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 #include <sys/eventhandler.h>
  53
  54 /*
  55  * thread related storage.
  56  */
  57 static uma_zone_t thread_zone;
  58
  59 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
  60
  61 int max_threads_per_proc = 1500;
  62 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
  63         &max_threads_per_proc, 0, "Limit on threads per proc");
  64
  65 int max_threads_hits;
  66 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
  67         &max_threads_hits, 0, "");
  68
  69 #ifdef KSE
  70 int virtual_cpu;
  71
  72 #endif
  73 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
  74 static struct mtx zombie_lock;
  75 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
  76
  77 static void thread_zombie(struct thread *);
  78
  79 #ifdef KSE
  80 static int
  81 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
  82 {
  83         int error, new_val;
  84         int def_val;
  85
  86         def_val = mp_ncpus;
  87         if (virtual_cpu == 0)
  88                 new_val = def_val;
  89         else
  90                 new_val = virtual_cpu;
  91         error = sysctl_handle_int(oidp, &new_val, 0, req);
  92         if (error != 0 || req->newptr == NULL)
  93                 return (error);
  94         if (new_val < 0)
  95                 return (EINVAL);
  96         virtual_cpu = new_val;
  97         return (0);
  98 }
  99
 100 /* DEBUG ONLY */
 101 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
 102         0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
 103         "debug virtual cpus");
 104 #endif
 105
 106 struct mtx tid_lock;
 107 static struct unrhdr *tid_unrhdr;
 108
 109 /*
 110  * Prepare a thread for use.
 111  */
 112 static int
 113 thread_ctor(void *mem, int size, void *arg, int flags)
 114 {
 115         struct thread   *td;
 116
 117         td = (struct thread *)mem;
 118         td->td_state = TDS_INACTIVE;
 119         td->td_oncpu = NOCPU;
 120
 121         td->td_tid = alloc_unr(tid_unrhdr);
 122         td->td_syscalls = 0;
 123
 124         /*
 125          * Note that td_critnest begins life as 1 because the thread is not
 126          * running and is thereby implicitly waiting to be on the receiving
 127          * end of a context switch.
 128          */
 129         td->td_critnest = 1;
 130         EVENTHANDLER_INVOKE(thread_ctor, td);
 131 #ifdef AUDIT
 132         audit_thread_alloc(td);
 133 #endif
 134         umtx_thread_alloc(td);
 135         return (0);
 136 }
 137
 138 /*
 139  * Reclaim a thread after use.
 140  */
 141 static void
 142 thread_dtor(void *mem, int size, void *arg)
 143 {
 144         struct thread *td;
 145
 146         td = (struct thread *)mem;
 147
 148 #ifdef INVARIANTS
 149         /* Verify that this thread is in a safe state to free. */
 150         switch (td->td_state) {
 151         case TDS_INHIBITED:
 152         case TDS_RUNNING:
 153         case TDS_CAN_RUN:
 154         case TDS_RUNQ:
 155                 /*
 156                  * We must never unlink a thread that is in one of
 157                  * these states, because it is currently active.
 158                  */
 159                 panic("bad state for thread unlinking");
 160                 /* NOTREACHED */
 161         case TDS_INACTIVE:
 162                 break;
 163         default:
 164                 panic("bad thread state");
 165                 /* NOTREACHED */
 166         }
 167 #endif
 168 #ifdef AUDIT
 169         audit_thread_free(td);
 170 #endif
 171         EVENTHANDLER_INVOKE(thread_dtor, td);
 172         free_unr(tid_unrhdr, td->td_tid);
 173         sched_newthread(td);
 174 }
 175
 176 /*
 177  * Initialize type-stable parts of a thread (when newly created).
