sys/kern/kern_thread.c

   1 /*-
   2  * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
   3  *  All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice(s), this list of conditions and the following disclaimer as
  10  *    the first lines of this file unmodified other than the possible
  11  *    addition of one or more copyright notices.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice(s), this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  23  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  26  * DAMAGE.
  27  */
  28
  29 #include "opt_witness.h"
  30 #include "opt_kdtrace.h"
  31 #include "opt_hwpmc_hooks.h"
  32
  33 #include <sys/cdefs.h>
  34 __FBSDID("$FreeBSD$");
  35
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/kernel.h>
  39 #include <sys/lock.h>
  40 #include <sys/mutex.h>
  41 #include <sys/proc.h>
  42 #include <sys/rangelock.h>
  43 #include <sys/resourcevar.h>
  44 #include <sys/sdt.h>
  45 #include <sys/smp.h>
  46 #include <sys/sched.h>
  47 #include <sys/sleepqueue.h>
  48 #include <sys/selinfo.h>
  49 #include <sys/turnstile.h>
  50 #include <sys/ktr.h>
  51 #include <sys/rwlock.h>
  52 #include <sys/umtx.h>
  53 #include <sys/cpuset.h>
  54 #ifdef  HWPMC_HOOKS
  55 #include <sys/pmckern.h>
  56 #endif
  57
  58 #include <security/audit/audit.h>
  59
  60 #include <vm/vm.h>
  61 #include <vm/vm_extern.h>
  62 #include <vm/uma.h>
  63 #include <sys/eventhandler.h>
  64
  65 SDT_PROVIDER_DECLARE(proc);
  66 SDT_PROBE_DEFINE(proc, , , lwp__exit);
  67
  68
  69 /*
  70  * thread related storage.
  71  */
  72 static uma_zone_t thread_zone;
  73
  74 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
  75 static struct mtx zombie_lock;
  76 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
  77
  78 static void thread_zombie(struct thread *);
  79
  80 #define TID_BUFFER_SIZE 1024
  81
  82 struct mtx tid_lock;
  83 static struct unrhdr *tid_unrhdr;
  84 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
  85 static int tid_head, tid_tail;
  86 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
  87
  88 struct  tidhashhead *tidhashtbl;
  89 u_long  tidhash;
  90 struct  rwlock tidhash_lock;
  91
  92 static lwpid_t
  93 tid_alloc(void)
  94 {
  95         lwpid_t tid;
  96
  97         tid = alloc_unr(tid_unrhdr);
  98         if (tid != -1)
  99                 return (tid);
 100         mtx_lock(&tid_lock);
 101         if (tid_head == tid_tail) {
 102                 mtx_unlock(&tid_lock);
 103                 return (-1);
 104         }
 105         tid = tid_buffer[tid_head];
 106         tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
 107         mtx_unlock(&tid_lock);
 108         return (tid);
 109 }
 110
 111 static void
 112 tid_free(lwpid_t tid)
 113 {
 114         lwpid_t tmp_tid = -1;
 115
 116         mtx_lock(&tid_lock);
 117         if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
 118                 tmp_tid = tid_buffer[tid_head];
 119                 tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
 120         }
 121         tid_buffer[tid_tail] = tid;
 122         tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE;
 123         mtx_unlock(&tid_lock);
 124         if (tmp_tid != -1)
 125                 free_unr(tid_unrhdr, tmp_tid);
 126 }
 127
 128 /*
 129  * Prepare a thread for use.
 130  */
 131 static int
 132 thread_ctor(void *mem, int size, void *arg, int flags)
 133 {
 134         struct thread   *td;
 135
 136         td = (struct thread *)mem;
 137         td->td_state = TDS_INACTIVE;
 138         td->td_oncpu = NOCPU;
 139
 140         td->td_tid = tid_alloc();
 141
 142         /*
 143          * Note that td_critnest begins life as 1 because the thread is not
 144          * running and is thereby implicitly waiting to be on the receiving
 145          * end of a context switch.
 146          */
 147         td->td_critnest = 1;
 148         td->td_lend_user_pri = PRI_MAX;
 149         EVENTHANDLER_INVOKE(thread_ctor, td);
 150 #ifdef AUDIT
 151         audit_thread_alloc(td);
 152 #endif
 153         umtx_thread_alloc(td);
 154         return (0);
 155 }
 156
 157 /*
 158  * Reclaim a thread after use.
