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 <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/systm.h>
  34 #include <sys/kernel.h>
  35 #include <sys/lock.h>
  36 #include <sys/mutex.h>
  37 #include <sys/proc.h>
  38 #include <sys/resourcevar.h>
  39 #include <sys/smp.h>
  40 #include <sys/sysctl.h>
  41 #include <sys/sched.h>
  42 #include <sys/sleepqueue.h>
  43 #include <sys/turnstile.h>
  44 #include <sys/ktr.h>
  45 #include <sys/umtx.h>
  46
  47 #include <security/audit/audit.h>
  48
  49 #include <vm/vm.h>
  50 #include <vm/vm_extern.h>
  51 #include <vm/uma.h>
  52
  53 /*
  54  * thread related storage.
  55  */
  56 static uma_zone_t thread_zone;
  57
  58 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
  59
  60 int max_threads_per_proc = 1500;
  61 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
  62         &max_threads_per_proc, 0, "Limit on threads per proc");
  63
  64 int max_threads_hits;
  65 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
  66         &max_threads_hits, 0, "");
  67
  68 #ifdef KSE
  69 int virtual_cpu;
  70
  71 #endif
  72 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
  73 struct mtx zombie_lock;
  74 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
  75
  76 static void thread_zombie(struct thread *);
  77
  78 #ifdef KSE
  79 static int
  80 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
  81 {
  82         int error, new_val;
  83         int def_val;
  84
  85         def_val = mp_ncpus;
  86         if (virtual_cpu == 0)
  87                 new_val = def_val;
  88         else
  89                 new_val = virtual_cpu;
  90         error = sysctl_handle_int(oidp, &new_val, 0, req);
  91         if (error != 0 || req->newptr == NULL)
  92                 return (error);
  93         if (new_val < 0)
  94                 return (EINVAL);
  95         virtual_cpu = new_val;
  96         return (0);
  97 }
  98
  99 /* DEBUG ONLY */
 100 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
 101         0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
 102         "debug virtual cpus");
 103 #endif
 104
 105 struct mtx tid_lock;
 106 static struct unrhdr *tid_unrhdr;
 107
 108 /*
 109  * Prepare a thread for use.
 110  */
 111 static int
 112 thread_ctor(void *mem, int size, void *arg, int flags)
 113 {
 114         struct thread   *td;
 115
 116         td = (struct thread *)mem;
 117         td->td_state = TDS_INACTIVE;
 118         td->td_oncpu = NOCPU;
 119
 120         td->td_tid = alloc_unr(tid_unrhdr);
 121         td->td_syscalls = 0;
 122
 123         /*
 124          * Note that td_critnest begins life as 1 because the thread is not
 125          * running and is thereby implicitly waiting to be on the receiving
 126          * end of a context switch.
 127          */
 128         td->td_critnest = 1;
 129
 130 #ifdef AUDIT
 131         audit_thread_alloc(td);
 132 #endif
 133         umtx_thread_alloc(td);
 134         return (0);
 135 }
 136
 137 /*
 138  * Reclaim a thread after use.
 139  */
 140 static void
 141 thread_dtor(void *mem, int size, void *arg)
 142 {
 143         struct thread *td;
 144
 145         td = (struct thread *)mem;
 146
 147 #ifdef INVARIANTS
 148         /* Verify that this thread is in a safe state to free. */
 149         switch (td->td_state) {
 150         case TDS_INHIBITED:
 151         case TDS_RUNNING:
 152         case TDS_CAN_RUN:
 153         case TDS_RUNQ:
 154                 /*
 155                  * We must never unlink a thread that is in one of
 156                  * these states, because it is currently active.
 157                  */
 158                 panic("bad state for thread unlinking");
 159                 /* NOTREACHED */
 160         case TDS_INACTIVE:
 161                 break;
 162         default:
 163                 panic("bad thread state");
 164                 /* NOTREACHED */
 165         }
 166 #endif
 167 #ifdef AUDIT
 168         audit_thread_free(td);
 169 #endif
 170         free_unr(tid_unrhdr, td->td_tid);
 171         sched_newthread(td);
 172 }
 173
 174 /*
 175  * Initialize type-stable parts of a thread (when newly created).
