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         tid_unrhdr = new_unrhdr(PID_MAX + 1, INT_MAX, &tid_lock);
 255
 256         thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
 257             thread_ctor, thread_dtor, thread_init, thread_fini,
 258             16 - 1, 0);
 259 #ifdef KSE
 260         kseinit();      /* set up kse specific stuff  e.g. upcall zone*/
 261 #endif
 262 }
 263
 264 /*
 265  * Place an unused thread on the zombie list.
 266  * Use the slpq as that must be unused by now.
 267  */
 268 void
 269 thread_zombie(struct thread *td)
 270 {
 271         mtx_lock_spin(&zombie_lock);
 272         TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
 273         mtx_unlock_spin(&zombie_lock);
 274 }
 275
 276 /*
 277  * Release a thread that has exited after cpu_throw().
 278  */
 279 void
 280 thread_stash(struct thread *td)
 281 {
 282         atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
 283         thread_zombie(td);
 284 }
 285
 286 /*
 287  * Reap zombie kse resource.
 288  */
 289 void
 290 thread_reap(void)
 291 {
 292         struct thread *td_first, *td_next;
 293
 294         /*
 295          * Don't even bother to lock if none at this instant,
 296          * we really don't care about the next instant..
 297          */
 298         if (!TAILQ_EMPTY(&zombie_threads)) {
 299                 mtx_lock_spin(&zombie_lock);
 300                 td_first = TAILQ_FIRST(&zombie_threads);
 301                 if (td_first)
 302                         TAILQ_INIT(&zombie_threads);
 303                 mtx_unlock_spin(&zombie_lock);
 304                 while (td_first) {
 305                         td_next = TAILQ_NEXT(td_first, td_slpq);
 306                         if (td_first->td_ucred)
 307                                 crfree(td_first->td_ucred);
 308                         thread_free(td_first);
 309                         td_first = td_next;
 310                 }
 311         }
 312 #ifdef KSE
 313         upcall_reap();
 314 #endif
 315 }
 316
 317 /*
 318  * Allocate a thread.
 319  */
 320 struct thread *
 321 thread_alloc(void)
 322 {
 323         struct thread *td;
 324
 325         thread_reap(); /* check if any zombies to get */
 326
 327         td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
 328         KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
 329         if (!vm_thread_new(td, 0)) {
 330                 uma_zfree(thread_zone, td);
 331                 return (NULL);
 332         }
 333         cpu_thread_alloc(td);
 334         return (td);
 335 }
 336
 337
 338 /*
 339  * Deallocate a thread.
 340  */
 341 void
 342 thread_free(struct thread *td)
 343 {
 344
 345         cpu_thread_free(td);
 346         if (td->td_altkstack != 0)
 347                 vm_thread_dispose_altkstack(td);
 348         if (td->td_kstack != 0)
 349                 vm_thread_dispose(td);
 350         uma_zfree(thread_zone, td);
 351 }
 352
 353 /*
 354  * Discard the current thread and exit from its context.
 355  * Always called with scheduler locked.
 356  *
 357  * Because we can't free a thread while we're operating under its context,
 358  * push the current thread into our CPU's deadthread holder. This means
 359  * we needn't worry about someone else grabbing our context before we
 360  * do a cpu_throw().  This may not be needed now as we are under schedlock.
 361  * Maybe we can just do a thread_stash() as thr_exit1 does.
 362  */
 363 /*  XXX
 364  * libthr expects its thread exit to return for the last
 365  * thread, meaning that the program is back to non-threaded
 366  * mode I guess. Because we do this (cpu_throw) unconditionally
 367  * here, they have their own version of it. (thr_exit1())
 368  * that doesn't do it all if this was the last thread.
 369  * It is also called from thread_suspend_check().
 370  * Of course in the end, they end up coming here through exit1
 371  * anyhow..  After fixing 'thr' to play by the rules we should be able
 372  * to merge these two functions together.