 159  */
 160 static void
 161 thread_dtor(void *mem, int size, void *arg)
 162 {
 163         struct thread *td;
 164
 165         td = (struct thread *)mem;
 166
 167 #ifdef INVARIANTS
 168         /* Verify that this thread is in a safe state to free. */
 169         switch (td->td_state) {
 170         case TDS_INHIBITED:
 171         case TDS_RUNNING:
 172         case TDS_CAN_RUN:
 173         case TDS_RUNQ:
 174                 /*
 175                  * We must never unlink a thread that is in one of
 176                  * these states, because it is currently active.
 177                  */
 178                 panic("bad state for thread unlinking");
 179                 /* NOTREACHED */
 180         case TDS_INACTIVE:
 181                 break;
 182         default:
 183                 panic("bad thread state");
 184                 /* NOTREACHED */
 185         }
 186 #endif
 187 #ifdef AUDIT
 188         audit_thread_free(td);
 189 #endif
 190         /* Free all OSD associated to this thread. */
 191         osd_thread_exit(td);
 192
 193         EVENTHANDLER_INVOKE(thread_dtor, td);
 194         tid_free(td->td_tid);
 195 }
 196
 197 /*
 198  * Initialize type-stable parts of a thread (when newly created).
 199  */
 200 static int
 201 thread_init(void *mem, int size, int flags)
 202 {
 203         struct thread *td;
 204
 205         td = (struct thread *)mem;
 206
 207         td->td_sleepqueue = sleepq_alloc();
 208         td->td_turnstile = turnstile_alloc();
 209         td->td_rlqe = NULL;
 210         td->td_vp_reserv = 0;
 211         EVENTHANDLER_INVOKE(thread_init, td);
 212         td->td_sched = (struct td_sched *)&td[1];
 213         umtx_thread_init(td);
 214         td->td_kstack = 0;
 215         td->td_sel = NULL;
 216         return (0);
 217 }
 218
 219 /*
 220  * Tear down type-stable parts of a thread (just before being discarded).
 221  */
 222 static void
 223 thread_fini(void *mem, int size)
 224 {
 225         struct thread *td;
 226
 227         td = (struct thread *)mem;
 228         EVENTHANDLER_INVOKE(thread_fini, td);
 229         rlqentry_free(td->td_rlqe);
 230         turnstile_free(td->td_turnstile);
 231         sleepq_free(td->td_sleepqueue);
 232         umtx_thread_fini(td);
 233         seltdfini(td);
 234 }
 235
 236 /*
 237  * For a newly created process,
 238  * link up all the structures and its initial threads etc.
 239  * called from:
 240  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
 241  * proc_dtor() (should go away)
 242  * proc_init()
 243  */
 244 void
 245 proc_linkup0(struct proc *p, struct thread *td)
 246 {
 247         TAILQ_INIT(&p->p_threads);           /* all threads in proc */
 248         proc_linkup(p, td);
 249 }
 250
 251 void
 252 proc_linkup(struct proc *p, struct thread *td)
 253 {
 254
 255         sigqueue_init(&p->p_sigqueue, p);
 256         p->p_ksi = ksiginfo_alloc(1);
 257         if (p->p_ksi != NULL) {
 258                 /* XXX p_ksi may be null if ksiginfo zone is not ready */
 259                 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
 260         }
 261         LIST_INIT(&p->p_mqnotifier);
 262         p->p_numthreads = 0;
 263         thread_link(td, p);
 264 }
 265
 266 /*
 267  * Initialize global thread allocation resources.
 268  */
 269 void
 270 threadinit(void)
 271 {
 272
 273         mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
 274
 275         /*
 276          * pid_max cannot be greater than PID_MAX.
 277          * leave one number for thread0.
 278          */
 279         tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
 280
 281         thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
 282             thread_ctor, thread_dtor, thread_init, thread_fini,
 283             16 - 1, 0);
 284         tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
 285         rw_init(&tidhash_lock, "tidhash");
 286 }
 287
 288 /*
 289  * Place an unused thread on the zombie list.
 290  * Use the slpq as that must be unused by now.
 291  */
 292 void
 293 thread_zombie(struct thread *td)
 294 {
 295         mtx_lock_spin(&zombie_lock);
 296         TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
 297         mtx_unlock_spin(&zombie_lock);
 298 }
 299
 300 /*
 301  * Release a thread that has exited after cpu_throw().
