbatch);
 641         credbatch_final(&credbatch);
 642         if (tdcount != 0) {
 643                 thread_count_sub(tdcount);
 644         }
 645         MPASS(limcount != 0);
 646         lim_freen(lim, limcount);
 647 }
 648
 649 /*
 650  * Reap zombies from all domains.
 651  */
 652 static void
 653 thread_reap_all(void)
 654 {
 655         struct thread_domain_data *tdd;
 656         int i, domain;
 657
 658         domain = PCPU_GET(domain);
 659         for (i = 0; i < vm_ndomains; i++) {
 660                 tdd = &thread_domain_data[(i + domain) % vm_ndomains];
 661                 thread_reap_domain(tdd);
 662         }
 663 }
 664
 665 /*
 666  * Reap zombies from local domain.
 667  */
 668 static void
 669 thread_reap(void)
 670 {
 671         struct thread_domain_data *tdd;
 672         int domain;
 673
 674         domain = PCPU_GET(domain);
 675         tdd = &thread_domain_data[domain];
 676
 677         thread_reap_domain(tdd);
 678 }
 679
 680 static void
 681 thread_reap_task_cb(void *arg __unused, int pending __unused)
 682 {
 683
 684         thread_reap_all();
 685 }
 686
 687 static void
 688 thread_reap_callout_cb(void *arg __unused)
 689 {
 690         struct thread_domain_data *tdd;
 691         int i, cticks, lticks;
 692         bool wantreap;
 693
 694         wantreap = false;
 695         cticks = atomic_load_int(&ticks);
 696         for (i = 0; i < vm_ndomains; i++) {
 697                 tdd = &thread_domain_data[i];
 698                 lticks = tdd->tdd_reapticks;
 699                 if (tdd->tdd_zombies != NULL &&
 700                     (u_int)(cticks - lticks) > 5 * hz) {
 701                         wantreap = true;
 702                         break;
 703                 }
 704         }
 705
 706         if (wantreap)
 707                 taskqueue_enqueue(taskqueue_thread, &thread_reap_task);
 708         callout_reset(&thread_reap_callout, 5 * hz,
 709             thread_reap_callout_cb, NULL);
 710 }
 711
 712 /*
 713  * Calling this function guarantees that any thread that exited before
 714  * the call is reaped when the function returns.  By 'exited' we mean
 715  * a thread removed from the process linkage with thread_unlink().
 716  * Practically this means that caller must lock/unlock corresponding
 717  * process lock before the call, to synchronize with thread_exit().
 718  */
 719 void
 720 thread_reap_barrier(void)
 721 {
 722         struct task *t;
 723
 724         /*
 725          * First do context switches to each CPU to ensure that all
 726          * PCPU pc_deadthreads are moved to zombie list.
 727          */
 728         quiesce_all_cpus("", PDROP);
 729
 730         /*
 731          * Second, fire the task in the same thread as normal
 732          * thread_reap() is done, to serialize reaping.
 733          */
 734         t = malloc(sizeof(*t), M_TEMP, M_WAITOK);
 735         TASK_INIT(t, 0, thread_reap_task_cb, t);
 736         taskqueue_enqueue(taskqueue_thread, t);
 737         taskqueue_drain(taskqueue_thread, t);
 738         free(t, M_TEMP);
 739 }
 740
 741 /*
 742  * Allocate a thread.
 743  */
 744 struct thread *
 745 thread_alloc(int pages)
 746 {
 747         struct thread *td;
 748         lwpid_t tid;
 749
 750         if (!thread_count_inc()) {
 751                 return (NULL);
 752         }
 753
 754         tid = tid_alloc();
 755         td = uma_zalloc(thread_zone, M_WAITOK);
 756         KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
 757         if (!vm_thread_new(td, pages)) {
 758                 uma_zfree(thread_zone, td);
 759                 tid_free(tid);
 760                 thread_count_dec();
 761                 return (NULL);
 762         }
 763         td->td_tid = tid;
 764         bzero(&td->td_sa.args, sizeof(td->td_sa.args));
 765         cpu_thread_alloc(td);
 766         EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td);
 767         return (td);
 768 }
 769
 770 int
 771 thread_alloc_stack(struct thread *td, int pages)
 772 {
 773
 774         KASSERT(td->td_kstack == 0,
 775             ("thread_alloc_stack called on a thread with kstack"));
 776         if (!vm_thread_new(td, pages))
 777                 return (0);
 778         cpu_thread_alloc(td);
 779         return (1);
 780 }
 781
 782 /*
 783  * Deallocate a thread.
 784  */
 785 static void
 786 thread_free_batched(struct thread *td)
 787 {
 788
 789         lock_profile_thread_exit(td);
 790         if (td->td_cpuset)
 791                 cpuset_rel(td->td_cpuset);
 792         td->td_cpuset = NULL;
 793         cpu_thread_free(td);
 794         if (td->td_kstack != 0)
 795                 vm_thread_dispose(td);
 796         callout_drain(&td->td_slpcallout);
 797         /*
 798          * Freeing handled by the caller.