 176  */
 177 static int
 178 thread_init(void *mem, int size, int flags)
 179 {
 180         struct thread *td;
 181
 182         td = (struct thread *)mem;
 183
 184         vm_thread_new(td, 0);
 185         cpu_thread_setup(td);
 186         td->td_sleepqueue = sleepq_alloc();
 187         td->td_turnstile = turnstile_alloc();
 188         td->td_sched = (struct td_sched *)&td[1];
 189         sched_newthread(td);
 190         umtx_thread_init(td);
 191         return (0);
 192 }
 193
 194 /*
 195  * Tear down type-stable parts of a thread (just before being discarded).
 196  */
 197 static void
 198 thread_fini(void *mem, int size)
 199 {
 200         struct thread *td;
 201
 202         td = (struct thread *)mem;
 203         turnstile_free(td->td_turnstile);
 204         sleepq_free(td->td_sleepqueue);
 205         umtx_thread_fini(td);
 206         vm_thread_dispose(td);
 207 }
 208
 209 /*
 210  * For a newly created process,
 211  * link up all the structures and its initial threads etc.
 212  * called from:
 213  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
 214  * proc_dtor() (should go away)
 215  * proc_init()
 216  */
 217 void
 218 proc_linkup(struct proc *p, struct thread *td)
 219 {
 220
 221         TAILQ_INIT(&p->p_threads);           /* all threads in proc */
 222 #ifdef KSE
 223         TAILQ_INIT(&p->p_upcalls);           /* upcall list */
 224 #endif
 225         sigqueue_init(&p->p_sigqueue, p);
 226         p->p_ksi = ksiginfo_alloc(1);
 227         if (p->p_ksi != NULL) {
 228                 /* XXX p_ksi may be null if ksiginfo zone is not ready */
 229                 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
 230         }
 231         LIST_INIT(&p->p_mqnotifier);
 232         p->p_numthreads = 0;
 233         thread_link(td, p);
 234 }
 235
 236 /*
 237  * Initialize global thread allocation resources.
 238  */
 239 void
 240 threadinit(void)
 241 {
 242
 243         mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
 244         tid_unrhdr = new_unrhdr(PID_MAX + 1, INT_MAX, &tid_lock);
 245
 246         thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
 247             thread_ctor, thread_dtor, thread_init, thread_fini,
 248             16 - 1, 0);
 249 #ifdef KSE
 250         kseinit();      /* set up kse specific stuff  e.g. upcall zone*/
 251 #endif
 252 }
 253
 254 /*
 255  * Place an unused thread on the zombie list.
 256  * Use the slpq as that must be unused by now.
 257  */
 258 void
 259 thread_zombie(struct thread *td)
 260 {
 261         mtx_lock_spin(&zombie_lock);
 262         TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
 263         mtx_unlock_spin(&zombie_lock);
 264 }
 265
 266 /*
 267  * Release a thread that has exited after cpu_throw().
 268  */
 269 void
 270 thread_stash(struct thread *td)
 271 {
 272         atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
 273         thread_zombie(td);
 274 }
 275
 276 /*
 277  * Reap zombie kse resource.
 278  */
 279 void
 280 thread_reap(void)
 281 {
 282         struct thread *td_first, *td_next;
 283
 284         /*
 285          * Don't even bother to lock if none at this instant,
 286          * we really don't care about the next instant..
 287          */
 288         if (!TAILQ_EMPTY(&zombie_threads)) {
 289                 mtx_lock_spin(&zombie_lock);
 290                 td_first = TAILQ_FIRST(&zombie_threads);
 291                 if (td_first)
 292                         TAILQ_INIT(&zombie_threads);
 293                 mtx_unlock_spin(&zombie_lock);
 294                 while (td_first) {
 295                         td_next = TAILQ_NEXT(td_first, td_slpq);
 296                         if (td_first->td_ucred)
 297                                 crfree(td_first->td_ucred);
 298                         thread_free(td_first);
 299                         td_first = td_next;
 300                 }
 301         }
 302 #ifdef KSE
 303         upcall_reap();
 304 #endif
 305 }
 306
 307 /*
 308  * Allocate a thread.
 309  */
 310 struct thread *
 311 thread_alloc(void)
 312 {
 313
 314         thread_reap(); /* check if any zombies to get */
 315         return (uma_zalloc(thread_zone, M_WAITOK));
 316 }
 317
 318
 319 /*
 320  * Deallocate a thread.
 321  */
 322 void
 323 thread_free(struct thread *td)
 324 {
 325
 326         cpu_thread_clean(td);
 327         uma_zfree(thread_zone, td);
 328 }
 329
 330 /*
 331  * Discard the current thread and exit from its context.