 302  */
 303 void
 304 thread_stash(struct thread *td)
 305 {
 306         atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
 307         thread_zombie(td);
 308 }
 309
 310 /*
 311  * Reap zombie resources.
 312  */
 313 void
 314 thread_reap(void)
 315 {
 316         struct thread *td_first, *td_next;
 317
 318         /*
 319          * Don't even bother to lock if none at this instant,
 320          * we really don't care about the next instant..
 321          */
 322         if (!TAILQ_EMPTY(&zombie_threads)) {
 323                 mtx_lock_spin(&zombie_lock);
 324                 td_first = TAILQ_FIRST(&zombie_threads);
 325                 if (td_first)
 326                         TAILQ_INIT(&zombie_threads);
 327                 mtx_unlock_spin(&zombie_lock);
 328                 while (td_first) {
 329                         td_next = TAILQ_NEXT(td_first, td_slpq);
 330                         if (td_first->td_ucred)
 331                                 crfree(td_first->td_ucred);
 332                         thread_free(td_first);
 333                         td_first = td_next;
 334                 }
 335         }
 336 }
 337
 338 /*
 339  * Allocate a thread.
 340  */
 341 struct thread *
 342 thread_alloc(int pages)
 343 {
 344         struct thread *td;
 345
 346         thread_reap(); /* check if any zombies to get */
 347
 348         td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
 349         KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
 350         if (!vm_thread_new(td, pages)) {
 351                 uma_zfree(thread_zone, td);
 352                 return (NULL);
 353         }
 354         cpu_thread_alloc(td);
 355         return (td);
 356 }
 357
 358 int
 359 thread_alloc_stack(struct thread *td, int pages)
 360 {
 361
 362         KASSERT(td->td_kstack == 0,
 363             ("thread_alloc_stack called on a thread with kstack"));
 364         if (!vm_thread_new(td, pages))
 365                 return (0);
 366         cpu_thread_alloc(td);
 367         return (1);
 368 }
 369
 370 /*
 371  * Deallocate a thread.
 372  */
 373 void
 374 thread_free(struct thread *td)
 375 {
 376
 377         lock_profile_thread_exit(td);
 378         if (td->td_cpuset)
 379                 cpuset_rel(td->td_cpuset);
 380         td->td_cpuset = NULL;
 381         cpu_thread_free(td);
 382         if (td->td_kstack != 0)
 383                 vm_thread_dispose(td);
 384         uma_zfree(thread_zone, td);
 385 }
 386
 387 /*
 388  * Discard the current thread and exit from its context.
 389  * Always called with scheduler locked.
 390  *
 391  * Because we can't free a thread while we're operating under its context,
 392  * push the current thread into our CPU's deadthread holder. This means
 393  * we needn't worry about someone else grabbing our context before we
 394  * do a cpu_throw().
 395  */
 396 void
 397 thread_exit(void)
 398 {
 399         uint64_t runtime, new_switchtime;
 400         struct thread *td;
 401         struct thread *td2;
 402         struct proc *p;
 403         int wakeup_swapper;
 404
 405         td = curthread;
 406         p = td->td_proc;
 407
 408         PROC_SLOCK_ASSERT(p, MA_OWNED);
 409         mtx_assert(&Giant, MA_NOTOWNED);
 410
 411         PROC_LOCK_ASSERT(p, MA_OWNED);
 412         KASSERT(p != NULL, ("thread exiting without a process"));
 413         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 414             (long)p->p_pid, td->td_name);
 415         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 416
 417 #ifdef AUDIT
 418         AUDIT_SYSCALL_EXIT(0, td);
 419 #endif
 420         umtx_thread_exit(td);
 421         /*
 422          * drop FPU & debug register state storage, or any other
 423          * architecture specific resources that
 424          * would not be on a new untouched process.
 425          */
 426         cpu_thread_exit(td);    /* XXXSMP */
 427
 428         /*
 429          * The last thread is left attached to the process
 430          * So that the whole bundle gets recycled. Skip
 431          * all this stuff if we never had threads.
 432          * EXIT clears all sign of other threads when
 433          * it goes to single threading, so the last thread always
 434          * takes the short path.
 435          */
 436         if (p->p_flag & P_HADTHREADS) {
 437                 if (p->p_numthreads > 1) {
 438                         thread_unlink(td);
 439                         td2 = FIRST_THREAD_IN_PROC(p);
 440                         sched_exit_thread(td2, td);
 441
 442                         /*
 443                          * The test below is NOT true if we are the
 444                          * sole exiting thread. P_STOPPED_SINGLE is unset
 445                          * in exit1() after it is the only survivor.