 799          */
 800         td->td_tid = -1;
 801         uma_zfree(thread_zone, td);
 802 }
 803
 804 void
 805 thread_free(struct thread *td)
 806 {
 807         lwpid_t tid;
 808
 809         EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td);
 810         tid = td->td_tid;
 811         thread_free_batched(td);
 812         tid_free(tid);
 813         thread_count_dec();
 814 }
 815
 816 void
 817 thread_cow_get_proc(struct thread *newtd, struct proc *p)
 818 {
 819
 820         PROC_LOCK_ASSERT(p, MA_OWNED);
 821         newtd->td_realucred = crcowget(p->p_ucred);
 822         newtd->td_ucred = newtd->td_realucred;
 823         newtd->td_limit = lim_hold(p->p_limit);
 824         newtd->td_cowgen = p->p_cowgen;
 825 }
 826
 827 void
 828 thread_cow_get(struct thread *newtd, struct thread *td)
 829 {
 830
 831         MPASS(td->td_realucred == td->td_ucred);
 832         newtd->td_realucred = crcowget(td->td_realucred);
 833         newtd->td_ucred = newtd->td_realucred;
 834         newtd->td_limit = lim_hold(td->td_limit);
 835         newtd->td_cowgen = td->td_cowgen;
 836 }
 837
 838 void
 839 thread_cow_free(struct thread *td)
 840 {
 841
 842         if (td->td_realucred != NULL)
 843                 crcowfree(td);
 844         if (td->td_limit != NULL)
 845                 lim_free(td->td_limit);
 846 }
 847
 848 void
 849 thread_cow_update(struct thread *td)
 850 {
 851         struct proc *p;
 852         struct ucred *oldcred;
 853         struct plimit *oldlimit;
 854
 855         p = td->td_proc;
 856         oldlimit = NULL;
 857         PROC_LOCK(p);
 858         oldcred = crcowsync();
 859         if (td->td_limit != p->p_limit) {
 860                 oldlimit = td->td_limit;
 861                 td->td_limit = lim_hold(p->p_limit);
 862         }
 863         td->td_cowgen = p->p_cowgen;
 864         PROC_UNLOCK(p);
 865         if (oldcred != NULL)
 866                 crfree(oldcred);
 867         if (oldlimit != NULL)
 868                 lim_free(oldlimit);
 869 }
 870
 871 /*
 872  * Discard the current thread and exit from its context.
 873  * Always called with scheduler locked.
 874  *
 875  * Because we can't free a thread while we're operating under its context,
 876  * push the current thread into our CPU's deadthread holder. This means
 877  * we needn't worry about someone else grabbing our context before we
 878  * do a cpu_throw().
 879  */
 880 void
 881 thread_exit(void)
 882 {
 883         uint64_t runtime, new_switchtime;
 884         struct thread *td;
 885         struct thread *td2;
 886         struct proc *p;
 887         int wakeup_swapper;
 888
 889         td = curthread;
 890         p = td->td_proc;
 891
 892         PROC_SLOCK_ASSERT(p, MA_OWNED);
 893         mtx_assert(&Giant, MA_NOTOWNED);
 894
 895         PROC_LOCK_ASSERT(p, MA_OWNED);
 896         KASSERT(p != NULL, ("thread exiting without a process"));
 897         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 898             (long)p->p_pid, td->td_name);
 899         SDT_PROBE0(proc, , , lwp__exit);
 900         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 901         MPASS(td->td_realucred == td->td_ucred);
 902
 903         /*
 904          * drop FPU & debug register state storage, or any other
 905          * architecture specific resources that
 906          * would not be on a new untouched process.
 907          */
 908         cpu_thread_exit(td);
 909
 910         /*
 911          * The last thread is left attached to the process
 912          * So that the whole bundle gets recycled. Skip
 913          * all this stuff if we never had threads.
 914          * EXIT clears all sign of other threads when
 915          * it goes to single threading, so the last thread always
 916          * takes the short path.
 917          */
 918         if (p->p_flag & P_HADTHREADS) {
 919                 if (p->p_numthreads > 1) {
 920                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 921                         thread_unlink(td);
 922                         td2 = FIRST_THREAD_IN_PROC(p);
 923                         sched_exit_thread(td2, td);
 924
 925                         /*
 926                          * The test below is NOT true if we are the
 927                          * sole exiting thread. P_STOPPED_SINGLE is unset
 928                          * in exit1() after it is the only survivor.
 929                          */
 930                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 931                                 if (p->p_numthreads == p->p_suspcount) {
 932                                         thread_lock(p->p_singlethread);
 933                                         wakeup_swapper = thread_unsuspend_one(
 934                                                 p->p_singlethread, p, false);
 935                                         if (wakeup_swapper)
 936                                                 kick_proc0();
 937                                 }
 938                         }
 939
 940                         PCPU_SET(deadthread, td);
 941                 } else {
 942                         /*
 943                          * The last thread is exiting.. but not through exit()
 944                          */
 945                         panic ("thread_exit: Last thread exiting on its own");
 946                 }
 947         }
 948 #ifdef  HWPMC_HOOKS
 949         /*
 950          * If this thread is part of a process that is being tracked by hwpmc(4),
 951          * inform the module of the thread's impending exit.