 178  */
 179 static int
 180 thread_init(void *mem, int size, int flags)
 181 {
 182         struct thread *td;
 183
 184         td = (struct thread *)mem;
 185
 186         td->td_sleepqueue = sleepq_alloc();
 187         td->td_turnstile = turnstile_alloc();
 188         EVENTHANDLER_INVOKE(thread_init, td);
 189         td->td_sched = (struct td_sched *)&td[1];
 190         sched_newthread(td);
 191         umtx_thread_init(td);
 192         td->td_kstack = 0;
 193         return (0);
 194 }
 195
 196 /*
 197  * Tear down type-stable parts of a thread (just before being discarded).
 198  */
 199 static void
 200 thread_fini(void *mem, int size)
 201 {
 202         struct thread *td;
 203
 204         td = (struct thread *)mem;
 205         EVENTHANDLER_INVOKE(thread_fini, td);
 206         turnstile_free(td->td_turnstile);
 207         sleepq_free(td->td_sleepqueue);
 208         umtx_thread_fini(td);
 209 }
 210
 211 /*
 212  * For a newly created process,
 213  * link up all the structures and its initial threads etc.
 214  * called from:
 215  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
 216  * proc_dtor() (should go away)
 217  * proc_init()
 218  */
 219 void
 220 proc_linkup0(struct proc *p, struct thread *td)
 221 {
 222         TAILQ_INIT(&p->p_threads);           /* all threads in proc */
 223         proc_linkup(p, td);
 224 }
 225
 226 void
 227 proc_linkup(struct proc *p, struct thread *td)
 228 {
 229
 230 #ifdef KSE
 231         TAILQ_INIT(&p->p_upcalls);           /* upcall list */
 232 #endif
 233         sigqueue_init(&p->p_sigqueue, p);
 234         p->p_ksi = ksiginfo_alloc(1);
 235         if (p->p_ksi != NULL) {
 236                 /* XXX p_ksi may be null if ksiginfo zone is not ready */
 237                 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
 238         }
 239         LIST_INIT(&p->p_mqnotifier);
 240         p->p_numthreads = 0;
 241         thread_link(td, p);
 242 }
 243
 244 /*
 245  * Initialize global thread allocation resources.
 246  */
 247 void
 248 threadinit(void)
 249 {
 250
 251         mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
 252         tid_unrhdr = new_unrhdr(PID_MAX + 1, INT_MAX, &tid_lock);
 253
 254         thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
 255             thread_ctor, thread_dtor, thread_init, thread_fini,
 256             16 - 1, 0);
 257 #ifdef KSE
 258         kseinit();      /* set up kse specific stuff  e.g. upcall zone*/
 259 #endif
 260 }
 261
 262 /*
 263  * Place an unused thread on the zombie list.
 264  * Use the slpq as that must be unused by now.
 265  */
 266 void
 267 thread_zombie(struct thread *td)
 268 {
 269         mtx_lock_spin(&zombie_lock);
 270         TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
 271         mtx_unlock_spin(&zombie_lock);
 272 }
 273
 274 /*
 275  * Release a thread that has exited after cpu_throw().
 276  */
 277 void
 278 thread_stash(struct thread *td)
 279 {
 280         atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
 281         thread_zombie(td);
 282 }
 283
 284 /*
 285  * Reap zombie kse resource.
 286  */
 287 void
 288 thread_reap(void)
 289 {
 290         struct thread *td_first, *td_next;
 291
 292         /*
 293          * Don't even bother to lock if none at this instant,
 294          * we really don't care about the next instant..
 295          */
 296         if (!TAILQ_EMPTY(&zombie_threads)) {
 297                 mtx_lock_spin(&zombie_lock);
 298                 td_first = TAILQ_FIRST(&zombie_threads);
 299                 if (td_first)
 300                         TAILQ_INIT(&zombie_threads);
 301                 mtx_unlock_spin(&zombie_lock);
 302                 while (td_first) {
 303                         td_next = TAILQ_NEXT(td_first, td_slpq);
 304                         if (td_first->td_ucred)
 305                                 crfree(td_first->td_ucred);
 306                         thread_free(td_first);
 307                         td_first = td_next;
 308                 }
 309         }
 310 #ifdef KSE
 311         upcall_reap();
 312 #endif
 313 }
 314
 315 /*
 316  * Allocate a thread.