 446                          */
 447                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 448                                 if (p->p_numthreads == p->p_suspcount) {
 449                                         thread_lock(p->p_singlethread);
 450                                         wakeup_swapper = thread_unsuspend_one(
 451                                                 p->p_singlethread);
 452                                         thread_unlock(p->p_singlethread);
 453                                         if (wakeup_swapper)
 454                                                 kick_proc0();
 455                                 }
 456                         }
 457
 458                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 459                         PCPU_SET(deadthread, td);
 460                 } else {
 461                         /*
 462                          * The last thread is exiting.. but not through exit()
 463                          */
 464                         panic ("thread_exit: Last thread exiting on its own");
 465                 }
 466         }
 467 #ifdef  HWPMC_HOOKS
 468         /*
 469          * If this thread is part of a process that is being tracked by hwpmc(4),
 470          * inform the module of the thread's impending exit.
 471          */
 472         if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 473                 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
 474 #endif
 475         PROC_UNLOCK(p);
 476
 477         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 478         new_switchtime = cpu_ticks();
 479         runtime = new_switchtime - PCPU_GET(switchtime);
 480         td->td_runtime += runtime;
 481         td->td_incruntime += runtime;
 482         PCPU_SET(switchtime, new_switchtime);
 483         PCPU_SET(switchticks, ticks);
 484         PCPU_INC(cnt.v_swtch);
 485
 486         /* Save our resource usage in our process. */
 487         td->td_ru.ru_nvcsw++;
 488         ruxagg(p, td);
 489         rucollect(&p->p_ru, &td->td_ru);
 490
 491         thread_lock(td);
 492         PROC_SUNLOCK(p);
 493         td->td_state = TDS_INACTIVE;
 494 #ifdef WITNESS
 495         witness_thread_exit(td);
 496 #endif
 497         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 498         sched_throw(td);
 499         panic("I'm a teapot!");
 500         /* NOTREACHED */
 501 }
 502
 503 /*
 504  * Do any thread specific cleanups that may be needed in wait()
 505  * called with Giant, proc and schedlock not held.
 506  */
 507 void
 508 thread_wait(struct proc *p)
 509 {
 510         struct thread *td;
 511
 512         mtx_assert(&Giant, MA_NOTOWNED);
 513         KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
 514         td = FIRST_THREAD_IN_PROC(p);
 515         /* Lock the last thread so we spin until it exits cpu_throw(). */
 516         thread_lock(td);
 517         thread_unlock(td);
 518         /* Wait for any remaining threads to exit cpu_throw(). */
 519         while (p->p_exitthreads)
 520                 sched_relinquish(curthread);
 521         lock_profile_thread_exit(td);
 522         cpuset_rel(td->td_cpuset);
 523         td->td_cpuset = NULL;
 524         cpu_thread_clean(td);
 525         crfree(td->td_ucred);
 526         thread_reap();  /* check for zombie threads etc. */
 527 }
 528
 529 /*
 530  * Link a thread to a process.
 531  * set up anything that needs to be initialized for it to
 532  * be used by the process.
 533  */
 534 void
 535 thread_link(struct thread *td, struct proc *p)
 536 {
 537
 538         /*
 539          * XXX This can't be enabled because it's called for proc0 before
 540          * its lock has been created.
 541          * PROC_LOCK_ASSERT(p, MA_OWNED);
 542          */
 543         td->td_state    = TDS_INACTIVE;
 544         td->td_proc     = p;
 545         td->td_flags    = TDF_INMEM;
 546
 547         LIST_INIT(&td->td_contested);
 548         LIST_INIT(&td->td_lprof[0]);
 549         LIST_INIT(&td->td_lprof[1]);
 550         sigqueue_init(&td->td_sigqueue, p);
 551         callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
 552         TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
 553         p->p_numthreads++;
 554 }
 555
 556 /*
 557  * Convert a process with one thread to an unthreaded process.