 952          */
 953         if (PMC_PROC_IS_USING_PMCS(td->td_proc)) {
 954                 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
 955                 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL);
 956         } else if (PMC_SYSTEM_SAMPLING_ACTIVE())
 957                 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL);
 958 #endif
 959         PROC_UNLOCK(p);
 960         PROC_STATLOCK(p);
 961         thread_lock(td);
 962         PROC_SUNLOCK(p);
 963
 964         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 965         new_switchtime = cpu_ticks();
 966         runtime = new_switchtime - PCPU_GET(switchtime);
 967         td->td_runtime += runtime;
 968         td->td_incruntime += runtime;
 969         PCPU_SET(switchtime, new_switchtime);
 970         PCPU_SET(switchticks, ticks);
 971         VM_CNT_INC(v_swtch);
 972
 973         /* Save our resource usage in our process. */
 974         td->td_ru.ru_nvcsw++;
 975         ruxagg_locked(p, td);
 976         rucollect(&p->p_ru, &td->td_ru);
 977         PROC_STATUNLOCK(p);
 978
 979         td->td_state = TDS_INACTIVE;
 980 #ifdef WITNESS
 981         witness_thread_exit(td);
 982 #endif
 983         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 984         sched_throw(td);
 985         panic("I'm a teapot!");
 986         /* NOTREACHED */
 987 }
 988
 989 /*
 990  * Do any thread specific cleanups that may be needed in wait()
 991  * called with Giant, proc and schedlock not held.
 992  */
 993 void
 994 thread_wait(struct proc *p)
 995 {
 996         struct thread *td;
 997
 998         mtx_assert(&Giant, MA_NOTOWNED);
 999         KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()"));
1000         KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking"));
1001         td = FIRST_THREAD_IN_PROC(p);
1002         /* Lock the last thread so we spin until it exits cpu_throw(). */
1003         thread_lock(td);
1004         thread_unlock(td);
1005         lock_profile_thread_exit(td);
1006         cpuset_rel(td->td_cpuset);
1007         td->td_cpuset = NULL;
1008         cpu_thread_clean(td);
1009         thread_cow_free(td);
1010         callout_drain(&td->td_slpcallout);
1011         thread_reap();  /* check for zombie threads etc. */
1012 }
1013
1014 /*
1015  * Link a thread to a process.
1016  * set up anything that needs to be initialized for it to
1017  * be used by the process.
1018  */
1019 void
1020 thread_link(struct thread *td, struct proc *p)
1021 {
1022
1023         /*
1024          * XXX This can't be enabled because it's called for proc0 before
1025          * its lock has been created.
1026          * PROC_LOCK_ASSERT(p, MA_OWNED);
1027          */
1028         td->td_state    = TDS_INACTIVE;
1029         td->td_proc     = p;
1030         td->td_flags    = TDF_INMEM;
1031
1032         LIST_INIT(&td->td_contested);
1033         LIST_INIT(&td->td_lprof[0]);
1034         LIST_INIT(&td->td_lprof[1]);
1035 #ifdef EPOCH_TRACE
1036         SLIST_INIT(&td->td_epochs);
1037 #endif
1038         sigqueue_init(&td->td_sigqueue, p);
1039         callout_init(&td->td_slpcallout, 1);
1040         TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist);
1041         p->p_numthreads++;
1042 }
1043
1044 /*
1045  * Called from:
1046  *  thread_exit()
1047  */
1048 void
1049 thread_unlink(struct thread *td)
1050 {
1051         struct proc *p = td->td_proc;
1052
1053         PROC_LOCK_ASSERT(p, MA_OWNED);
1054 #ifdef EPOCH_TRACE
1055         MPASS(SLIST_EMPTY(&td->td_epochs));
1056 #endif
1057
1058         TAILQ_REMOVE(&p->p_threads, td, td_plist);
1059         p->p_numthreads--;
1060         /* could clear a few other things here */
1061         /* Must  NOT clear links to proc! */
1062 }
1063
1064 static int
1065 calc_remaining(struct proc *p, int mode)
1066 {
1067         int remaining;
1068
1069         PROC_LOCK_ASSERT(p, MA_OWNED);
1070         PROC_SLOCK_ASSERT(p, MA_OWNED);
1071         if (mode == SINGLE_EXIT)
1072                 remaining = p->p_numthreads;
1073         else if (mode == SINGLE_BOUNDARY)
1074                 remaining = p->p_numthreads - p->p_boundary_count;
1075         else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC)
1076                 remaining = p->p_numthreads - p->p_suspcount;
1077         else
1078                 panic("calc_remaining: wrong mode %d", mode);
1079         return (remaining);
1080 }
1081
1082 static int
1083 remain_for_mode(int mode)
1084 {
1085
1086         return (mode == SINGLE_ALLPROC ? 0 : 1);
1087 }
1088
1089 static int
1090 weed_inhib(int mode, struct thread *td2, struct proc *p)
1091 {
1092         int wakeup_swapper;
1093
1094         PROC_LOCK_ASSERT(p, MA_OWNED);
1095         PROC_SLOCK_ASSERT(p, MA_OWNED);
1096         THREAD_LOCK_ASSERT(td2, MA_OWNED);
1097
1098         wakeup_swapper = 0;
1099
1100         /*
1101          * Since the thread lock is dropped by the scheduler we have
1102          * to retry to check for races.
