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/selinfo.h>
  44 #include <sys/turnstile.h>
  45 #include <sys/ktr.h>
  46 #include <sys/umtx.h>
  47
  48 #include <security/audit/audit.h>
  49
  50 #include <vm/vm.h>
  51 #include <vm/vm_extern.h>
  52 #include <vm/uma.h>
  53 #include <sys/eventhandler.h>
  54
  55 /*
  56  * thread related storage.
  57  */
  58 static uma_zone_t thread_zone;
  59
  60 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
  61
  62 int max_threads_per_proc = 1500;
  63 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
  64         &max_threads_per_proc, 0, "Limit on threads per proc");
  65
  66 int max_threads_hits;
  67 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
  68         &max_threads_hits, 0, "");
  69
  70 #ifdef KSE
  71 int virtual_cpu;
  72
  73 #endif
  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 #ifdef KSE
  81 static int
  82 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
  83 {
  84         int error, new_val;
  85         int def_val;
  86
  87         def_val = mp_ncpus;
  88         if (virtual_cpu == 0)
  89                 new_val = def_val;
  90         else
  91                 new_val = virtual_cpu;
  92         error = sysctl_handle_int(oidp, &new_val, 0, req);
  93         if (error != 0 || req->newptr == NULL)
  94                 return (error);
  95         if (new_val < 0)
  96                 return (EINVAL);
  97         virtual_cpu = new_val;
  98         return (0);
  99 }
 100
 101 /* DEBUG ONLY */
 102 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
 103         0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
 104         "debug virtual cpus");
 105 #endif
 106
 107 struct mtx tid_lock;
 108 static struct unrhdr *tid_unrhdr;
 109
 110 /*
 111  * Prepare a thread for use.
 112  */
 113 static int
 114 thread_ctor(void *mem, int size, void *arg, int flags)
 115 {
 116         struct thread   *td;
 117
 118         td = (struct thread *)mem;
 119         td->td_state = TDS_INACTIVE;
 120         td->td_oncpu = NOCPU;
 121
 122         td->td_tid = alloc_unr(tid_unrhdr);
 123         td->td_syscalls = 0;
 124
 125         /*
 126          * Note that td_critnest begins life as 1 because the thread is not
 127          * running and is thereby implicitly waiting to be on the receiving
 128          * end of a context switch.
 129          */
 130         td->td_critnest = 1;
 131         EVENTHANDLER_INVOKE(thread_ctor, td);
 132 #ifdef AUDIT
 133         audit_thread_alloc(td);
 134 #endif
 135         umtx_thread_alloc(td);
 136         return (0);
 137 }
 138
 139 /*
 140  * Reclaim a thread after use.
 141  */
 142 static void
 143 thread_dtor(void *mem, int size, void *arg)
 144 {
 145         struct thread *td;
 146
 147         td = (struct thread *)mem;
 148
 149 #ifdef INVARIANTS
 150         /* Verify that this thread is in a safe state to free. */
 151         switch (td->td_state) {
 152         case TDS_INHIBITED:
 153         case TDS_RUNNING:
 154         case TDS_CAN_RUN:
 155         case TDS_RUNQ:
 156                 /*
 157                  * We must never unlink a thread that is in one of
 158                  * these states, because it is currently active.
 159                  */
 160                 panic("bad state for thread unlinking");
 161                 /* NOTREACHED */
 162         case TDS_INACTIVE:
 163                 break;
 164         default:
 165                 panic("bad thread state");
 166                 /* NOTREACHED */
 167         }
 168 #endif
 169 #ifdef AUDIT
 170         audit_thread_free(td);
 171 #endif
 172         EVENTHANDLER_INVOKE(thread_dtor, td);
 173         free_unr(tid_unrhdr, td->td_tid);
 174         sched_newthread(td);
 175 }
 176
 177 /*
 178  * Initialize type-stable parts of a thread (when newly created).