 317  */
 318 struct thread *
 319 thread_alloc(void)
 320 {
 321         struct thread *td;
 322
 323         thread_reap(); /* check if any zombies to get */
 324
 325         td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
 326         KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
 327         if (!vm_thread_new(td, 0)) {
 328                 uma_zfree(thread_zone, td);
 329                 return (NULL);
 330         }
 331         cpu_thread_alloc(td);
 332         return (td);
 333 }
 334
 335
 336 /*
 337  * Deallocate a thread.
 338  */
 339 void
 340 thread_free(struct thread *td)
 341 {
 342
 343         cpu_thread_free(td);
 344         if (td->td_altkstack != 0)
 345                 vm_thread_dispose_altkstack(td);
 346         if (td->td_kstack != 0)
 347                 vm_thread_dispose(td);
 348         uma_zfree(thread_zone, td);
 349 }
 350
 351 /*
 352  * Discard the current thread and exit from its context.
 353  * Always called with scheduler locked.
 354  *
 355  * Because we can't free a thread while we're operating under its context,
 356  * push the current thread into our CPU's deadthread holder. This means
 357  * we needn't worry about someone else grabbing our context before we
 358  * do a cpu_throw().  This may not be needed now as we are under schedlock.
 359  * Maybe we can just do a thread_stash() as thr_exit1 does.
 360  */
 361 /*  XXX
 362  * libthr expects its thread exit to return for the last
 363  * thread, meaning that the program is back to non-threaded
 364  * mode I guess. Because we do this (cpu_throw) unconditionally
 365  * here, they have their own version of it. (thr_exit1())
 366  * that doesn't do it all if this was the last thread.
 367  * It is also called from thread_suspend_check().
 368  * Of course in the end, they end up coming here through exit1
 369  * anyhow..  After fixing 'thr' to play by the rules we should be able
 370  * to merge these two functions together.
 371  *
 372  * called from:
 373  * exit1()
 374  * kse_exit()
 375  * thr_exit()
 376  * ifdef KSE
 377  * thread_user_enter()
 378  * thread_userret()
 379  * endif
 380  * thread_suspend_check()
 381  */
 382 void
 383 thread_exit(void)
 384 {
 385         uint64_t new_switchtime;
 386         struct thread *td;
 387         struct thread *td2;
 388         struct proc *p;
 389
 390         td = curthread;
 391         p = td->td_proc;
 392
 393         PROC_SLOCK_ASSERT(p, MA_OWNED);
 394         mtx_assert(&Giant, MA_NOTOWNED);
 395
 396         PROC_LOCK_ASSERT(p, MA_OWNED);
 397         KASSERT(p != NULL, ("thread exiting without a process"));
 398         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 399             (long)p->p_pid, td->td_name);
 400         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 401
 402 #ifdef AUDIT
 403         AUDIT_SYSCALL_EXIT(0, td);
 404 #endif
 405
 406 #ifdef KSE
 407         if (td->td_standin != NULL) {
 408                 /*
 409                  * Note that we don't need to free the cred here as it
 410                  * is done in thread_reap().
 411                  */
 412                 thread_zombie(td->td_standin);
 413                 td->td_standin = NULL;
 414         }
 415 #endif
 416
 417         umtx_thread_exit(td);
 418
 419         /*
 420          * drop FPU & debug register state storage, or any other
 421          * architecture specific resources that
 422          * would not be on a new untouched process.
 423          */
 424         cpu_thread_exit(td);    /* XXXSMP */
 425
 426         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 427         new_switchtime = cpu_ticks();
 428         p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
 429         PCPU_SET(switchtime, new_switchtime);
 430         PCPU_SET(switchticks, ticks);
 431         PCPU_INC(cnt.v_swtch);
 432         /* Save our resource usage in our process. */
 433         td->td_ru.ru_nvcsw++;
 434         rucollect(&p->p_ru, &td->td_ru);
 435         /*
 436          * The last thread is left attached to the process
 437          * So that the whole bundle gets recycled. Skip
 438          * all this stuff if we never had threads.