 332  * Always called with scheduler locked.
 333  *
 334  * Because we can't free a thread while we're operating under its context,
 335  * push the current thread into our CPU's deadthread holder. This means
 336  * we needn't worry about someone else grabbing our context before we
 337  * do a cpu_throw().  This may not be needed now as we are under schedlock.
 338  * Maybe we can just do a thread_stash() as thr_exit1 does.
 339  */
 340 /*  XXX
 341  * libthr expects its thread exit to return for the last
 342  * thread, meaning that the program is back to non-threaded
 343  * mode I guess. Because we do this (cpu_throw) unconditionally
 344  * here, they have their own version of it. (thr_exit1())
 345  * that doesn't do it all if this was the last thread.
 346  * It is also called from thread_suspend_check().
 347  * Of course in the end, they end up coming here through exit1
 348  * anyhow..  After fixing 'thr' to play by the rules we should be able
 349  * to merge these two functions together.
 350  *
 351  * called from:
 352  * exit1()
 353  * kse_exit()
 354  * thr_exit()
 355  * ifdef KSE
 356  * thread_user_enter()
 357  * thread_userret()
 358  * endif
 359  * thread_suspend_check()
 360  */
 361 void
 362 thread_exit(void)
 363 {
 364         uint64_t new_switchtime;
 365         struct thread *td;
 366         struct thread *td2;
 367         struct proc *p;
 368
 369         td = curthread;
 370         p = td->td_proc;
 371
 372         PROC_SLOCK_ASSERT(p, MA_OWNED);
 373         mtx_assert(&Giant, MA_NOTOWNED);
 374
 375         PROC_LOCK_ASSERT(p, MA_OWNED);
 376         KASSERT(p != NULL, ("thread exiting without a process"));
 377         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 378             (long)p->p_pid, p->p_comm);
 379         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 380
 381 #ifdef AUDIT
 382         AUDIT_SYSCALL_EXIT(0, td);
 383 #endif
 384
 385 #ifdef KSE
 386         if (td->td_standin != NULL) {
 387                 /*
 388                  * Note that we don't need to free the cred here as it
 389                  * is done in thread_reap().
 390                  */
 391                 thread_zombie(td->td_standin);
 392                 td->td_standin = NULL;
 393         }
 394 #endif
 395
 396         umtx_thread_exit(td);
 397
 398         /*
 399          * drop FPU & debug register state storage, or any other
 400          * architecture specific resources that
 401          * would not be on a new untouched process.
 402          */
 403         cpu_thread_exit(td);    /* XXXSMP */
 404
 405         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 406         new_switchtime = cpu_ticks();
 407         p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
 408         PCPU_SET(switchtime, new_switchtime);
 409         PCPU_SET(switchticks, ticks);
 410         PCPU_INC(cnt.v_swtch);
 411         /* Save our resource usage in our process. */
 412         td->td_ru.ru_nvcsw++;
 413         rucollect(&p->p_ru, &td->td_ru);
 414         /*
 415          * The last thread is left attached to the process
 416          * So that the whole bundle gets recycled. Skip
 417          * all this stuff if we never had threads.
 418          * EXIT clears all sign of other threads when
 419          * it goes to single threading, so the last thread always
 420          * takes the short path.
 421          */
 422         if (p->p_flag & P_HADTHREADS) {
 423                 if (p->p_numthreads > 1) {
 424                         thread_lock(td);
 425 #ifdef KSE
 426                         kse_unlink(td);
 427 #else
 428                         thread_unlink(td);
 429 #endif
 430                         thread_unlock(td);
 431                         td2 = FIRST_THREAD_IN_PROC(p);
 432                         sched_exit_thread(td2, td);
 433
 434                         /*
 435                          * The test below is NOT true if we are the
 436                          * sole exiting thread. P_STOPPED_SNGL is unset
 437                          * in exit1() after it is the only survivor.
 438                          */
 439                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 440                                 if (p->p_numthreads == p->p_suspcount) {
 441                                         thread_lock(p->p_singlethread);
 442                                         thread_unsuspend_one(p->p_singlethread);
 443                                         thread_unlock(p->p_singlethread);
 444                                 }
 445                         }
 446
 447                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 448                         PCPU_SET(deadthread, td);
 449                 } else {
 450                         /*
 451                          * The last thread is exiting.. but not through exit()
 452                          * what should we do?