 179  */
 180 static int
 181 thread_init(void *mem, int size, int flags)
 182 {
 183         struct thread *td;
 184
 185         td = (struct thread *)mem;
 186
 187         td->td_sleepqueue = sleepq_alloc();
 188         td->td_turnstile = turnstile_alloc();
 189         EVENTHANDLER_INVOKE(thread_init, td);
 190         td->td_sched = (struct td_sched *)&td[1];
 191         sched_newthread(td);
 192         umtx_thread_init(td);
 193         td->td_kstack = 0;
 194         return (0);
 195 }
 196
 197 /*
 198  * Tear down type-stable parts of a thread (just before being discarded).
 199  */
 200 static void
 201 thread_fini(void *mem, int size)
 202 {
 203         struct thread *td;
 204
 205         td = (struct thread *)mem;
 206         EVENTHANDLER_INVOKE(thread_fini, td);
 207         turnstile_free(td->td_turnstile);
 208         sleepq_free(td->td_sleepqueue);
 209         umtx_thread_fini(td);
 210         seltdfini(td);
 211 }
 212
 213 /*
 214  * For a newly created process,
 215  * link up all the structures and its initial threads etc.
 216  * called from:
 217  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
 218  * proc_dtor() (should go away)
 219  * proc_init()
 220  */
 221 void
 222 proc_linkup0(struct proc *p, struct thread *td)
 223 {
 224         TAILQ_INIT(&p->p_threads);           /* all threads in proc */
 225         proc_linkup(p, td);
 226 }
 227
 228 void
 229 proc_linkup(struct proc *p, struct thread *td)
 230 {
 231
 232 #ifdef KSE
 233         TAILQ_INIT(&p->p_upcalls);           /* upcall list */
 234 #endif
 235         sigqueue_init(&p->p_sigqueue, p);
 236         p->p_ksi = ksiginfo_alloc(1);
 237         if (p->p_ksi != NULL) {
 238                 /* XXX p_ksi may be null if ksiginfo zone is not ready */
 239                 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
 240         }
 241         LIST_INIT(&p->p_mqnotifier);
 242         p->p_numthreads = 0;
 243         thread_link(td, p);
 244 }
 245
 246 /*
 247  * Initialize global thread allocation resources.
 248  */
 249 void
 250 threadinit(void)
 251 {
 252
 253         mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
 254         /* leave one number for thread0 */
 255         tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
 256
 257         thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
 258             thread_ctor, thread_dtor, thread_init, thread_fini,
 259             16 - 1, 0);
 260 #ifdef KSE
 261         kseinit();      /* set up kse specific stuff  e.g. upcall zone*/
 262 #endif
 263 }
 264
 265 /*
 266  * Place an unused thread on the zombie list.
 267  * Use the slpq as that must be unused by now.
 268  */
 269 void
 270 thread_zombie(struct thread *td)
 271 {
 272         mtx_lock_spin(&zombie_lock);
 273         TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
 274         mtx_unlock_spin(&zombie_lock);
 275 }
 276
 277 /*
 278  * Release a thread that has exited after cpu_throw().
 279  */
 280 void
 281 thread_stash(struct thread *td)
 282 {
 283         atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
 284         thread_zombie(td);
 285 }
 286
 287 /*
 288  * Reap zombie kse resource.
 289  */
 290 void
 291 thread_reap(void)
 292 {
 293         struct thread *td_first, *td_next;
 294
 295         /*
 296          * Don't even bother to lock if none at this instant,
 297          * we really don't care about the next instant..
 298          */
 299         if (!TAILQ_EMPTY(&zombie_threads)) {
 300                 mtx_lock_spin(&zombie_lock);
 301                 td_first = TAILQ_FIRST(&zombie_threads);
 302                 if (td_first)
 303                         TAILQ_INIT(&zombie_threads);
 304                 mtx_unlock_spin(&zombie_lock);
 305                 while (td_first) {
 306                         td_next = TAILQ_NEXT(td_first, td_slpq);
 307                         if (td_first->td_ucred)
 308                                 crfree(td_first->td_ucred);
 309                         thread_free(td_first);
 310                         td_first = td_next;
 311                 }
 312         }
 313 #ifdef KSE
 314         upcall_reap();
 315 #endif
 316 }
 317
 318 /*
 319  * Allocate a thread.