1103          */
1104 restart:
1105         switch (mode) {
1106         case SINGLE_EXIT:
1107                 if (TD_IS_SUSPENDED(td2)) {
1108                         wakeup_swapper |= thread_unsuspend_one(td2, p, true);
1109                         thread_lock(td2);
1110                         goto restart;
1111                 }
1112                 if (TD_CAN_ABORT(td2)) {
1113                         wakeup_swapper |= sleepq_abort(td2, EINTR);
1114                         return (wakeup_swapper);
1115                 }
1116                 break;
1117         case SINGLE_BOUNDARY:
1118         case SINGLE_NO_EXIT:
1119                 if (TD_IS_SUSPENDED(td2) &&
1120                     (td2->td_flags & TDF_BOUNDARY) == 0) {
1121                         wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1122                         thread_lock(td2);
1123                         goto restart;
1124                 }
1125                 if (TD_CAN_ABORT(td2)) {
1126                         wakeup_swapper |= sleepq_abort(td2, ERESTART);
1127                         return (wakeup_swapper);
1128                 }
1129                 break;
1130         case SINGLE_ALLPROC:
1131                 /*
1132                  * ALLPROC suspend tries to avoid spurious EINTR for
1133                  * threads sleeping interruptable, by suspending the
1134                  * thread directly, similarly to sig_suspend_threads().
1135                  * Since such sleep is not neccessary performed at the user
1136                  * boundary, TDF_ALLPROCSUSP is used to avoid immediate
1137                  * un-suspend.
1138                  */
1139                 if (TD_IS_SUSPENDED(td2) && (td2->td_flags &
1140                     TDF_ALLPROCSUSP) == 0) {
1141                         wakeup_swapper |= thread_unsuspend_one(td2, p, false);
1142                         thread_lock(td2);
1143                         goto restart;
1144                 }
1145                 if (TD_CAN_ABORT(td2)) {
1146                         td2->td_flags |= TDF_ALLPROCSUSP;
1147                         wakeup_swapper |= sleepq_abort(td2, ERESTART);
1148                         return (wakeup_swapper);
1149                 }
1150                 break;
1151         default:
1152                 break;
1153         }
1154         thread_unlock(td2);
1155         return (wakeup_swapper);
1156 }
1157
1158 /*
1159  * Enforce single-threading.
1160  *
1161  * Returns 1 if the caller must abort (another thread is waiting to
1162  * exit the process or similar). Process is locked!
1163  * Returns 0 when you are successfully the only thread running.
1164  * A process has successfully single threaded in the suspend mode when
1165  * There are no threads in user mode. Threads in the kernel must be
1166  * allowed to continue until they get to the user boundary. They may even
1167  * copy out their return values and data before suspending. They may however be
1168  * accelerated in reaching the user boundary as we will wake up
1169  * any sleeping threads that are interruptable. (PCATCH).
1170  */
1171 int
1172 thread_single(struct proc *p, int mode)
1173 {
1174         struct thread *td;
1175         struct thread *td2;
1176         int remaining, wakeup_swapper;
1177
1178         td = curthread;
1179         KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1180             mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1181             ("invalid mode %d", mode));
1182         /*
1183          * If allowing non-ALLPROC singlethreading for non-curproc
1184          * callers, calc_remaining() and remain_for_mode() should be
1185          * adjusted to also account for td->td_proc != p.  For now
1186          * this is not implemented because it is not used.
1187          */
1188         KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) ||
1189             (mode != SINGLE_ALLPROC && td->td_proc == p),
1190             ("mode %d proc %p curproc %p", mode, p, td->td_proc));
1191         mtx_assert(&Giant, MA_NOTOWNED);
1192         PROC_LOCK_ASSERT(p, MA_OWNED);
1193
1194         /*
1195          * Is someone already single threading?
1196          * Or may be singlethreading is not needed at all.