 558  */
 559 void
 560 thread_unthread(struct thread *td)
 561 {
 562         struct proc *p = td->td_proc;
 563
 564         KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
 565         p->p_flag &= ~P_HADTHREADS;
 566 }
 567
 568 /*
 569  * Called from:
 570  *  thread_exit()
 571  */
 572 void
 573 thread_unlink(struct thread *td)
 574 {
 575         struct proc *p = td->td_proc;
 576
 577         PROC_LOCK_ASSERT(p, MA_OWNED);
 578         TAILQ_REMOVE(&p->p_threads, td, td_plist);
 579         p->p_numthreads--;
 580         /* could clear a few other things here */
 581         /* Must  NOT clear links to proc! */
 582 }
 583
 584 static int
 585 calc_remaining(struct proc *p, int mode)
 586 {
 587         int remaining;
 588
 589         PROC_LOCK_ASSERT(p, MA_OWNED);
 590         PROC_SLOCK_ASSERT(p, MA_OWNED);
 591         if (mode == SINGLE_EXIT)
 592                 remaining = p->p_numthreads;
 593         else if (mode == SINGLE_BOUNDARY)
 594                 remaining = p->p_numthreads - p->p_boundary_count;
 595         else if (mode == SINGLE_NO_EXIT)
 596                 remaining = p->p_numthreads - p->p_suspcount;
 597         else
 598                 panic("calc_remaining: wrong mode %d", mode);
 599         return (remaining);
 600 }
 601
 602 /*
 603  * Enforce single-threading.
 604  *
 605  * Returns 1 if the caller must abort (another thread is waiting to
 606  * exit the process or similar). Process is locked!
 607  * Returns 0 when you are successfully the only thread running.
 608  * A process has successfully single threaded in the suspend mode when
 609  * There are no threads in user mode. Threads in the kernel must be
 610  * allowed to continue until they get to the user boundary. They may even
 611  * copy out their return values and data before suspending. They may however be
 612  * accelerated in reaching the user boundary as we will wake up
 613  * any sleeping threads that are interruptable. (PCATCH).
 614  */
 615 int
 616 thread_single(int mode)
 617 {
 618         struct thread *td;
 619         struct thread *td2;
 620         struct proc *p;
 621         int remaining, wakeup_swapper;
 622
 623         td = curthread;
 624         p = td->td_proc;
 625         mtx_assert(&Giant, MA_NOTOWNED);
 626         PROC_LOCK_ASSERT(p, MA_OWNED);
 627
 628         if ((p->p_flag & P_HADTHREADS) == 0)
 629                 return (0);
 630
 631         /* Is someone already single threading? */
 632         if (p->p_singlethread != NULL && p->p_singlethread != td)
 633                 return (1);
 634
 635         if (mode == SINGLE_EXIT) {
 636                 p->p_flag |= P_SINGLE_EXIT;
 637                 p->p_flag &= ~P_SINGLE_BOUNDARY;
 638         } else {
 639                 p->p_flag &= ~P_SINGLE_EXIT;
 640                 if (mode == SINGLE_BOUNDARY)
 641                         p->p_flag |= P_SINGLE_BOUNDARY;
 642                 else
 643                         p->p_flag &= ~P_SINGLE_BOUNDARY;
 644         }
 645         p->p_flag |= P_STOPPED_SINGLE;
 646         PROC_SLOCK(p);
 647         p->p_singlethread = td;
 648         remaining = calc_remaining(p, mode);
 649         while (remaining != 1) {
 650                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
 651                         goto stopme;
 652                 wakeup_swapper = 0;
 653                 FOREACH_THREAD_IN_PROC(p, td2) {
 654                         if (td2 == td)
 655                                 continue;
 656                         thread_lock(td2);
 657                         td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
 658                         if (TD_IS_INHIBITED(td2)) {
 659                                 switch (mode) {
 660                                 case SINGLE_EXIT:
 661                                         if (TD_IS_SUSPENDED(td2))
 662                                                 wakeup_swapper |=
 663                                                     thread_unsuspend_one(td2);
 664                                         if (TD_ON_SLEEPQ(td2) &&
 665                                             (td2->td_flags & TDF_SINTR))
 666                                                 wakeup_swapper |=
 667                                                     sleepq_abort(td2, EINTR);
 668                                         break;
 669                                 case SINGLE_BOUNDARY:
 670                                         if (TD_IS_SUSPENDED(td2) &&
 671                                             !(td2->td_flags & TDF_BOUNDARY))
 672                                                 wakeup_swapper |=
 673                                                     thread_unsuspend_one(td2);
 674                                         if (TD_ON_SLEEPQ(td2) &&
 675                                             (td2->td_flags & TDF_SINTR))
 676                                                 wakeup_swapper |=
 677                                                     sleepq_abort(td2, ERESTART);
 678                                         break;
 679                                 case SINGLE_NO_EXIT:
 680                                         if (TD_IS_SUSPENDED(td2) &&
 681                                             !(td2->td_flags & TDF_BOUNDARY))
 682                                                 wakeup_swapper |=
 683                                                     thread_unsuspend_one(td2);
 684                                         if (TD_ON_SLEEPQ(td2) &&
 685                                             (td2->td_flags & TDF_SINTR))
 686                                                 wakeup_swapper |=
 687                                                     sleepq_abort(td2, ERESTART);
 688                                         break;
 689                                 default:
 690                                         break;
 691                                 }
 692                         }
 693 #ifdef SMP
 694                         else if (TD_IS_RUNNING(td2) && td != td2) {
 695                                 forward_signal(td2);
 696                         }
 697 #endif
 698                         thread_unlock(td2);
 699                 }
 700                 if (wakeup_swapper)
 701                         kick_proc0();
 702                 remaining = calc_remaining(p, mode);
 703
 704                 /*
 705                  * Maybe we suspended some threads.. was it enough?