 373  *
 374  * called from:
 375  * exit1()
 376  * kse_exit()
 377  * thr_exit()
 378  * ifdef KSE
 379  * thread_user_enter()
 380  * thread_userret()
 381  * endif
 382  * thread_suspend_check()
 383  */
 384 void
 385 thread_exit(void)
 386 {
 387         uint64_t new_switchtime;
 388         struct thread *td;
 389         struct thread *td2;
 390         struct proc *p;
 391
 392         td = curthread;
 393         p = td->td_proc;
 394
 395         PROC_SLOCK_ASSERT(p, MA_OWNED);
 396         mtx_assert(&Giant, MA_NOTOWNED);
 397
 398         PROC_LOCK_ASSERT(p, MA_OWNED);
 399         KASSERT(p != NULL, ("thread exiting without a process"));
 400         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 401             (long)p->p_pid, td->td_name);
 402         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 403
 404 #ifdef AUDIT
 405         AUDIT_SYSCALL_EXIT(0, td);
 406 #endif
 407
 408 #ifdef KSE
 409         if (td->td_standin != NULL) {
 410                 /*
 411                  * Note that we don't need to free the cred here as it
 412                  * is done in thread_reap().
 413                  */
 414                 thread_zombie(td->td_standin);
 415                 td->td_standin = NULL;
 416         }
 417 #endif
 418
 419         umtx_thread_exit(td);
 420
 421         /*
 422          * drop FPU & debug register state storage, or any other
 423          * architecture specific resources that
 424          * would not be on a new untouched process.
 425          */
 426         cpu_thread_exit(td);    /* XXXSMP */
 427
 428         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 429         new_switchtime = cpu_ticks();
 430         p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
 431         PCPU_SET(switchtime, new_switchtime);
 432         PCPU_SET(switchticks, ticks);
 433         PCPU_INC(cnt.v_swtch);
 434         /* Save our resource usage in our process. */
 435         td->td_ru.ru_nvcsw++;
 436         rucollect(&p->p_ru, &td->td_ru);
 437         /*
 438          * The last thread is left attached to the process
 439          * So that the whole bundle gets recycled. Skip
 440          * all this stuff if we never had threads.
 441          * EXIT clears all sign of other threads when
 442          * it goes to single threading, so the last thread always
 443          * takes the short path.
 444          */
 445         if (p->p_flag & P_HADTHREADS) {
 446                 if (p->p_numthreads > 1) {
 447                         thread_lock(td);
 448 #ifdef KSE
 449                         kse_unlink(td);
 450 #else
 451                         thread_unlink(td);
 452 #endif
 453                         thread_unlock(td);
 454                         td2 = FIRST_THREAD_IN_PROC(p);
 455                         sched_exit_thread(td2, td);
 456
 457                         /*
 458                          * The test below is NOT true if we are the
 459                          * sole exiting thread. P_STOPPED_SNGL is unset
 460                          * in exit1() after it is the only survivor.
 461                          */
 462                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 463                                 if (p->p_numthreads == p->p_suspcount) {
 464                                         thread_lock(p->p_singlethread);
 465                                         thread_unsuspend_one(p->p_singlethread);
 466                                         thread_unlock(p->p_singlethread);
 467                                 }
 468                         }
 469
 470                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 471                         PCPU_SET(deadthread, td);
 472                 } else {
 473                         /*
 474                          * The last thread is exiting.. but not through exit()
 475                          * what should we do?
 476                          * Theoretically this can't happen
 477                          * exit1() - clears threading flags before coming here
 478                          * kse_exit() - treats last thread specially
 479                          * thr_exit() - treats last thread specially
 480                          * ifdef KSE
 481                          * thread_user_enter() - only if more exist
 482                          * thread_userret() - only if more exist
 483                          * endif
 484                          * thread_suspend_check() - only if more exist
 485                          */
 486                         panic ("thread_exit: Last thread exiting on its own");
 487                 }
 488         }
 489         PROC_UNLOCK(p);
 490         thread_lock(td);
 491         /* Save our tick information with both the thread and proc locked */
 492         ruxagg(&p->p_rux, td);
 493         PROC_SUNLOCK(p);
 494         td->td_state = TDS_INACTIVE;
 495         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 496         sched_throw(td);
 497         panic("I'm a teapot!");
 498         /* NOTREACHED */
 499 }
 500
 501 /*
 502  * Do any thread specific cleanups that may be needed in wait()
 503  * called with Giant, proc and schedlock not held.