 453                          * Theoretically this can't happen
 454                          * exit1() - clears threading flags before coming here
 455                          * kse_exit() - treats last thread specially
 456                          * thr_exit() - treats last thread specially
 457                          * ifdef KSE
 458                          * thread_user_enter() - only if more exist
 459                          * thread_userret() - only if more exist
 460                          * endif
 461                          * thread_suspend_check() - only if more exist
 462                          */
 463                         panic ("thread_exit: Last thread exiting on its own");
 464                 }
 465         }
 466         PROC_UNLOCK(p);
 467         thread_lock(td);
 468         /* Save our tick information with both the thread and proc locked */
 469         ruxagg(&p->p_rux, td);
 470         PROC_SUNLOCK(p);
 471         td->td_state = TDS_INACTIVE;
 472         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 473         sched_throw(td);
 474         panic("I'm a teapot!");
 475         /* NOTREACHED */
 476 }
 477
 478 /*
 479  * Do any thread specific cleanups that may be needed in wait()
 480  * called with Giant, proc and schedlock not held.
 481  */
 482 void
 483 thread_wait(struct proc *p)
 484 {
 485         struct thread *td;
 486
 487         mtx_assert(&Giant, MA_NOTOWNED);
 488         KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
 489         td = FIRST_THREAD_IN_PROC(p);
 490 #ifdef KSE
 491         if (td->td_standin != NULL) {
 492                 if (td->td_standin->td_ucred != NULL) {
 493                         crfree(td->td_standin->td_ucred);
 494                         td->td_standin->td_ucred = NULL;
 495                 }
 496                 thread_free(td->td_standin);
 497                 td->td_standin = NULL;
 498         }
 499 #endif
 500         /* Lock the last thread so we spin until it exits cpu_throw(). */
 501         thread_lock(td);
 502         thread_unlock(td);
 503         /* Wait for any remaining threads to exit cpu_throw(). */
 504         while (p->p_exitthreads)
 505                 sched_relinquish(curthread);
 506         cpu_thread_clean(td);
 507         crfree(td->td_ucred);
 508         thread_reap();  /* check for zombie threads etc. */
 509 }
 510
 511 /*
 512  * Link a thread to a process.
 513  * set up anything that needs to be initialized for it to
 514  * be used by the process.
 515  *
 516  * Note that we do not link to the proc's ucred here.
 517  * The thread is linked as if running but no KSE assigned.
 518  * Called from:
 519  *  proc_linkup()
 520  *  thread_schedule_upcall()
 521  *  thr_create()
 522  */
 523 void
 524 thread_link(struct thread *td, struct proc *p)
 525 {
 526
 527         /*
 528          * XXX This can't be enabled because it's called for proc0 before
 529          * it's spinlock has been created.
 530          * PROC_SLOCK_ASSERT(p, MA_OWNED);
 531          */
 532         td->td_state    = TDS_INACTIVE;
 533         td->td_proc     = p;
 534         td->td_flags    = 0;
 535
 536         LIST_INIT(&td->td_contested);
 537         sigqueue_init(&td->td_sigqueue, p);
 538         callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
 539         TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
 540         p->p_numthreads++;
 541 }
 542
 543 /*
 544  * Convert a process with one thread to an unthreaded process.
 545  * Called from:
 546  *  thread_single(exit)  (called from execve and exit)
 547  *  kse_exit()          XXX may need cleaning up wrt KSE stuff
 548  */
 549 void
 550 thread_unthread(struct thread *td)
 551 {
 552         struct proc *p = td->td_proc;
 553
 554         KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
 555 #ifdef KSE
 556         thread_lock(td);
 557         upcall_remove(td);
 558         thread_unlock(td);
 559         p->p_flag &= ~(P_SA|P_HADTHREADS);
 560         td->td_mailbox = NULL;
 561         td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
 562         if (td->td_standin != NULL) {
 563                 thread_zombie(td->td_standin);
 564                 td->td_standin = NULL;
 565         }
 566 #else
 567         p->p_flag &= ~P_HADTHREADS;
 568 #endif
 569 }
 570
 571 /*
 572  * Called from:
 573  *  thread_exit()
 574  */
 575 void
 576 thread_unlink(struct thread *td)
 577 {
 578         struct proc *p = td->td_proc;
 579
 580         PROC_SLOCK_ASSERT(p, MA_OWNED);
 581         TAILQ_REMOVE(&p->p_threads, td, td_plist);
 582         p->p_numthreads--;
 583         /* could clear a few other things here */
 584         /* Must  NOT clear links to proc! */
 585 }
 586
 587 /*
 588  * Enforce single-threading.