1197          */
1198         if (mode == SINGLE_ALLPROC) {
1199                 while ((p->p_flag & P_STOPPED_SINGLE) != 0) {
1200                         if ((p->p_flag2 & P2_WEXIT) != 0)
1201                                 return (1);
1202                         msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0);
1203                 }
1204         } else if ((p->p_flag & P_HADTHREADS) == 0)
1205                 return (0);
1206         if (p->p_singlethread != NULL && p->p_singlethread != td)
1207                 return (1);
1208
1209         if (mode == SINGLE_EXIT) {
1210                 p->p_flag |= P_SINGLE_EXIT;
1211                 p->p_flag &= ~P_SINGLE_BOUNDARY;
1212         } else {
1213                 p->p_flag &= ~P_SINGLE_EXIT;
1214                 if (mode == SINGLE_BOUNDARY)
1215                         p->p_flag |= P_SINGLE_BOUNDARY;
1216                 else
1217                         p->p_flag &= ~P_SINGLE_BOUNDARY;
1218         }
1219         if (mode == SINGLE_ALLPROC) {
1220                 p->p_flag |= P_TOTAL_STOP;
1221                 thread_lock(td);
1222                 td->td_flags |= TDF_DOING_SA;
1223                 thread_unlock(td);
1224         }
1225         p->p_flag |= P_STOPPED_SINGLE;
1226         PROC_SLOCK(p);
1227         p->p_singlethread = td;
1228         remaining = calc_remaining(p, mode);
1229         while (remaining != remain_for_mode(mode)) {
1230                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
1231                         goto stopme;
1232                 wakeup_swapper = 0;
1233                 FOREACH_THREAD_IN_PROC(p, td2) {
1234                         if (td2 == td)
1235                                 continue;
1236                         thread_lock(td2);
1237                         td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
1238                         if (TD_IS_INHIBITED(td2)) {
1239                                 wakeup_swapper |= weed_inhib(mode, td2, p);
1240 #ifdef SMP
1241                         } else if (TD_IS_RUNNING(td2)) {
1242                                 forward_signal(td2);
1243                                 thread_unlock(td2);
1244 #endif
1245                         } else
1246                                 thread_unlock(td2);
1247                 }
1248                 if (wakeup_swapper)
1249                         kick_proc0();
1250                 remaining = calc_remaining(p, mode);
1251
1252                 /*
1253                  * Maybe we suspended some threads.. was it enough?
1254                  */
1255                 if (remaining == remain_for_mode(mode))
1256                         break;
1257
1258 stopme:
1259                 /*
1260                  * Wake us up when everyone else has suspended.
1261                  * In the mean time we suspend as well.
1262                  */
1263                 thread_suspend_switch(td, p);
1264                 remaining = calc_remaining(p, mode);
1265         }
1266         if (mode == SINGLE_EXIT) {
1267                 /*
1268                  * Convert the process to an unthreaded process.  The
1269                  * SINGLE_EXIT is called by exit1() or execve(), in
1270                  * both cases other threads must be retired.
1271                  */
1272                 KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads"));
1273                 p->p_singlethread = NULL;
1274                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS);
1275
1276                 /*
1277                  * Wait for any remaining threads to exit cpu_throw().
1278                  */
1279                 while (p->p_exitthreads != 0) {
1280                         PROC_SUNLOCK(p);
1281                         PROC_UNLOCK(p);
1282                         sched_relinquish(td);
1283                         PROC_LOCK(p);
1284                         PROC_SLOCK(p);
1285                 }
1286         } else if (mode == SINGLE_BOUNDARY) {
1287                 /*
1288                  * Wait until all suspended threads are removed from
1289                  * the processors.  The thread_suspend_check()
1290                  * increments p_boundary_count while it is still
1291                  * running, which makes it possible for the execve()
1292                  * to destroy vmspace while our other threads are
1293                  * still using the address space.
1294                  *
1295                  * We lock the thread, which is only allowed to
1296                  * succeed after context switch code finished using
1297                  * the address space.
1298                  */
1299                 FOREACH_THREAD_IN_PROC(p, td2) {
1300                         if (td2 == td)
1301                                 continue;
1302                         thread_lock(td2);
1303                         KASSERT((td2->td_flags & TDF_BOUNDARY) != 0,
1304                             ("td %p not on boundary", td2));
1305                         KASSERT(TD_IS_SUSPENDED(td2),
1306                             ("td %p is not suspended", td2));
1307                         thread_unlock(td2);
1308                 }
1309         }
1310         PROC_SUNLOCK(p);
1311         if (mode == SINGLE_ALLPROC) {
1312                 thread_lock(td);
1313                 td->td_flags &= ~TDF_DOING_SA;
1314                 thread_unlock(td);
1315         }
1316         return (0);
1317 }
1318
1319 bool
1320 thread_suspend_check_needed(void)
1321 {
1322         struct proc *p;
1323         struct thread *td;
1324
1325         td = curthread;
1326         p = td->td_proc;
1327         PROC_LOCK_ASSERT(p, MA_OWNED);
1328         return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 &&
1329             (td->td_dbgflags & TDB_SUSPEND) != 0));
1330 }
1331
1332 /*
1333  * Called in from locations that can safely check to see
1334  * whether we have to suspend or at least throttle for a
1335  * single-thread event (e.g. fork).
1336  *
1337  * Such locations include userret().
1338  * If the "return_instead" argument is non zero, the thread must be able to
1339  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
1340  *
1341  * The 'return_instead' argument tells the function if it may do a
1342  * thread_exit() or suspend, or whether the caller must abort and back
1343  * out instead.