 504  */
 505 void
 506 thread_wait(struct proc *p)
 507 {
 508         struct thread *td;
 509
 510         mtx_assert(&Giant, MA_NOTOWNED);
 511         KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
 512         td = FIRST_THREAD_IN_PROC(p);
 513 #ifdef KSE
 514         if (td->td_standin != NULL) {
 515                 if (td->td_standin->td_ucred != NULL) {
 516                         crfree(td->td_standin->td_ucred);
 517                         td->td_standin->td_ucred = NULL;
 518                 }
 519                 thread_free(td->td_standin);
 520                 td->td_standin = NULL;
 521         }
 522 #endif
 523         /* Lock the last thread so we spin until it exits cpu_throw(). */
 524         thread_lock(td);
 525         thread_unlock(td);
 526         /* Wait for any remaining threads to exit cpu_throw(). */
 527         while (p->p_exitthreads)
 528                 sched_relinquish(curthread);
 529         cpu_thread_clean(td);
 530         crfree(td->td_ucred);
 531         thread_reap();  /* check for zombie threads etc. */
 532 }
 533
 534 /*
 535  * Link a thread to a process.
 536  * set up anything that needs to be initialized for it to
 537  * be used by the process.
 538  *
 539  * Note that we do not link to the proc's ucred here.
 540  * The thread is linked as if running but no KSE assigned.
 541  * Called from:
 542  *  proc_linkup()
 543  *  thread_schedule_upcall()
 544  *  thr_create()
 545  */
 546 void
 547 thread_link(struct thread *td, struct proc *p)
 548 {
 549
 550         /*
 551          * XXX This can't be enabled because it's called for proc0 before
 552          * it's spinlock has been created.
 553          * PROC_SLOCK_ASSERT(p, MA_OWNED);
 554          */
 555         td->td_state    = TDS_INACTIVE;
 556         td->td_proc     = p;
 557         td->td_flags    = TDF_INMEM;
 558
 559         LIST_INIT(&td->td_contested);
 560         LIST_INIT(&td->td_lprof[0]);
 561         LIST_INIT(&td->td_lprof[1]);
 562         sigqueue_init(&td->td_sigqueue, p);
 563         callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
 564         TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
 565         p->p_numthreads++;
 566 }
 567
 568 /*
 569  * Convert a process with one thread to an unthreaded process.
 570  * Called from:
 571  *  thread_single(exit)  (called from execve and exit)
 572  *  kse_exit()          XXX may need cleaning up wrt KSE stuff
 573  */
 574 void
 575 thread_unthread(struct thread *td)
 576 {
 577         struct proc *p = td->td_proc;
 578
 579         KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
 580 #ifdef KSE
 581         thread_lock(td);
 582         upcall_remove(td);
 583         thread_unlock(td);
 584         p->p_flag &= ~(P_SA|P_HADTHREADS);
 585         td->td_mailbox = NULL;
 586         td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
 587         if (td->td_standin != NULL) {
 588                 thread_zombie(td->td_standin);
 589                 td->td_standin = NULL;
 590         }
 591 #else
 592         p->p_flag &= ~P_HADTHREADS;
 593 #endif
 594 }
 595
 596 /*
 597  * Called from:
 598  *  thread_exit()
 599  */
 600 void
 601 thread_unlink(struct thread *td)
 602 {
 603         struct proc *p = td->td_proc;
 604
 605         PROC_SLOCK_ASSERT(p, MA_OWNED);
 606         TAILQ_REMOVE(&p->p_threads, td, td_plist);
 607         p->p_numthreads--;
 608         /* could clear a few other things here */
 609         /* Must  NOT clear links to proc! */
 610 }
 611
 612 /*
 613  * Enforce single-threading.
 614  *
 615  * Returns 1 if the caller must abort (another thread is waiting to
 616  * exit the process or similar). Process is locked!
 617  * Returns 0 when you are successfully the only thread running.
 618  * A process has successfully single threaded in the suspend mode when
 619  * There are no threads in user mode. Threads in the kernel must be
 620  * allowed to continue until they get to the user boundary. They may even
 621  * copy out their return values and data before suspending. They may however be
 622  * accelerated in reaching the user boundary as we will wake up
 623  * any sleeping threads that are interruptable. (PCATCH).