 706                  */
 707                 if (remaining == 1)
 708                         break;
 709
 710 stopme:
 711                 /*
 712                  * Wake us up when everyone else has suspended.
 713                  * In the mean time we suspend as well.
 714                  */
 715                 thread_suspend_switch(td);
 716                 remaining = calc_remaining(p, mode);
 717         }
 718         if (mode == SINGLE_EXIT) {
 719                 /*
 720                  * We have gotten rid of all the other threads and we
 721                  * are about to either exit or exec. In either case,
 722                  * we try our utmost  to revert to being a non-threaded
 723                  * process.
 724                  */
 725                 p->p_singlethread = NULL;
 726                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
 727                 thread_unthread(td);
 728         }
 729         PROC_SUNLOCK(p);
 730         return (0);
 731 }
 732
 733 /*
 734  * Called in from locations that can safely check to see
 735  * whether we have to suspend or at least throttle for a
 736  * single-thread event (e.g. fork).
 737  *
 738  * Such locations include userret().
 739  * If the "return_instead" argument is non zero, the thread must be able to
 740  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
 741  *
 742  * The 'return_instead' argument tells the function if it may do a
 743  * thread_exit() or suspend, or whether the caller must abort and back
 744  * out instead.
 745  *
 746  * If the thread that set the single_threading request has set the
 747  * P_SINGLE_EXIT bit in the process flags then this call will never return
 748  * if 'return_instead' is false, but will exit.
 749  *
 750  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
 751  *---------------+--------------------+---------------------
 752  *       0       | returns 0          |   returns 0 or 1
 753  *               | when ST ends       |   immediatly
 754  *---------------+--------------------+---------------------
 755  *       1       | thread exits       |   returns 1
 756  *               |                    |  immediatly
 757  * 0 = thread_exit() or suspension ok,
 758  * other = return error instead of stopping the thread.
 759  *
 760  * While a full suspension is under effect, even a single threading
 761  * thread would be suspended if it made this call (but it shouldn't).
 762  * This call should only be made from places where
 763  * thread_exit() would be safe as that may be the outcome unless
 764  * return_instead is set.
 765  */
 766 int
 767 thread_suspend_check(int return_instead)
 768 {
 769         struct thread *td;
 770         struct proc *p;
 771         int wakeup_swapper;
 772
 773         td = curthread;
 774         p = td->td_proc;
 775         mtx_assert(&Giant, MA_NOTOWNED);
 776         PROC_LOCK_ASSERT(p, MA_OWNED);
 777         while (P_SHOULDSTOP(p) ||
 778               ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
 779                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 780                         KASSERT(p->p_singlethread != NULL,
 781                             ("singlethread not set"));
 782                         /*
 783                          * The only suspension in action is a
 784                          * single-threading. Single threader need not stop.
 785                          * XXX Should be safe to access unlocked
 786                          * as it can only be set to be true by us.
 787                          */
 788                         if (p->p_singlethread == td)
 789                                 return (0);     /* Exempt from stopping. */
 790                 }
 791                 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
 792                         return (EINTR);
 793
 794                 /* Should we goto user boundary if we didn't come from there? */
 795                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
 796                     (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
 797                         return (ERESTART);
 798
 799                 /*
 800                  * Ignore suspend requests for stop signals if they
 801                  * are deferred.