 320  */
 321 struct thread *
 322 thread_alloc(void)
 323 {
 324         struct thread *td;
 325
 326         thread_reap(); /* check if any zombies to get */
 327
 328         td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
 329         KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
 330         if (!vm_thread_new(td, 0)) {
 331                 uma_zfree(thread_zone, td);
 332                 return (NULL);
 333         }
 334         cpu_thread_alloc(td);
 335         return (td);
 336 }
 337
 338
 339 /*
 340  * Deallocate a thread.
 341  */
 342 void
 343 thread_free(struct thread *td)
 344 {
 345
 346         cpu_thread_free(td);
 347         if (td->td_altkstack != 0)
 348                 vm_thread_dispose_altkstack(td);
 349         if (td->td_kstack != 0)
 350                 vm_thread_dispose(td);
 351         uma_zfree(thread_zone, td);
 352 }
 353
 354 /*
 355  * Discard the current thread and exit from its context.
 356  * Always called with scheduler locked.
 357  *
 358  * Because we can't free a thread while we're operating under its context,
 359  * push the current thread into our CPU's deadthread holder. This means
 360  * we needn't worry about someone else grabbing our context before we
 361  * do a cpu_throw().  This may not be needed now as we are under schedlock.
 362  * Maybe we can just do a thread_stash() as thr_exit1 does.
 363  */
 364 /*  XXX
 365  * libthr expects its thread exit to return for the last
 366  * thread, meaning that the program is back to non-threaded
 367  * mode I guess. Because we do this (cpu_throw) unconditionally
 368  * here, they have their own version of it. (thr_exit1())
 369  * that doesn't do it all if this was the last thread.
 370  * It is also called from thread_suspend_check().
 371  * Of course in the end, they end up coming here through exit1
 372  * anyhow..  After fixing 'thr' to play by the rules we should be able
 373  * to merge these two functions together.
 374  *
 375  * called from:
 376  * exit1()
 377  * kse_exit()
 378  * thr_exit()
 379  * ifdef KSE
 380  * thread_user_enter()
 381  * thread_userret()
 382  * endif
 383  * thread_suspend_check()
 384  */
 385 void
 386 thread_exit(void)
 387 {
 388         uint64_t new_switchtime;
 389         struct thread *td;
 390         struct thread *td2;
 391         struct proc *p;
 392
 393         td = curthread;
 394         p = td->td_proc;
 395
 396         PROC_SLOCK_ASSERT(p, MA_OWNED);
 397         mtx_assert(&Giant, MA_NOTOWNED);
 398
 399         PROC_LOCK_ASSERT(p, MA_OWNED);
 400         KASSERT(p != NULL, ("thread exiting without a process"));
 401         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 402             (long)p->p_pid, td->td_name);
 403         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 404
 405 #ifdef AUDIT
 406         AUDIT_SYSCALL_EXIT(0, td);
 407 #endif
 408
 409 #ifdef KSE
 410         if (td->td_standin != NULL) {
 411                 /*
 412                  * Note that we don't need to free the cred here as it
 413                  * is done in thread_reap().
 414                  */
 415                 thread_zombie(td->td_standin);
 416                 td->td_standin = NULL;
 417         }
 418 #endif
 419
 420         umtx_thread_exit(td);
 421
 422         /*
 423          * drop FPU & debug register state storage, or any other
 424          * architecture specific resources that
 425          * would not be on a new untouched process.