 624  */
 625 int
 626 thread_single(int mode)
 627 {
 628         struct thread *td;
 629         struct thread *td2;
 630         struct proc *p;
 631         int remaining;
 632
 633         td = curthread;
 634         p = td->td_proc;
 635         mtx_assert(&Giant, MA_NOTOWNED);
 636         PROC_LOCK_ASSERT(p, MA_OWNED);
 637         KASSERT((td != NULL), ("curthread is NULL"));
 638
 639         if ((p->p_flag & P_HADTHREADS) == 0)
 640                 return (0);
 641
 642         /* Is someone already single threading? */
 643         if (p->p_singlethread != NULL && p->p_singlethread != td)
 644                 return (1);
 645
 646         if (mode == SINGLE_EXIT) {
 647                 p->p_flag |= P_SINGLE_EXIT;
 648                 p->p_flag &= ~P_SINGLE_BOUNDARY;
 649         } else {
 650                 p->p_flag &= ~P_SINGLE_EXIT;
 651                 if (mode == SINGLE_BOUNDARY)
 652                         p->p_flag |= P_SINGLE_BOUNDARY;
 653                 else
 654                         p->p_flag &= ~P_SINGLE_BOUNDARY;
 655         }
 656         p->p_flag |= P_STOPPED_SINGLE;
 657         PROC_SLOCK(p);
 658         p->p_singlethread = td;
 659         if (mode == SINGLE_EXIT)
 660                 remaining = p->p_numthreads;
 661         else if (mode == SINGLE_BOUNDARY)
 662                 remaining = p->p_numthreads - p->p_boundary_count;
 663         else
 664                 remaining = p->p_numthreads - p->p_suspcount;
 665         while (remaining != 1) {
 666                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
 667                         goto stopme;
 668                 FOREACH_THREAD_IN_PROC(p, td2) {
 669                         if (td2 == td)
 670                                 continue;
 671                         thread_lock(td2);
 672                         td2->td_flags |= TDF_ASTPENDING;
 673                         if (TD_IS_INHIBITED(td2)) {
 674                                 switch (mode) {
 675                                 case SINGLE_EXIT:
 676                                         if (td->td_flags & TDF_DBSUSPEND)
 677                                                 td->td_flags &= ~TDF_DBSUSPEND;
 678                                         if (TD_IS_SUSPENDED(td2))
 679                                                 thread_unsuspend_one(td2);
 680                                         if (TD_ON_SLEEPQ(td2) &&
 681                                             (td2->td_flags & TDF_SINTR))
 682                                                 sleepq_abort(td2, EINTR);
 683                                         break;
 684                                 case SINGLE_BOUNDARY:
 685                                         break;
 686                                 default:
 687                                         if (TD_IS_SUSPENDED(td2)) {
 688                                                 thread_unlock(td2);
 689                                                 continue;
 690                                         }
 691                                         /*
 692                                          * maybe other inhibited states too?
 693                                          */
 694                                         if ((td2->td_flags & TDF_SINTR) &&
 695                                             (td2->td_inhibitors &
 696                                             (TDI_SLEEPING | TDI_SWAPPED)))
 697                                                 thread_suspend_one(td2);
 698                                         break;
 699                                 }
 700                         }
 701 #ifdef SMP
 702                         else if (TD_IS_RUNNING(td2) && td != td2) {
 703                                 forward_signal(td2);
 704                         }
 705 #endif
 706                         thread_unlock(td2);
 707                 }
 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                 /*
 716                  * Maybe we suspended some threads.. was it enough?
 717                  */
 718                 if (remaining == 1)
 719                         break;
 720
 721 stopme:
 722                 /*
 723                  * Wake us up when everyone else has suspended.
 724                  * In the mean time we suspend as well.
 725                  */
 726                 thread_suspend_switch(td);
 727                 if (mode == SINGLE_EXIT)
 728                         remaining = p->p_numthreads;
 729                 else if (mode == SINGLE_BOUNDARY)
 730                         remaining = p->p_numthreads - p->p_boundary_count;
 731                 else
 732                         remaining = p->p_numthreads - p->p_suspcount;
 733         }
 734         if (mode == SINGLE_EXIT) {
 735                 /*
 736                  * We have gotten rid of all the other threads and we
 737                  * are about to either exit or exec. In either case,
 738                  * we try our utmost  to revert to being a non-threaded
 739                  * process.