 589  *
 590  * Returns 1 if the caller must abort (another thread is waiting to
 591  * exit the process or similar). Process is locked!
 592  * Returns 0 when you are successfully the only thread running.
 593  * A process has successfully single threaded in the suspend mode when
 594  * There are no threads in user mode. Threads in the kernel must be
 595  * allowed to continue until they get to the user boundary. They may even
 596  * copy out their return values and data before suspending. They may however be
 597  * accelerated in reaching the user boundary as we will wake up
 598  * any sleeping threads that are interruptable. (PCATCH).
 599  */
 600 int
 601 thread_single(int mode)
 602 {
 603         struct thread *td;
 604         struct thread *td2;
 605         struct proc *p;
 606         int remaining;
 607
 608         td = curthread;
 609         p = td->td_proc;
 610         mtx_assert(&Giant, MA_NOTOWNED);
 611         PROC_LOCK_ASSERT(p, MA_OWNED);
 612         KASSERT((td != NULL), ("curthread is NULL"));
 613
 614         if ((p->p_flag & P_HADTHREADS) == 0)
 615                 return (0);
 616
 617         /* Is someone already single threading? */
 618         if (p->p_singlethread != NULL && p->p_singlethread != td)
 619                 return (1);
 620
 621         if (mode == SINGLE_EXIT) {
 622                 p->p_flag |= P_SINGLE_EXIT;
 623                 p->p_flag &= ~P_SINGLE_BOUNDARY;
 624         } else {
 625                 p->p_flag &= ~P_SINGLE_EXIT;
 626                 if (mode == SINGLE_BOUNDARY)
 627                         p->p_flag |= P_SINGLE_BOUNDARY;
 628                 else
 629                         p->p_flag &= ~P_SINGLE_BOUNDARY;
 630         }
 631         p->p_flag |= P_STOPPED_SINGLE;
 632         PROC_SLOCK(p);
 633         p->p_singlethread = td;
 634         if (mode == SINGLE_EXIT)
 635                 remaining = p->p_numthreads;
 636         else if (mode == SINGLE_BOUNDARY)
 637                 remaining = p->p_numthreads - p->p_boundary_count;
 638         else
 639                 remaining = p->p_numthreads - p->p_suspcount;
 640         while (remaining != 1) {
 641                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
 642                         goto stopme;
 643                 FOREACH_THREAD_IN_PROC(p, td2) {
 644                         if (td2 == td)
 645                                 continue;
 646                         thread_lock(td2);
 647                         td2->td_flags |= TDF_ASTPENDING;
 648                         if (TD_IS_INHIBITED(td2)) {
 649                                 switch (mode) {
 650                                 case SINGLE_EXIT:
 651                                         if (td->td_flags & TDF_DBSUSPEND)
 652                                                 td->td_flags &= ~TDF_DBSUSPEND;
 653                                         if (TD_IS_SUSPENDED(td2))
 654                                                 thread_unsuspend_one(td2);
 655                                         if (TD_ON_SLEEPQ(td2) &&
 656                                             (td2->td_flags & TDF_SINTR))
 657                                                 sleepq_abort(td2, EINTR);
 658                                         break;
 659                                 case SINGLE_BOUNDARY:
 660                                         if (TD_IS_SUSPENDED(td2) &&
 661                                             !(td2->td_flags & TDF_BOUNDARY))
 662                                                 thread_unsuspend_one(td2);
 663                                         if (TD_ON_SLEEPQ(td2) &&
 664                                             (td2->td_flags & TDF_SINTR))
 665                                                 sleepq_abort(td2, ERESTART);
 666                                         break;
 667                                 default:
 668                                         if (TD_IS_SUSPENDED(td2)) {
 669                                                 thread_unlock(td2);
 670                                                 continue;
 671                                         }
 672                                         /*
 673                                          * maybe other inhibited states too?