 426          */
 427         cpu_thread_exit(td);    /* XXXSMP */
 428
 429         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 430         new_switchtime = cpu_ticks();
 431         p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
 432         PCPU_SET(switchtime, new_switchtime);
 433         PCPU_SET(switchticks, ticks);
 434         PCPU_INC(cnt.v_swtch);
 435         /* Save our resource usage in our process. */
 436         td->td_ru.ru_nvcsw++;
 437         rucollect(&p->p_ru, &td->td_ru);
 438         /*
 439          * The last thread is left attached to the process
 440          * So that the whole bundle gets recycled. Skip
 441          * all this stuff if we never had threads.
 442          * EXIT clears all sign of other threads when
 443          * it goes to single threading, so the last thread always
 444          * takes the short path.
 445          */
 446         if (p->p_flag & P_HADTHREADS) {
 447                 if (p->p_numthreads > 1) {
 448                         thread_lock(td);
 449 #ifdef KSE
 450                         kse_unlink(td);
 451 #else
 452                         thread_unlink(td);
 453 #endif
 454                         thread_unlock(td);
 455                         td2 = FIRST_THREAD_IN_PROC(p);
 456                         sched_exit_thread(td2, td);
 457
 458                         /*
 459                          * The test below is NOT true if we are the
 460                          * sole exiting thread. P_STOPPED_SNGL is unset
 461                          * in exit1() after it is the only survivor.
 462                          */
 463                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 464                                 if (p->p_numthreads == p->p_suspcount) {
 465                                         thread_lock(p->p_singlethread);
 466                                         thread_unsuspend_one(p->p_singlethread);
 467                                         thread_unlock(p->p_singlethread);
 468                                 }
 469                         }
 470
 471                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 472                         PCPU_SET(deadthread, td);
 473                 } else {
 474                         /*
 475                          * The last thread is exiting.. but not through exit()
 476                          * what should we do?
 477                          * Theoretically this can't happen
 478                          * exit1() - clears threading flags before coming here
 479                          * kse_exit() - treats last thread specially
 480                          * thr_exit() - treats last thread specially
 481                          * ifdef KSE
 482                          * thread_user_enter() - only if more exist
 483                          * thread_userret() - only if more exist
 484                          * endif
 485                          * thread_suspend_check() - only if more exist
 486                          */
 487                         panic ("thread_exit: Last thread exiting on its own");
 488                 }
 489         }
 490         PROC_UNLOCK(p);
 491         thread_lock(td);
 492         /* Save our tick information with both the thread and proc locked */
 493         ruxagg(&p->p_rux, td);
 494         PROC_SUNLOCK(p);
 495         td->td_state = TDS_INACTIVE;
 496         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 497         sched_throw(td);
 498         panic("I'm a teapot!");
 499         /* NOTREACHED */
 500 }
 501
 502 /*
 503  * Do any thread specific cleanups that may be needed in wait()
 504  * called with Giant, proc and schedlock not held.
 505  */
 506 void
 507 thread_wait(struct proc *p)
 508 {
 509         struct thread *td;
 510
 511         mtx_assert(&Giant, MA_NOTOWNED);
 512         KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
 513         td = FIRST_THREAD_IN_PROC(p);
 514 #ifdef KSE
 515         if (td->td_standin != NULL) {
 516                 if (td->td_standin->td_ucred != NULL) {
 517                         crfree(td->td_standin->td_ucred);
 518                         td->td_standin->td_ucred = NULL;
 519                 }
 520                 thread_free(td->td_standin);
 521                 td->td_standin = NULL;
 522         }
 523 #endif
 524         /* Lock the last thread so we spin until it exits cpu_throw(). */
 525         thread_lock(td);
 526         thread_unlock(td);
 527         /* Wait for any remaining threads to exit cpu_throw(). */
 528         while (p->p_exitthreads)
 529                 sched_relinquish(curthread);
 530         cpu_thread_clean(td);
 531         crfree(td->td_ucred);
 532         thread_reap();  /* check for zombie threads etc. */
 533 }
 534
 535 /*
 536  * Link a thread to a process.
 537  * set up anything that needs to be initialized for it to
 538  * be used by the process.
 539  *
 540  * Note that we do not link to the proc's ucred here.
 541  * The thread is linked as if running but no KSE assigned.