 439          * EXIT clears all sign of other threads when
 440          * it goes to single threading, so the last thread always
 441          * takes the short path.
 442          */
 443         if (p->p_flag & P_HADTHREADS) {
 444                 if (p->p_numthreads > 1) {
 445                         thread_lock(td);
 446 #ifdef KSE
 447                         kse_unlink(td);
 448 #else
 449                         thread_unlink(td);
 450 #endif
 451                         thread_unlock(td);
 452                         td2 = FIRST_THREAD_IN_PROC(p);
 453                         sched_exit_thread(td2, td);
 454
 455                         /*
 456                          * The test below is NOT true if we are the
 457                          * sole exiting thread. P_STOPPED_SNGL is unset
 458                          * in exit1() after it is the only survivor.
 459                          */
 460                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 461                                 if (p->p_numthreads == p->p_suspcount) {
 462                                         thread_lock(p->p_singlethread);
 463                                         thread_unsuspend_one(p->p_singlethread);
 464                                         thread_unlock(p->p_singlethread);
 465                                 }
 466                         }
 467
 468                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 469                         PCPU_SET(deadthread, td);
 470                 } else {
 471                         /*
 472                          * The last thread is exiting.. but not through exit()
 473                          * what should we do?
 474                          * Theoretically this can't happen
 475                          * exit1() - clears threading flags before coming here
 476                          * kse_exit() - treats last thread specially
 477                          * thr_exit() - treats last thread specially
 478                          * ifdef KSE
 479                          * thread_user_enter() - only if more exist
 480                          * thread_userret() - only if more exist
 481                          * endif
 482                          * thread_suspend_check() - only if more exist
 483                          */
 484                         panic ("thread_exit: Last thread exiting on its own");
 485                 }
 486         }
 487         PROC_UNLOCK(p);
 488         thread_lock(td);
 489         /* Save our tick information with both the thread and proc locked */
 490         ruxagg(&p->p_rux, td);
 491         PROC_SUNLOCK(p);
 492         td->td_state = TDS_INACTIVE;
 493         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 494         sched_throw(td);
 495         panic("I'm a teapot!");
 496         /* NOTREACHED */
 497 }
 498
 499 /*
 500  * Do any thread specific cleanups that may be needed in wait()
 501  * called with Giant, proc and schedlock not held.
 502  */
 503 void
 504 thread_wait(struct proc *p)
 505 {
 506         struct thread *td;
 507
 508         mtx_assert(&Giant, MA_NOTOWNED);
 509         KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
 510         td = FIRST_THREAD_IN_PROC(p);
 511 #ifdef KSE
 512         if (td->td_standin != NULL) {
 513                 if (td->td_standin->td_ucred != NULL) {
 514                         crfree(td->td_standin->td_ucred);
 515                         td->td_standin->td_ucred = NULL;
 516                 }
 517                 thread_free(td->td_standin);
 518                 td->td_standin = NULL;
 519         }
 520 #endif
 521         /* Lock the last thread so we spin until it exits cpu_throw(). */
 522         thread_lock(td);
 523         thread_unlock(td);
 524         /* Wait for any remaining threads to exit cpu_throw(). */
 525         while (p->p_exitthreads)
 526                 sched_relinquish(curthread);
 527         cpu_thread_clean(td);
 528         crfree(td->td_ucred);
 529         thread_reap();  /* check for zombie threads etc. */
 530 }
 531
 532 /*
 533  * Link a thread to a process.
 534  * set up anything that needs to be initialized for it to
 535  * be used by the process.
 536  *
 537  * Note that we do not link to the proc's ucred here.
 538  * The thread is linked as if running but no KSE assigned.