1344  *
1345  * If the thread that set the single_threading request has set the
1346  * P_SINGLE_EXIT bit in the process flags then this call will never return
1347  * if 'return_instead' is false, but will exit.
1348  *
1349  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
1350  *---------------+--------------------+---------------------
1351  *       0       | returns 0          |   returns 0 or 1
1352  *               | when ST ends       |   immediately
1353  *---------------+--------------------+---------------------
1354  *       1       | thread exits       |   returns 1
1355  *               |                    |  immediately
1356  * 0 = thread_exit() or suspension ok,
1357  * other = return error instead of stopping the thread.
1358  *
1359  * While a full suspension is under effect, even a single threading
1360  * thread would be suspended if it made this call (but it shouldn't).
1361  * This call should only be made from places where
1362  * thread_exit() would be safe as that may be the outcome unless
1363  * return_instead is set.
1364  */
1365 int
1366 thread_suspend_check(int return_instead)
1367 {
1368         struct thread *td;
1369         struct proc *p;
1370         int wakeup_swapper;
1371
1372         td = curthread;
1373         p = td->td_proc;
1374         mtx_assert(&Giant, MA_NOTOWNED);
1375         PROC_LOCK_ASSERT(p, MA_OWNED);
1376         while (thread_suspend_check_needed()) {
1377                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1378                         KASSERT(p->p_singlethread != NULL,
1379                             ("singlethread not set"));
1380                         /*
1381                          * The only suspension in action is a
1382                          * single-threading. Single threader need not stop.
1383                          * It is safe to access p->p_singlethread unlocked
1384                          * because it can only be set to our address by us.
1385                          */
1386                         if (p->p_singlethread == td)
1387                                 return (0);     /* Exempt from stopping. */
1388                 }
1389                 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
1390                         return (EINTR);
1391
1392                 /* Should we goto user boundary if we didn't come from there? */
1393                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1394                     (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
1395                         return (ERESTART);
1396
1397                 /*
1398                  * Ignore suspend requests if they are deferred.
1399                  */
1400                 if ((td->td_flags & TDF_SBDRY) != 0) {
1401                         KASSERT(return_instead,
1402                             ("TDF_SBDRY set for unsafe thread_suspend_check"));
1403                         KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
1404                             (TDF_SEINTR | TDF_SERESTART),
1405                             ("both TDF_SEINTR and TDF_SERESTART"));
1406                         return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0);
1407                 }
1408
1409                 /*
1410                  * If the process is waiting for us to exit,
1411                  * this thread should just suicide.
1412                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
1413                  */
1414                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
1415                         PROC_UNLOCK(p);
1416
1417                         /*
1418                          * Allow Linux emulation layer to do some work
1419                          * before thread suicide.
1420                          */
1421                         if (__predict_false(p->p_sysent->sv_thread_detach != NULL))
1422                                 (p->p_sysent->sv_thread_detach)(td);
1423                         umtx_thread_exit(td);
1424                         kern_thr_exit(td);
1425                         panic("stopped thread did not exit");
1426                 }
1427
1428                 PROC_SLOCK(p);
1429                 thread_stopped(p);
1430                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
1431                         if (p->p_numthreads == p->p_suspcount + 1) {
1432                                 thread_lock(p->p_singlethread);
1433                                 wakeup_swapper = thread_unsuspend_one(
1434                                     p->p_singlethread, p, false);
1435                                 if (wakeup_swapper)
1436                                         kick_proc0();
1437                         }
1438                 }
1439                 PROC_UNLOCK(p);
1440                 thread_lock(td);
1441                 /*
1442                  * When a thread suspends, it just
1443                  * gets taken off all queues.
1444                  */
1445                 thread_suspend_one(td);
1446                 if (return_instead == 0) {
1447                         p->p_boundary_count++;
1448                         td->td_flags |= TDF_BOUNDARY;
1449                 }
1450                 PROC_SUNLOCK(p);
1451                 mi_switch(SW_INVOL | SWT_SUSPEND);
1452                 PROC_LOCK(p);
1453         }
1454         return (0);
1455 }
1456
1457 /*
1458  * Check for possible stops and suspensions while executing a
1459  * casueword or similar transiently failing operation.
1460  *
1461  * The sleep argument controls whether the function can handle a stop
1462  * request itself or it should return ERESTART and the request is
1463  * proceed at the kernel/user boundary in ast.
1464  *
1465  * Typically, when retrying due to casueword(9) failure (rv == 1), we
1466  * should handle the stop requests there, with exception of cases when
1467  * the thread owns a kernel resource, for instance busied the umtx
1468  * key, or when functions return immediately if thread_check_susp()
1469  * returned non-zero.  On the other hand, retrying the whole lock
1470  * operation, we better not stop there but delegate the handling to
1471  * ast.
1472  *
1473  * If the request is for thread termination P_SINGLE_EXIT, we cannot
1474  * handle it at all, and simply return EINTR.
1475  */
1476 int
1477 thread_check_susp(struct thread *td, bool sleep)
1478 {
1479         struct proc *p;
1480         int error;
1481
1482         /*
1483          * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
1484          * eventually break the lockstep loop.