 542  * Called from:
 543  *  proc_linkup()
 544  *  thread_schedule_upcall()
 545  *  thr_create()
 546  */
 547 void
 548 thread_link(struct thread *td, struct proc *p)
 549 {
 550
 551         /*
 552          * XXX This can't be enabled because it's called for proc0 before
 553          * it's spinlock has been created.
 554          * PROC_SLOCK_ASSERT(p, MA_OWNED);
 555          */
 556         td->td_state    = TDS_INACTIVE;
 557         td->td_proc     = p;
 558         td->td_flags    = TDF_INMEM;
 559
 560         LIST_INIT(&td->td_contested);
 561         LIST_INIT(&td->td_lprof[0]);
 562         LIST_INIT(&td->td_lprof[1]);
 563         sigqueue_init(&td->td_sigqueue, p);
 564         callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
 565         TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
 566         p->p_numthreads++;
 567 }
 568
 569 /*
 570  * Convert a process with one thread to an unthreaded process.
 571  * Called from:
 572  *  thread_single(exit)  (called from execve and exit)
 573  *  kse_exit()          XXX may need cleaning up wrt KSE stuff
 574  */
 575 void
 576 thread_unthread(struct thread *td)
 577 {
 578         struct proc *p = td->td_proc;
 579
 580         KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
 581 #ifdef KSE
 582         thread_lock(td);
 583         upcall_remove(td);
 584         thread_unlock(td);
 585         p->p_flag &= ~(P_SA|P_HADTHREADS);
 586         td->td_mailbox = NULL;
 587         td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
 588         if (td->td_standin != NULL) {
 589                 thread_zombie(td->td_standin);
 590                 td->td_standin = NULL;
 591         }
 592 #else
 593         p->p_flag &= ~P_HADTHREADS;
 594 #endif
 595 }
 596
 597 /*
 598  * Called from:
 599  *  thread_exit()
 600  */
 601 void
 602 thread_unlink(struct thread *td)
 603 {
 604         struct proc *p = td->td_proc;
 605
 606         PROC_SLOCK_ASSERT(p, MA_OWNED);
 607         TAILQ_REMOVE(&p->p_threads, td, td_plist);
 608         p->p_numthreads--;
 609         /* could clear a few other things here */
 610         /* Must  NOT clear links to proc! */
 611 }
 612
 613 /*
 614  * Enforce single-threading.
 615  *
 616  * Returns 1 if the caller must abort (another thread is waiting to
 617  * exit the process or similar). Process is locked!
 618  * Returns 0 when you are successfully the only thread running.
 619  * A process has successfully single threaded in the suspend mode when
 620  * There are no threads in user mode. Threads in the kernel must be
 621  * allowed to continue until they get to the user boundary. They may even
 622  * copy out their return values and data before suspending. They may however be
 623  * accelerated in reaching the user boundary as we will wake up
 624  * any sleeping threads that are interruptable. (PCATCH).