 674                                          */
 675                                         if ((td2->td_flags & TDF_SINTR) &&
 676                                             (td2->td_inhibitors &
 677                                             (TDI_SLEEPING | TDI_SWAPPED)))
 678                                                 thread_suspend_one(td2);
 679                                         break;
 680                                 }
 681                         }
 682 #ifdef SMP
 683                         else if (TD_IS_RUNNING(td2) && td != td2) {
 684                                 forward_signal(td2);
 685                         }
 686 #endif
 687                         thread_unlock(td2);
 688                 }
 689                 if (mode == SINGLE_EXIT)
 690                         remaining = p->p_numthreads;
 691                 else if (mode == SINGLE_BOUNDARY)
 692                         remaining = p->p_numthreads - p->p_boundary_count;
 693                 else
 694                         remaining = p->p_numthreads - p->p_suspcount;
 695
 696                 /*
 697                  * Maybe we suspended some threads.. was it enough?
 698                  */
 699                 if (remaining == 1)
 700                         break;
 701
 702 stopme:
 703                 /*
 704                  * Wake us up when everyone else has suspended.
 705                  * In the mean time we suspend as well.
 706                  */
 707                 thread_suspend_switch(td);
 708                 if (mode == SINGLE_EXIT)
 709                         remaining = p->p_numthreads;
 710                 else if (mode == SINGLE_BOUNDARY)
 711                         remaining = p->p_numthreads - p->p_boundary_count;
 712                 else
 713                         remaining = p->p_numthreads - p->p_suspcount;
 714         }
 715         if (mode == SINGLE_EXIT) {
 716                 /*
 717                  * We have gotten rid of all the other threads and we
 718                  * are about to either exit or exec. In either case,
 719                  * we try our utmost  to revert to being a non-threaded
 720                  * process.
 721                  */
 722                 p->p_singlethread = NULL;
 723                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
 724                 thread_unthread(td);
 725         }
 726         PROC_SUNLOCK(p);
 727         return (0);
 728 }
 729
 730 /*
 731  * Called in from locations that can safely check to see
 732  * whether we have to suspend or at least throttle for a
 733  * single-thread event (e.g. fork).
 734  *
 735  * Such locations include userret().
 736  * If the "return_instead" argument is non zero, the thread must be able to
 737  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
 738  *
 739  * The 'return_instead' argument tells the function if it may do a
 740  * thread_exit() or suspend, or whether the caller must abort and back
 741  * out instead.
 742  *
 743  * If the thread that set the single_threading request has set the
 744  * P_SINGLE_EXIT bit in the process flags then this call will never return
 745  * if 'return_instead' is false, but will exit.
 746  *
 747  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
 748  *---------------+--------------------+---------------------
 749  *       0       | returns 0          |   returns 0 or 1
 750  *               | when ST ends       |   immediatly
 751  *---------------+--------------------+---------------------
 752  *       1       | thread exits       |   returns 1
 753  *               |                    |  immediatly
 754  * 0 = thread_exit() or suspension ok,
 755  * other = return error instead of stopping the thread.
 756  *
 757  * While a full suspension is under effect, even a single threading
 758  * thread would be suspended if it made this call (but it shouldn't).
 759  * This call should only be made from places where
 760  * thread_exit() would be safe as that may be the outcome unless
 761  * return_instead is set.
 762  */
 763 int
 764 thread_suspend_check(int return_instead)
 765 {
 766         struct thread *td;
 767         struct proc *p;
 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_flags & TDF_DBSUSPEND))) {
 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                 /* If thread will exit, flush its pending signals */
 796                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 797                         sigqueue_flush(&td->td_sigqueue);
 798
 799                 PROC_SLOCK(p);
 800                 thread_stopped(p);
 801                 /*
 802                  * If the process is waiting for us to exit,
 803                  * this thread should just suicide.
 804                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
 805                  */
 806                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 807                         thread_exit();
 808                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 809                         if (p->p_numthreads == p->p_suspcount + 1) {
 810                                 thread_lock(p->p_singlethread);
 811                                 thread_unsuspend_one(p->p_singlethread);
 812                                 thread_unlock(p->p_singlethread);
 813                         }
 814                 }
 815                 PROC_UNLOCK(p);
 816                 thread_lock(td);
 817                 /*
 818                  * When a thread suspends, it just
 819                  * gets taken off all queues.