 539  * Called from:
 540  *  proc_linkup()
 541  *  thread_schedule_upcall()
 542  *  thr_create()
 543  */
 544 void
 545 thread_link(struct thread *td, struct proc *p)
 546 {
 547
 548         /*
 549          * XXX This can't be enabled because it's called for proc0 before
 550          * it's spinlock has been created.
 551          * PROC_SLOCK_ASSERT(p, MA_OWNED);
 552          */
 553         td->td_state    = TDS_INACTIVE;
 554         td->td_proc     = p;
 555         td->td_flags    = TDF_INMEM;
 556
 557         LIST_INIT(&td->td_contested);
 558         sigqueue_init(&td->td_sigqueue, p);
 559         callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
 560         TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
 561         p->p_numthreads++;
 562 }
 563
 564 /*
 565  * Convert a process with one thread to an unthreaded process.
 566  * Called from:
 567  *  thread_single(exit)  (called from execve and exit)
 568  *  kse_exit()          XXX may need cleaning up wrt KSE stuff
 569  */
 570 void
 571 thread_unthread(struct thread *td)
 572 {
 573         struct proc *p = td->td_proc;
 574
 575         KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
 576 #ifdef KSE
 577         thread_lock(td);
 578         upcall_remove(td);
 579         thread_unlock(td);
 580         p->p_flag &= ~(P_SA|P_HADTHREADS);
 581         td->td_mailbox = NULL;
 582         td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
 583         if (td->td_standin != NULL) {
 584                 thread_zombie(td->td_standin);
 585                 td->td_standin = NULL;
 586         }
 587 #else
 588         p->p_flag &= ~P_HADTHREADS;
 589 #endif
 590 }
 591
 592 /*
 593  * Called from:
 594  *  thread_exit()
 595  */
 596 void
 597 thread_unlink(struct thread *td)
 598 {
 599         struct proc *p = td->td_proc;
 600
 601         PROC_SLOCK_ASSERT(p, MA_OWNED);
 602         TAILQ_REMOVE(&p->p_threads, td, td_plist);
 603         p->p_numthreads--;
 604         /* could clear a few other things here */
 605         /* Must  NOT clear links to proc! */
 606 }
 607
 608 /*
 609  * Enforce single-threading.
 610  *
 611  * Returns 1 if the caller must abort (another thread is waiting to
 612  * exit the process or similar). Process is locked!
 613  * Returns 0 when you are successfully the only thread running.
 614  * A process has successfully single threaded in the suspend mode when
 615  * There are no threads in user mode. Threads in the kernel must be
 616  * allowed to continue until they get to the user boundary. They may even
 617  * copy out their return values and data before suspending. They may however be
 618  * accelerated in reaching the user boundary as we will wake up
 619  * any sleeping threads that are interruptable. (PCATCH).