1485          */
1486         if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
1487                 return (0);
1488         error = 0;
1489         p = td->td_proc;
1490         PROC_LOCK(p);
1491         if (p->p_flag & P_SINGLE_EXIT)
1492                 error = EINTR;
1493         else if (P_SHOULDSTOP(p) ||
1494             ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
1495                 error = sleep ? thread_suspend_check(0) : ERESTART;
1496         PROC_UNLOCK(p);
1497         return (error);
1498 }
1499
1500 void
1501 thread_suspend_switch(struct thread *td, struct proc *p)
1502 {
1503
1504         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1505         PROC_LOCK_ASSERT(p, MA_OWNED);
1506         PROC_SLOCK_ASSERT(p, MA_OWNED);
1507         /*
1508          * We implement thread_suspend_one in stages here to avoid
1509          * dropping the proc lock while the thread lock is owned.
1510          */
1511         if (p == td->td_proc) {
1512                 thread_stopped(p);
1513                 p->p_suspcount++;
1514         }
1515         PROC_UNLOCK(p);
1516         thread_lock(td);
1517         td->td_flags &= ~TDF_NEEDSUSPCHK;
1518         TD_SET_SUSPENDED(td);
1519         sched_sleep(td, 0);
1520         PROC_SUNLOCK(p);
1521         DROP_GIANT();
1522         mi_switch(SW_VOL | SWT_SUSPEND);
1523         PICKUP_GIANT();
1524         PROC_LOCK(p);
1525         PROC_SLOCK(p);
1526 }
1527
1528 void
1529 thread_suspend_one(struct thread *td)
1530 {
1531         struct proc *p;
1532
1533         p = td->td_proc;
1534         PROC_SLOCK_ASSERT(p, MA_OWNED);
1535         THREAD_LOCK_ASSERT(td, MA_OWNED);
1536         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
1537         p->p_suspcount++;
1538         td->td_flags &= ~TDF_NEEDSUSPCHK;
1539         TD_SET_SUSPENDED(td);
1540         sched_sleep(td, 0);
1541 }
1542
1543 static int
1544 thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary)
1545 {
1546
1547         THREAD_LOCK_ASSERT(td, MA_OWNED);
1548         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1549         TD_CLR_SUSPENDED(td);
1550         td->td_flags &= ~TDF_ALLPROCSUSP;
1551         if (td->td_proc == p) {
1552                 PROC_SLOCK_ASSERT(p, MA_OWNED);
1553                 p->p_suspcount--;
1554                 if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) {
1555                         td->td_flags &= ~TDF_BOUNDARY;
1556                         p->p_boundary_count--;
1557                 }
1558         }
1559         return (setrunnable(td, 0));
1560 }
1561
1562 void
1563 thread_run_flash(struct thread *td)
1564 {
1565         struct proc *p;
1566
1567         p = td->td_proc;
1568         PROC_LOCK_ASSERT(p, MA_OWNED);
1569
1570         if (TD_ON_SLEEPQ(td))
1571                 sleepq_remove_nested(td);
1572         else
1573                 thread_lock(td);
1574
1575         THREAD_LOCK_ASSERT(td, MA_OWNED);
1576         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
1577
1578         TD_CLR_SUSPENDED(td);
1579         PROC_SLOCK(p);
1580         MPASS(p->p_suspcount > 0);
1581         p->p_suspcount--;
1582         PROC_SUNLOCK(p);
1583         if (setrunnable(td, 0))
1584                 kick_proc0();
1585 }
1586
1587 /*
1588  * Allow all threads blocked by single threading to continue running.
1589  */
1590 void
1591 thread_unsuspend(struct proc *p)
1592 {
1593         struct thread *td;
1594         int wakeup_swapper;
1595
1596         PROC_LOCK_ASSERT(p, MA_OWNED);
1597         PROC_SLOCK_ASSERT(p, MA_OWNED);
1598         wakeup_swapper = 0;
1599         if (!P_SHOULDSTOP(p)) {
1600                 FOREACH_THREAD_IN_PROC(p, td) {
1601                         thread_lock(td);
1602                         if (TD_IS_SUSPENDED(td) && (td->td_flags &
1603                             TDF_DOING_SA) == 0) {
1604                                 wakeup_swapper |= thread_unsuspend_one(td, p,
1605                                     true);
1606                         } else
1607                                 thread_unlock(td);
1608                 }
1609         } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
1610             p->p_numthreads == p->p_suspcount) {
1611                 /*
1612                  * Stopping everything also did the job for the single
1613                  * threading request. Now we've downgraded to single-threaded,
1614                  * let it continue.
1615                  */
1616                 if (p->p_singlethread->td_proc == p) {
1617                         thread_lock(p->p_singlethread);
1618                         wakeup_swapper = thread_unsuspend_one(
1619                             p->p_singlethread, p, false);
1620                 }
1621         }
1622         if (wakeup_swapper)
1623                 kick_proc0();
1624 }
1625
1626 /*
1627  * End the single threading mode..