 625  */
 626 int
 627 thread_single(int mode)
 628 {
 629         struct thread *td;
 630         struct thread *td2;
 631         struct proc *p;
 632         int remaining;
 633
 634         td = curthread;
 635         p = td->td_proc;
 636         mtx_assert(&Giant, MA_NOTOWNED);
 637         PROC_LOCK_ASSERT(p, MA_OWNED);
 638         KASSERT((td != NULL), ("curthread is NULL"));
 639
 640         if ((p->p_flag & P_HADTHREADS) == 0)
 641                 return (0);
 642
 643         /* Is someone already single threading? */
 644         if (p->p_singlethread != NULL && p->p_singlethread != td)
 645                 return (1);
 646
 647         if (mode == SINGLE_EXIT) {
 648                 p->p_flag |= P_SINGLE_EXIT;
 649                 p->p_flag &= ~P_SINGLE_BOUNDARY;
 650         } else {
 651                 p->p_flag &= ~P_SINGLE_EXIT;
 652                 if (mode == SINGLE_BOUNDARY)
 653                         p->p_flag |= P_SINGLE_BOUNDARY;
 654                 else
 655                         p->p_flag &= ~P_SINGLE_BOUNDARY;
 656         }
 657         p->p_flag |= P_STOPPED_SINGLE;
 658         PROC_SLOCK(p);
 659         p->p_singlethread = td;
 660         if (mode == SINGLE_EXIT)
 661                 remaining = p->p_numthreads;
 662         else if (mode == SINGLE_BOUNDARY)
 663                 remaining = p->p_numthreads - p->p_boundary_count;
 664         else
 665                 remaining = p->p_numthreads - p->p_suspcount;
 666         while (remaining != 1) {
 667                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
 668                         goto stopme;
 669                 FOREACH_THREAD_IN_PROC(p, td2) {
 670                         if (td2 == td)
 671                                 continue;
 672                         thread_lock(td2);
 673                         td2->td_flags |= TDF_ASTPENDING;
 674                         if (TD_IS_INHIBITED(td2)) {
 675                                 switch (mode) {
 676                                 case SINGLE_EXIT:
 677                                         if (td->td_flags & TDF_DBSUSPEND)
 678                                                 td->td_flags &= ~TDF_DBSUSPEND;
 679                                         if (TD_IS_SUSPENDED(td2))
 680                                                 thread_unsuspend_one(td2);
 681                                         if (TD_ON_SLEEPQ(td2) &&
 682                                             (td2->td_flags & TDF_SINTR))
 683                                                 sleepq_abort(td2, EINTR);
 684                                         break;
 685                                 case SINGLE_BOUNDARY:
 686                                         break;
 687                                 default:
 688                                         if (TD_IS_SUSPENDED(td2)) {
 689                                                 thread_unlock(td2);
 690                                                 continue;
 691                                         }
 692                                         /*
 693                                          * maybe other inhibited states too?
 694                                          */
 695                                         if ((td2->td_flags & TDF_SINTR) &&
 696                                             (td2->td_inhibitors &
 697                                             (TDI_SLEEPING | TDI_SWAPPED)))
 698                                                 thread_suspend_one(td2);
 699                                         break;
 700                                 }
 701                         }
 702 #ifdef SMP
 703                         else if (TD_IS_RUNNING(td2) && td != td2) {
 704                                 forward_signal(td2);
 705                         }
 706 #endif
 707                         thread_unlock(td2);
 708                 }
 709                 if (mode == SINGLE_EXIT)
 710                         remaining = p->p_numthreads;
 711                 else if (mode == SINGLE_BOUNDARY)
 712                         remaining = p->p_numthreads - p->p_boundary_count;
 713                 else
 714                         remaining = p->p_numthreads - p->p_suspcount;
 715
 716                 /*
 717                  * Maybe we suspended some threads.. was it enough?
 718                  */
 719                 if (remaining == 1)
 720                         break;
 721
 722 stopme:
 723                 /*
 724                  * Wake us up when everyone else has suspended.
 725                  * In the mean time we suspend as well.
 726                  */
 727                 thread_suspend_switch(td);
 728                 if (mode == SINGLE_EXIT)
 729                         remaining = p->p_numthreads;
 730                 else if (mode == SINGLE_BOUNDARY)
 731                         remaining = p->p_numthreads - p->p_boundary_count;
 732                 else
 733                         remaining = p->p_numthreads - p->p_suspcount;
 734         }
 735         if (mode == SINGLE_EXIT) {
 736                 /*
 737                  * We have gotten rid of all the other threads and we
 738                  * are about to either exit or exec. In either case,
 739                  * we try our utmost  to revert to being a non-threaded
 740                  * process.
 741                  */
 742                 p->p_singlethread = NULL;
 743                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
 744                 thread_unthread(td);
 745         }
 746         PROC_SUNLOCK(p);
 747         return (0);
 748 }
 749
 750 /*
 751  * Called in from locations that can safely check to see
 752  * whether we have to suspend or at least throttle for a
 753  * single-thread event (e.g. fork).