 740                  */
 741                 p->p_singlethread = NULL;
 742                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
 743                 thread_unthread(td);
 744         }
 745         PROC_SUNLOCK(p);
 746         return (0);
 747 }
 748
 749 /*
 750  * Called in from locations that can safely check to see
 751  * whether we have to suspend or at least throttle for a
 752  * single-thread event (e.g. fork).
 753  *
 754  * Such locations include userret().
 755  * If the "return_instead" argument is non zero, the thread must be able to
 756  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
 757  *
 758  * The 'return_instead' argument tells the function if it may do a
 759  * thread_exit() or suspend, or whether the caller must abort and back
 760  * out instead.
 761  *
 762  * If the thread that set the single_threading request has set the
 763  * P_SINGLE_EXIT bit in the process flags then this call will never return
 764  * if 'return_instead' is false, but will exit.
 765  *
 766  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
 767  *---------------+--------------------+---------------------
 768  *       0       | returns 0          |   returns 0 or 1
 769  *               | when ST ends       |   immediatly
 770  *---------------+--------------------+---------------------
 771  *       1       | thread exits       |   returns 1
 772  *               |                    |  immediatly
 773  * 0 = thread_exit() or suspension ok,
 774  * other = return error instead of stopping the thread.
 775  *
 776  * While a full suspension is under effect, even a single threading
 777  * thread would be suspended if it made this call (but it shouldn't).
 778  * This call should only be made from places where
 779  * thread_exit() would be safe as that may be the outcome unless
 780  * return_instead is set.
 781  */
 782 int
 783 thread_suspend_check(int return_instead)
 784 {
 785         struct thread *td;
 786         struct proc *p;
 787
 788         td = curthread;
 789         p = td->td_proc;
 790         mtx_assert(&Giant, MA_NOTOWNED);
 791         PROC_LOCK_ASSERT(p, MA_OWNED);
 792         while (P_SHOULDSTOP(p) ||
 793               ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
 794                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 795                         KASSERT(p->p_singlethread != NULL,
 796                             ("singlethread not set"));
 797                         /*
 798                          * The only suspension in action is a
 799                          * single-threading. Single threader need not stop.
 800                          * XXX Should be safe to access unlocked
 801                          * as it can only be set to be true by us.
 802                          */
 803                         if (p->p_singlethread == td)
 804                                 return (0);     /* Exempt from stopping. */
 805                 }
 806                 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
 807                         return (EINTR);
 808
 809                 /* Should we goto user boundary if we didn't come from there? */
 810                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
 811                     (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
 812                         return (ERESTART);
 813
 814                 /* If thread will exit, flush its pending signals */
 815                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 816                         sigqueue_flush(&td->td_sigqueue);
 817
 818                 PROC_SLOCK(p);
 819                 thread_stopped(p);
 820                 /*
 821                  * If the process is waiting for us to exit,
 822                  * this thread should just suicide.
 823                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
 824                  */
 825                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 826                         thread_exit();
 827                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 828                         if (p->p_numthreads == p->p_suspcount + 1) {
 829                                 thread_lock(p->p_singlethread);
 830                                 thread_unsuspend_one(p->p_singlethread);
 831                                 thread_unlock(p->p_singlethread);
 832                         }
 833                 }
 834                 PROC_UNLOCK(p);
 835                 thread_lock(td);
 836                 /*
 837                  * When a thread suspends, it just
 838                  * gets taken off all queues.
 839                  */
 840                 thread_suspend_one(td);
 841                 if (return_instead == 0) {
 842                         p->p_boundary_count++;
 843                         td->td_flags |= TDF_BOUNDARY;
 844                 }
 845                 PROC_SUNLOCK(p);
 846                 mi_switch(SW_INVOL, NULL);
 847                 if (return_instead == 0)
 848                         td->td_flags &= ~TDF_BOUNDARY;
 849                 thread_unlock(td);
 850                 PROC_LOCK(p);
 851                 if (return_instead == 0)
 852                         p->p_boundary_count--;
 853         }
 854         return (0);
 855 }
 856
 857 void
 858 thread_suspend_switch(struct thread *td)
 859 {
 860         struct proc *p;
 861
 862         p = td->td_proc;
 863         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 864         PROC_LOCK_ASSERT(p, MA_OWNED);
 865         PROC_SLOCK_ASSERT(p, MA_OWNED);
 866         /*
 867          * We implement thread_suspend_one in stages here to avoid
 868          * dropping the proc lock while the thread lock is owned.