 620  */
 621 int
 622 thread_single(int mode)
 623 {
 624         struct thread *td;
 625         struct thread *td2;
 626         struct proc *p;
 627         int remaining;
 628
 629         td = curthread;
 630         p = td->td_proc;
 631         mtx_assert(&Giant, MA_NOTOWNED);
 632         PROC_LOCK_ASSERT(p, MA_OWNED);
 633         KASSERT((td != NULL), ("curthread is NULL"));
 634
 635         if ((p->p_flag & P_HADTHREADS) == 0)
 636                 return (0);
 637
 638         /* Is someone already single threading? */
 639         if (p->p_singlethread != NULL && p->p_singlethread != td)
 640                 return (1);
 641
 642         if (mode == SINGLE_EXIT) {
 643                 p->p_flag |= P_SINGLE_EXIT;
 644                 p->p_flag &= ~P_SINGLE_BOUNDARY;
 645         } else {
 646                 p->p_flag &= ~P_SINGLE_EXIT;
 647                 if (mode == SINGLE_BOUNDARY)
 648                         p->p_flag |= P_SINGLE_BOUNDARY;
 649                 else
 650                         p->p_flag &= ~P_SINGLE_BOUNDARY;
 651         }
 652         p->p_flag |= P_STOPPED_SINGLE;
 653         PROC_SLOCK(p);
 654         p->p_singlethread = td;
 655         if (mode == SINGLE_EXIT)
 656                 remaining = p->p_numthreads;
 657         else if (mode == SINGLE_BOUNDARY)
 658                 remaining = p->p_numthreads - p->p_boundary_count;
 659         else
 660                 remaining = p->p_numthreads - p->p_suspcount;
 661         while (remaining != 1) {
 662                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
 663                         goto stopme;
 664                 FOREACH_THREAD_IN_PROC(p, td2) {
 665                         if (td2 == td)
 666                                 continue;
 667                         thread_lock(td2);
 668                         td2->td_flags |= TDF_ASTPENDING;
 669                         if (TD_IS_INHIBITED(td2)) {
 670                                 switch (mode) {
 671                                 case SINGLE_EXIT:
 672                                         if (td->td_flags & TDF_DBSUSPEND)
 673                                                 td->td_flags &= ~TDF_DBSUSPEND;
 674                                         if (TD_IS_SUSPENDED(td2))
 675                                                 thread_unsuspend_one(td2);
 676                                         if (TD_ON_SLEEPQ(td2) &&
 677                                             (td2->td_flags & TDF_SINTR))
 678                                                 sleepq_abort(td2, EINTR);
 679                                         break;
 680                                 case SINGLE_BOUNDARY:
 681                                         break;
 682                                 default:
 683                                         if (TD_IS_SUSPENDED(td2)) {
 684                                                 thread_unlock(td2);
 685                                                 continue;
 686                                         }
 687                                         /*
 688                                          * maybe other inhibited states too?
 689                                          */
 690                                         if ((td2->td_flags & TDF_SINTR) &&
 691                                             (td2->td_inhibitors &
 692                                             (TDI_SLEEPING | TDI_SWAPPED)))
 693                                                 thread_suspend_one(td2);
 694                                         break;
 695                                 }
 696                         }
 697 #ifdef SMP
 698                         else if (TD_IS_RUNNING(td2) && td != td2) {
 699                                 forward_signal(td2);
 700                         }
 701 #endif
 702                         thread_unlock(td2);
 703                 }
 704                 if (mode == SINGLE_EXIT)
 705                         remaining = p->p_numthreads;
 706                 else if (mode == SINGLE_BOUNDARY)
 707                         remaining = p->p_numthreads - p->p_boundary_count;
 708                 else
 709                         remaining = p->p_numthreads - p->p_suspcount;
 710
 711                 /*
 712                  * Maybe we suspended some threads.. was it enough?
 713                  */
 714                 if (remaining == 1)
 715                         break;
 716
 717 stopme:
 718                 /*
 719                  * Wake us up when everyone else has suspended.
 720                  * In the mean time we suspend as well.
 721                  */
 722                 thread_suspend_switch(td);
 723                 if (mode == SINGLE_EXIT)
 724                         remaining = p->p_numthreads;
 725                 else if (mode == SINGLE_BOUNDARY)
 726                         remaining = p->p_numthreads - p->p_boundary_count;
 727                 else
 728                         remaining = p->p_numthreads - p->p_suspcount;
 729         }
 730         if (mode == SINGLE_EXIT) {
 731                 /*
 732                  * We have gotten rid of all the other threads and we
 733                  * are about to either exit or exec. In either case,
 734                  * we try our utmost  to revert to being a non-threaded
 735                  * process.
 736                  */
 737                 p->p_singlethread = NULL;
 738                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
 739                 thread_unthread(td);
 740         }
 741         PROC_SUNLOCK(p);
 742         return (0);
 743 }
 744
 745 /*
 746  * Called in from locations that can safely check to see
 747  * whether we have to suspend or at least throttle for a
 748  * single-thread event (e.g. fork).