 754  *
 755  * Such locations include userret().
 756  * If the "return_instead" argument is non zero, the thread must be able to
 757  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
 758  *
 759  * The 'return_instead' argument tells the function if it may do a
 760  * thread_exit() or suspend, or whether the caller must abort and back
 761  * out instead.
 762  *
 763  * If the thread that set the single_threading request has set the
 764  * P_SINGLE_EXIT bit in the process flags then this call will never return
 765  * if 'return_instead' is false, but will exit.
 766  *
 767  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
 768  *---------------+--------------------+---------------------
 769  *       0       | returns 0          |   returns 0 or 1
 770  *               | when ST ends       |   immediatly
 771  *---------------+--------------------+---------------------
 772  *       1       | thread exits       |   returns 1
 773  *               |                    |  immediatly
 774  * 0 = thread_exit() or suspension ok,
 775  * other = return error instead of stopping the thread.
 776  *
 777  * While a full suspension is under effect, even a single threading
 778  * thread would be suspended if it made this call (but it shouldn't).
 779  * This call should only be made from places where
 780  * thread_exit() would be safe as that may be the outcome unless
 781  * return_instead is set.
 782  */
 783 int
 784 thread_suspend_check(int return_instead)
 785 {
 786         struct thread *td;
 787         struct proc *p;
 788
 789         td = curthread;
 790         p = td->td_proc;
 791         mtx_assert(&Giant, MA_NOTOWNED);
 792         PROC_LOCK_ASSERT(p, MA_OWNED);
 793         while (P_SHOULDSTOP(p) ||
 794               ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
 795                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 796                         KASSERT(p->p_singlethread != NULL,
 797                             ("singlethread not set"));
 798                         /*
 799                          * The only suspension in action is a
 800                          * single-threading. Single threader need not stop.
 801                          * XXX Should be safe to access unlocked
 802                          * as it can only be set to be true by us.
 803                          */
 804                         if (p->p_singlethread == td)
 805                                 return (0);     /* Exempt from stopping. */
 806                 }
 807                 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
 808                         return (EINTR);
 809
 810                 /* Should we goto user boundary if we didn't come from there? */
 811                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
 812                     (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
 813                         return (ERESTART);
 814
 815                 /* If thread will exit, flush its pending signals */
 816                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 817                         sigqueue_flush(&td->td_sigqueue);
 818
 819                 PROC_SLOCK(p);
 820                 thread_stopped(p);
 821                 /*
 822                  * If the process is waiting for us to exit,
 823                  * this thread should just suicide.
 824                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
 825                  */
 826                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 827                         thread_exit();
 828                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 829                         if (p->p_numthreads == p->p_suspcount + 1) {
 830                                 thread_lock(p->p_singlethread);
 831                                 thread_unsuspend_one(p->p_singlethread);
 832                                 thread_unlock(p->p_singlethread);
 833                         }
 834                 }
 835                 PROC_UNLOCK(p);
 836                 thread_lock(td);
 837                 /*
 838                  * When a thread suspends, it just
 839                  * gets taken off all queues.
 840                  */
 841                 thread_suspend_one(td);
 842                 if (return_instead == 0) {
 843                         p->p_boundary_count++;
 844                         td->td_flags |= TDF_BOUNDARY;
 845                 }
 846                 PROC_SUNLOCK(p);
 847                 mi_switch(SW_INVOL, NULL);
 848                 if (return_instead == 0)
 849                         td->td_flags &= ~TDF_BOUNDARY;
 850                 thread_unlock(td);
 851                 PROC_LOCK(p);
 852                 if (return_instead == 0)
 853                         p->p_boundary_count--;
 854         }
 855         return (0);
 856 }
 857
 858 void
 859 thread_suspend_switch(struct thread *td)
 860 {
 861         struct proc *p;
 862
 863         p = td->td_proc;
 864         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 865         PROC_LOCK_ASSERT(p, MA_OWNED);
 866         PROC_SLOCK_ASSERT(p, MA_OWNED);
 867         /*
 868          * We implement thread_suspend_one in stages here to avoid
 869          * dropping the proc lock while the thread lock is owned.