 869          */
 870         thread_stopped(p);
 871         p->p_suspcount++;
 872         PROC_UNLOCK(p);
 873         thread_lock(td);
 874         sched_sleep(td);
 875         TD_SET_SUSPENDED(td);
 876         PROC_SUNLOCK(p);
 877         DROP_GIANT();
 878         mi_switch(SW_VOL, NULL);
 879         thread_unlock(td);
 880         PICKUP_GIANT();
 881         PROC_LOCK(p);
 882         PROC_SLOCK(p);
 883 }
 884
 885 void
 886 thread_suspend_one(struct thread *td)
 887 {
 888         struct proc *p = td->td_proc;
 889
 890         PROC_SLOCK_ASSERT(p, MA_OWNED);
 891         THREAD_LOCK_ASSERT(td, MA_OWNED);
 892         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 893         p->p_suspcount++;
 894         sched_sleep(td);
 895         TD_SET_SUSPENDED(td);
 896 }
 897
 898 void
 899 thread_unsuspend_one(struct thread *td)
 900 {
 901         struct proc *p = td->td_proc;
 902
 903         PROC_SLOCK_ASSERT(p, MA_OWNED);
 904         THREAD_LOCK_ASSERT(td, MA_OWNED);
 905         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
 906         TD_CLR_SUSPENDED(td);
 907         p->p_suspcount--;
 908         setrunnable(td);
 909 }
 910
 911 /*
 912  * Allow all threads blocked by single threading to continue running.
 913  */
 914 void
 915 thread_unsuspend(struct proc *p)
 916 {
 917         struct thread *td;
 918
 919         PROC_LOCK_ASSERT(p, MA_OWNED);
 920         PROC_SLOCK_ASSERT(p, MA_OWNED);
 921         if (!P_SHOULDSTOP(p)) {
 922                 FOREACH_THREAD_IN_PROC(p, td) {
 923                         thread_lock(td);
 924                         if (TD_IS_SUSPENDED(td)) {
 925                                 thread_unsuspend_one(td);
 926                         }
 927                         thread_unlock(td);
 928                 }
 929         } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
 930             (p->p_numthreads == p->p_suspcount)) {
 931                 /*
 932                  * Stopping everything also did the job for the single
 933                  * threading request. Now we've downgraded to single-threaded,
 934                  * let it continue.
 935                  */
 936                 thread_lock(p->p_singlethread);
 937                 thread_unsuspend_one(p->p_singlethread);
 938                 thread_unlock(p->p_singlethread);
 939         }
 940 }
 941
 942 /*
 943  * End the single threading mode..
 944  */
 945 void
 946 thread_single_end(void)
 947 {
 948         struct thread *td;
 949         struct proc *p;
 950
 951         td = curthread;
 952         p = td->td_proc;
 953         PROC_LOCK_ASSERT(p, MA_OWNED);
 954         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
 955         PROC_SLOCK(p);
 956         p->p_singlethread = NULL;
 957         /*
 958          * If there are other threads they mey now run,
 959          * unless of course there is a blanket 'stop order'
 960          * on the process. The single threader must be allowed
 961          * to continue however as this is a bad place to stop.
 962          */
 963         if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
 964                 FOREACH_THREAD_IN_PROC(p, td) {
 965                         thread_lock(td);
 966                         if (TD_IS_SUSPENDED(td)) {
 967                                 thread_unsuspend_one(td);
 968                         }
 969                         thread_unlock(td);
 970                 }
 971         }
 972         PROC_SUNLOCK(p);
 973 }
 974
 975 struct thread *
 976 thread_find(struct proc *p, lwpid_t tid)
 977 {
 978         struct thread *td;
 979
 980         PROC_LOCK_ASSERT(p, MA_OWNED);
 981         PROC_SLOCK(p);
 982         FOREACH_THREAD_IN_PROC(p, td) {
 983                 if (td->td_tid == tid)
 984                         break;
 985         }
 986         PROC_SUNLOCK(p);
 987         return (td);
 988 }