 802                  */
 803                 if (P_SHOULDSTOP(p) == P_STOPPED_SIG &&
 804                     td->td_flags & TDF_SBDRY) {
 805                         KASSERT(return_instead,
 806                             ("TDF_SBDRY set for unsafe thread_suspend_check"));
 807                         return (0);
 808                 }
 809
 810                 /*
 811                  * If the process is waiting for us to exit,
 812                  * this thread should just suicide.
 813                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
 814                  */
 815                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
 816                         PROC_UNLOCK(p);
 817                         tidhash_remove(td);
 818                         PROC_LOCK(p);
 819                         tdsigcleanup(td);
 820                         PROC_SLOCK(p);
 821                         thread_stopped(p);
 822                         thread_exit();
 823                 }
 824
 825                 PROC_SLOCK(p);
 826                 thread_stopped(p);
 827                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 828                         if (p->p_numthreads == p->p_suspcount + 1) {
 829                                 thread_lock(p->p_singlethread);
 830                                 wakeup_swapper =
 831                                     thread_unsuspend_one(p->p_singlethread);
 832                                 thread_unlock(p->p_singlethread);
 833                                 if (wakeup_swapper)
 834                                         kick_proc0();
 835                         }
 836                 }
 837                 PROC_UNLOCK(p);
 838                 thread_lock(td);
 839                 /*
 840                  * When a thread suspends, it just
 841                  * gets taken off all queues.
 842                  */
 843                 thread_suspend_one(td);
 844                 if (return_instead == 0) {
 845                         p->p_boundary_count++;
 846                         td->td_flags |= TDF_BOUNDARY;
 847                 }
 848                 PROC_SUNLOCK(p);
 849                 mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
 850                 if (return_instead == 0)
 851                         td->td_flags &= ~TDF_BOUNDARY;
 852                 thread_unlock(td);
 853                 PROC_LOCK(p);
 854                 if (return_instead == 0) {
 855                         PROC_SLOCK(p);
 856                         p->p_boundary_count--;
 857                         PROC_SUNLOCK(p);
 858                 }
 859         }
 860         return (0);
 861 }
 862
 863 void
 864 thread_suspend_switch(struct thread *td)
 865 {
 866         struct proc *p;
 867
 868         p = td->td_proc;
 869         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 870         PROC_LOCK_ASSERT(p, MA_OWNED);
 871         PROC_SLOCK_ASSERT(p, MA_OWNED);
 872         /*
 873          * We implement thread_suspend_one in stages here to avoid
 874          * dropping the proc lock while the thread lock is owned.
 875          */
 876         thread_stopped(p);
 877         p->p_suspcount++;
 878         PROC_UNLOCK(p);
 879         thread_lock(td);
 880         td->td_flags &= ~TDF_NEEDSUSPCHK;
 881         TD_SET_SUSPENDED(td);
 882         sched_sleep(td, 0);
 883         PROC_SUNLOCK(p);
 884         DROP_GIANT();
 885         mi_switch(SW_VOL | SWT_SUSPEND, NULL);
 886         thread_unlock(td);
 887         PICKUP_GIANT();
 888         PROC_LOCK(p);
 889         PROC_SLOCK(p);
 890 }
 891
 892 void
 893 thread_suspend_one(struct thread *td)
 894 {
 895         struct proc *p = td->td_proc;
 896
 897         PROC_SLOCK_ASSERT(p, MA_OWNED);
 898         THREAD_LOCK_ASSERT(td, MA_OWNED);
 899         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 900         p->p_suspcount++;
 901         td->td_flags &= ~TDF_NEEDSUSPCHK;
 902         TD_SET_SUSPENDED(td);
 903         sched_sleep(td, 0);
 904 }
 905
 906 int
 907 thread_unsuspend_one(struct thread *td)
 908 {
 909         struct proc *p = td->td_proc;
 910
 911         PROC_SLOCK_ASSERT(p, MA_OWNED);
 912         THREAD_LOCK_ASSERT(td, MA_OWNED);
 913         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
 914         TD_CLR_SUSPENDED(td);
 915         p->p_suspcount--;
 916         return (setrunnable(td));
 917 }
 918
 919 /*
 920  * Allow all threads blocked by single threading to continue running.
