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