1628  */
1629 void
1630 thread_single_end(struct proc *p, int mode)
1631 {
1632         struct thread *td;
1633         int wakeup_swapper;
1634
1635         KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY ||
1636             mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT,
1637             ("invalid mode %d", mode));
1638         PROC_LOCK_ASSERT(p, MA_OWNED);
1639         KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) ||
1640             (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0),
1641             ("mode %d does not match P_TOTAL_STOP", mode));
1642         KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread,
1643             ("thread_single_end from other thread %p %p",
1644             curthread, p->p_singlethread));
1645         KASSERT(mode != SINGLE_BOUNDARY ||
1646             (p->p_flag & P_SINGLE_BOUNDARY) != 0,
1647             ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag));
1648         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY |
1649             P_TOTAL_STOP);
1650         PROC_SLOCK(p);
1651         p->p_singlethread = NULL;
1652         wakeup_swapper = 0;
1653         /*
1654          * If there are other threads they may now run,
1655          * unless of course there is a blanket 'stop order'
1656          * on the process. The single threader must be allowed
1657          * to continue however as this is a bad place to stop.
1658          */
1659         if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) {
1660                 FOREACH_THREAD_IN_PROC(p, td) {
1661                         thread_lock(td);
1662                         if (TD_IS_SUSPENDED(td)) {
1663                                 wakeup_swapper |= thread_unsuspend_one(td, p,
1664                                     true);
1665                         } else
1666                                 thread_unlock(td);
1667                 }
1668         }
1669         KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0,
1670             ("inconsistent boundary count %d", p->p_boundary_count));
1671         PROC_SUNLOCK(p);
1672         if (wakeup_swapper)
1673                 kick_proc0();
1674         wakeup(&p->p_flag);
1675 }
1676
1677 /*
1678  * Locate a thread by number and return with proc lock held.
1679  *
1680  * thread exit establishes proc -> tidhash lock ordering, but lookup
1681  * takes tidhash first and needs to return locked proc.
1682  *
1683  * The problem is worked around by relying on type-safety of both
1684  * structures and doing the work in 2 steps:
1685  * - tidhash-locked lookup which saves both thread and proc pointers
1686  * - proc-locked verification that the found thread still matches
1687  */
1688 static bool
1689 tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp)
1690 {
1691 #define RUN_THRESH      16
1692         struct proc *p;
1693         struct thread *td;
1694         int run;
1695         bool locked;
1696
1697         run = 0;
1698         rw_rlock(TIDHASHLOCK(tid));
1699         locked = true;
1700         LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1701                 if (td->td_tid != tid) {
1702                         run++;
1703                         continue;
1704                 }
1705                 p = td->td_proc;
1706                 if (pid != -1 && p->p_pid != pid) {
1707                         td = NULL;
1708                         break;
1709                 }
1710                 if (run > RUN_THRESH) {
1711                         if (rw_try_upgrade(TIDHASHLOCK(tid))) {
1712                                 LIST_REMOVE(td, td_hash);
1713                                 LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1714                                         td, td_hash);
1715                                 rw_wunlock(TIDHASHLOCK(tid));
1716                                 locked = false;
1717                                 break;
1718                         }
1719                 }
1720                 break;
1721         }
1722         if (locked)
1723                 rw_runlock(TIDHASHLOCK(tid));
1724         if (td == NULL)
1725                 return (false);
1726         *pp = p;
1727         *tdp = td;
1728         return (true);
1729 }
1730
1731 struct thread *
1732 tdfind(lwpid_t tid, pid_t pid)
1733 {
1734         struct proc *p;
1735         struct thread *td;
1736
1737         td = curthread;
1738         if (td->td_tid == tid) {
1739                 if (pid != -1 && td->td_proc->p_pid != pid)
1740                         return (NULL);
1741                 PROC_LOCK(td->td_proc);
1742                 return (td);
1743         }
1744
1745         for (;;) {
1746                 if (!tdfind_hash(tid, pid, &p, &td))
1747                         return (NULL);
1748                 PROC_LOCK(p);
1749                 if (td->td_tid != tid) {
1750                         PROC_UNLOCK(p);
1751                         continue;
1752                 }
1753                 if (td->td_proc != p) {
1754                         PROC_UNLOCK(p);
1755                         continue;
1756                 }
1757                 if (p->p_state == PRS_NEW) {
1758                         PROC_UNLOCK(p);
1759                         return (NULL);
1760                 }
1761                 return (td);
1762         }
1763 }
1764
1765 void
1766 tidhash_add(struct thread *td)
1767 {
1768         rw_wlock(TIDHASHLOCK(td->td_tid));
1769         LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1770         rw_wunlock(TIDHASHLOCK(td->td_tid));
1771 }
1772
1773 void
1774 tidhash_remove(struct thread *td)
1775 {
1776
1777         rw_wlock(TIDHASHLOCK(td->td_tid));
1778         LIST_REMOVE(td, td_hash);
1779         rw_wunlock(TIDHASHLOCK(td->td_tid));
1780 }