 921  */
 922 void
 923 thread_unsuspend(struct proc *p)
 924 {
 925         struct thread *td;
 926         int wakeup_swapper;
 927
 928         PROC_LOCK_ASSERT(p, MA_OWNED);
 929         PROC_SLOCK_ASSERT(p, MA_OWNED);
 930         wakeup_swapper = 0;
 931         if (!P_SHOULDSTOP(p)) {
 932                 FOREACH_THREAD_IN_PROC(p, td) {
 933                         thread_lock(td);
 934                         if (TD_IS_SUSPENDED(td)) {
 935                                 wakeup_swapper |= thread_unsuspend_one(td);
 936                         }
 937                         thread_unlock(td);
 938                 }
 939         } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
 940             (p->p_numthreads == p->p_suspcount)) {
 941                 /*
 942                  * Stopping everything also did the job for the single
 943                  * threading request. Now we've downgraded to single-threaded,
 944                  * let it continue.
 945                  */
 946                 thread_lock(p->p_singlethread);
 947                 wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
 948                 thread_unlock(p->p_singlethread);
 949         }
 950         if (wakeup_swapper)
 951                 kick_proc0();
 952 }
 953
 954 /*
 955  * End the single threading mode..
 956  */
 957 void
 958 thread_single_end(void)
 959 {
 960         struct thread *td;
 961         struct proc *p;
 962         int wakeup_swapper;
 963
 964         td = curthread;
 965         p = td->td_proc;
 966         PROC_LOCK_ASSERT(p, MA_OWNED);
 967         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
 968         PROC_SLOCK(p);
 969         p->p_singlethread = NULL;
 970         wakeup_swapper = 0;
 971         /*
 972          * If there are other threads they may now run,
 973          * unless of course there is a blanket 'stop order'
 974          * on the process. The single threader must be allowed
 975          * to continue however as this is a bad place to stop.
 976          */
 977         if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
 978                 FOREACH_THREAD_IN_PROC(p, td) {
 979                         thread_lock(td);
 980                         if (TD_IS_SUSPENDED(td)) {
 981                                 wakeup_swapper |= thread_unsuspend_one(td);
 982                         }
 983                         thread_unlock(td);
 984                 }
 985         }
 986         PROC_SUNLOCK(p);
 987         if (wakeup_swapper)
 988                 kick_proc0();
 989 }
 990
 991 struct thread *
 992 thread_find(struct proc *p, lwpid_t tid)
 993 {
 994         struct thread *td;
 995
 996         PROC_LOCK_ASSERT(p, MA_OWNED);
 997         FOREACH_THREAD_IN_PROC(p, td) {
 998                 if (td->td_tid == tid)
 999                         break;
1000         }
1001         return (td);
1002 }
1003
1004 /* Locate a thread by number; return with proc lock held. */
1005 struct thread *
1006 tdfind(lwpid_t tid, pid_t pid)
1007 {
1008 #define RUN_THRESH      16
1009         struct thread *td;
1010         int run = 0;
1011
1012         rw_rlock(&tidhash_lock);
1013         LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1014                 if (td->td_tid == tid) {
1015                         if (pid != -1 && td->td_proc->p_pid != pid) {
1016                                 td = NULL;
1017                                 break;
1018                         }
1019                         PROC_LOCK(td->td_proc);
1020                         if (td->td_proc->p_state == PRS_NEW) {
1021                                 PROC_UNLOCK(td->td_proc);
1022                                 td = NULL;
1023                                 break;
1024                         }
1025                         if (run > RUN_THRESH) {
1026                                 if (rw_try_upgrade(&tidhash_lock)) {
1027                                         LIST_REMOVE(td, td_hash);
1028                                         LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1029                                                 td, td_hash);
1030                                         rw_wunlock(&tidhash_lock);
1031                                         return (td);
1032                                 }
1033                         }
1034                         break;
1035                 }
1036                 run++;
1037         }
1038         rw_runlock(&tidhash_lock);
1039         return (td);
1040 }
1041
1042 void
1043 tidhash_add(struct thread *td)
1044 {
1045         rw_wlock(&tidhash_lock);
1046         LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1047         rw_wunlock(&tidhash_lock);
1048 }
1049
1050 void
1051 tidhash_remove(struct thread *td)
1052 {
1053         rw_wlock(&tidhash_lock);
1054         LIST_REMOVE(td, td_hash);
1055         rw_wunlock(&tidhash_lock);
1056 }