 870          */
 871         thread_stopped(p);
 872         p->p_suspcount++;
 873         PROC_UNLOCK(p);
 874         thread_lock(td);
 875         sched_sleep(td);
 876         TD_SET_SUSPENDED(td);
 877         PROC_SUNLOCK(p);
 878         DROP_GIANT();
 879         mi_switch(SW_VOL, NULL);
 880         thread_unlock(td);
 881         PICKUP_GIANT();
 882         PROC_LOCK(p);
 883         PROC_SLOCK(p);
 884 }
 885
 886 void
 887 thread_suspend_one(struct thread *td)
 888 {
 889         struct proc *p = td->td_proc;
 890
 891         PROC_SLOCK_ASSERT(p, MA_OWNED);
 892         THREAD_LOCK_ASSERT(td, MA_OWNED);
 893         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 894         p->p_suspcount++;
 895         sched_sleep(td);
 896         TD_SET_SUSPENDED(td);
 897 }
 898
 899 void
 900 thread_unsuspend_one(struct thread *td)
 901 {
 902         struct proc *p = td->td_proc;
 903
 904         PROC_SLOCK_ASSERT(p, MA_OWNED);
 905         THREAD_LOCK_ASSERT(td, MA_OWNED);
 906         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
 907         TD_CLR_SUSPENDED(td);
 908         p->p_suspcount--;
 909         setrunnable(td);
 910 }
 911
 912 /*
 913  * Allow all threads blocked by single threading to continue running.
 914  */
 915 void
 916 thread_unsuspend(struct proc *p)
 917 {
 918         struct thread *td;
 919
 920         PROC_LOCK_ASSERT(p, MA_OWNED);
 921         PROC_SLOCK_ASSERT(p, MA_OWNED);
 922         if (!P_SHOULDSTOP(p)) {
 923                 FOREACH_THREAD_IN_PROC(p, td) {
 924                         thread_lock(td);
 925                         if (TD_IS_SUSPENDED(td)) {
 926                                 thread_unsuspend_one(td);
 927                         }
 928                         thread_unlock(td);
 929                 }
 930         } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
 931             (p->p_numthreads == p->p_suspcount)) {
 932                 /*
 933                  * Stopping everything also did the job for the single
 934                  * threading request. Now we've downgraded to single-threaded,
 935                  * let it continue.
 936                  */
 937                 thread_lock(p->p_singlethread);
 938                 thread_unsuspend_one(p->p_singlethread);
 939                 thread_unlock(p->p_singlethread);
 940         }
 941 }
 942
 943 /*
 944  * End the single threading mode..
 945  */
 946 void
 947 thread_single_end(void)
 948 {
 949         struct thread *td;
 950         struct proc *p;
 951
 952         td = curthread;
 953         p = td->td_proc;
 954         PROC_LOCK_ASSERT(p, MA_OWNED);
 955         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
 956         PROC_SLOCK(p);
 957         p->p_singlethread = NULL;
 958         /*
 959          * If there are other threads they mey now run,
 960          * unless of course there is a blanket 'stop order'
 961          * on the process. The single threader must be allowed
 962          * to continue however as this is a bad place to stop.
 963          */
 964         if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
 965                 FOREACH_THREAD_IN_PROC(p, td) {
 966                         thread_lock(td);
 967                         if (TD_IS_SUSPENDED(td)) {
 968                                 thread_unsuspend_one(td);
 969                         }
 970                         thread_unlock(td);
 971                 }
 972         }
 973         PROC_SUNLOCK(p);
 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         PROC_SLOCK(p);
 983         FOREACH_THREAD_IN_PROC(p, td) {
 984                 if (td->td_tid == tid)
 985                         break;
 986         }
 987         PROC_SUNLOCK(p);
 988         return (td);
 989 }