 820                  */
 821                 thread_suspend_one(td);
 822                 if (return_instead == 0) {
 823                         p->p_boundary_count++;
 824                         td->td_flags |= TDF_BOUNDARY;
 825                 }
 826                 PROC_SUNLOCK(p);
 827                 mi_switch(SW_INVOL, NULL);
 828                 if (return_instead == 0)
 829                         td->td_flags &= ~TDF_BOUNDARY;
 830                 thread_unlock(td);
 831                 PROC_LOCK(p);
 832                 if (return_instead == 0)
 833                         p->p_boundary_count--;
 834         }
 835         return (0);
 836 }
 837
 838 void
 839 thread_suspend_switch(struct thread *td)
 840 {
 841         struct proc *p;
 842
 843         p = td->td_proc;
 844         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 845         PROC_LOCK_ASSERT(p, MA_OWNED);
 846         PROC_SLOCK_ASSERT(p, MA_OWNED);
 847         /*
 848          * We implement thread_suspend_one in stages here to avoid
 849          * dropping the proc lock while the thread lock is owned.
 850          */
 851         thread_stopped(p);
 852         p->p_suspcount++;
 853         PROC_UNLOCK(p);
 854         thread_lock(td);
 855         TD_SET_SUSPENDED(td);
 856         PROC_SUNLOCK(p);
 857         DROP_GIANT();
 858         mi_switch(SW_VOL, NULL);
 859         thread_unlock(td);
 860         PICKUP_GIANT();
 861         PROC_LOCK(p);
 862         PROC_SLOCK(p);
 863 }
 864
 865 void
 866 thread_suspend_one(struct thread *td)
 867 {
 868         struct proc *p = td->td_proc;
 869
 870         PROC_SLOCK_ASSERT(p, MA_OWNED);
 871         THREAD_LOCK_ASSERT(td, MA_OWNED);
 872         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 873         p->p_suspcount++;
 874         TD_SET_SUSPENDED(td);
 875 }
 876
 877 void
 878 thread_unsuspend_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), ("Thread not suspended"));
 885         TD_CLR_SUSPENDED(td);
 886         p->p_suspcount--;
 887         setrunnable(td);
 888 }
 889
 890 /*
 891  * Allow all threads blocked by single threading to continue running.
 892  */
 893 void
 894 thread_unsuspend(struct proc *p)
 895 {
 896         struct thread *td;
 897
 898         PROC_LOCK_ASSERT(p, MA_OWNED);
 899         PROC_SLOCK_ASSERT(p, MA_OWNED);
 900         if (!P_SHOULDSTOP(p)) {
 901                 FOREACH_THREAD_IN_PROC(p, td) {
 902                         thread_lock(td);
 903                         if (TD_IS_SUSPENDED(td)) {
 904                                 thread_unsuspend_one(td);
 905                         }
 906                         thread_unlock(td);
 907                 }
 908         } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
 909             (p->p_numthreads == p->p_suspcount)) {
 910                 /*
 911                  * Stopping everything also did the job for the single
 912                  * threading request. Now we've downgraded to single-threaded,
 913                  * let it continue.
 914                  */
 915                 thread_lock(p->p_singlethread);
 916                 thread_unsuspend_one(p->p_singlethread);
 917                 thread_unlock(p->p_singlethread);
 918         }
 919 }
 920
 921 /*
 922  * End the single threading mode..
 923  */
 924 void
 925 thread_single_end(void)
 926 {
 927         struct thread *td;
 928         struct proc *p;
 929
 930         td = curthread;
 931         p = td->td_proc;
 932         PROC_LOCK_ASSERT(p, MA_OWNED);
 933         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
 934         PROC_SLOCK(p);
 935         p->p_singlethread = NULL;
 936         /*
 937          * If there are other threads they mey now run,
 938          * unless of course there is a blanket 'stop order'
 939          * on the process. The single threader must be allowed
 940          * to continue however as this is a bad place to stop.
 941          */
 942         if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
 943                 FOREACH_THREAD_IN_PROC(p, td) {
 944                         thread_lock(td);
 945                         if (TD_IS_SUSPENDED(td)) {
 946                                 thread_unsuspend_one(td);
 947                         }
 948                         thread_unlock(td);
 949                 }
 950         }
 951         PROC_SUNLOCK(p);
 952 }
 953
 954 struct thread *
 955 thread_find(struct proc *p, lwpid_t tid)
 956 {
 957         struct thread *td;
 958
 959         PROC_LOCK_ASSERT(p, MA_OWNED);
 960         PROC_SLOCK(p);
 961         FOREACH_THREAD_IN_PROC(p, td) {
 962                 if (td->td_tid == tid)
 963                         break;
 964         }
 965         PROC_SUNLOCK(p);
 966         return (td);
 967 }