 749  *
 750  * Such locations include userret().
 751  * If the "return_instead" argument is non zero, the thread must be able to
 752  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
 753  *
 754  * The 'return_instead' argument tells the function if it may do a
 755  * thread_exit() or suspend, or whether the caller must abort and back
 756  * out instead.
 757  *
 758  * If the thread that set the single_threading request has set the
 759  * P_SINGLE_EXIT bit in the process flags then this call will never return
 760  * if 'return_instead' is false, but will exit.
 761  *
 762  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
 763  *---------------+--------------------+---------------------
 764  *       0       | returns 0          |   returns 0 or 1
 765  *               | when ST ends       |   immediatly
 766  *---------------+--------------------+---------------------
 767  *       1       | thread exits       |   returns 1
 768  *               |                    |  immediatly
 769  * 0 = thread_exit() or suspension ok,
 770  * other = return error instead of stopping the thread.
 771  *
 772  * While a full suspension is under effect, even a single threading
 773  * thread would be suspended if it made this call (but it shouldn't).
 774  * This call should only be made from places where
 775  * thread_exit() would be safe as that may be the outcome unless
 776  * return_instead is set.
 777  */
 778 int
 779 thread_suspend_check(int return_instead)
 780 {
 781         struct thread *td;
 782         struct proc *p;
 783
 784         td = curthread;
 785         p = td->td_proc;
 786         mtx_assert(&Giant, MA_NOTOWNED);
 787         PROC_LOCK_ASSERT(p, MA_OWNED);
 788         while (P_SHOULDSTOP(p) ||
 789               ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
 790                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 791                         KASSERT(p->p_singlethread != NULL,
 792                             ("singlethread not set"));
 793                         /*
 794                          * The only suspension in action is a
 795                          * single-threading. Single threader need not stop.
 796                          * XXX Should be safe to access unlocked
 797                          * as it can only be set to be true by us.
 798                          */
 799                         if (p->p_singlethread == td)
 800                                 return (0);     /* Exempt from stopping. */
 801                 }
 802                 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
 803                         return (EINTR);
 804
 805                 /* Should we goto user boundary if we didn't come from there? */
 806                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
 807                     (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
 808                         return (ERESTART);
 809
 810                 /* If thread will exit, flush its pending signals */
 811                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 812                         sigqueue_flush(&td->td_sigqueue);
 813
 814                 PROC_SLOCK(p);
 815                 thread_stopped(p);
 816                 /*
 817                  * If the process is waiting for us to exit,
 818                  * this thread should just suicide.
 819                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
 820                  */
 821                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 822                         thread_exit();
 823                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 824                         if (p->p_numthreads == p->p_suspcount + 1) {
 825                                 thread_lock(p->p_singlethread);
 826                                 thread_unsuspend_one(p->p_singlethread);
 827                                 thread_unlock(p->p_singlethread);
 828                         }
 829                 }
 830                 PROC_UNLOCK(p);
 831                 thread_lock(td);
 832                 /*
 833                  * When a thread suspends, it just
 834                  * gets taken off all queues.
 835                  */
 836                 thread_suspend_one(td);
 837                 if (return_instead == 0) {
 838                         p->p_boundary_count++;
 839                         td->td_flags |= TDF_BOUNDARY;
 840                 }
 841                 PROC_SUNLOCK(p);
 842                 mi_switch(SW_INVOL, NULL);
 843                 if (return_instead == 0)
 844                         td->td_flags &= ~TDF_BOUNDARY;
 845                 thread_unlock(td);
 846                 PROC_LOCK(p);
 847                 if (return_instead == 0)
 848                         p->p_boundary_count--;
 849         }
 850         return (0);
 851 }
 852
 853 void
 854 thread_suspend_switch(struct thread *td)
 855 {
 856         struct proc *p;
 857
 858         p = td->td_proc;
 859         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 860         PROC_LOCK_ASSERT(p, MA_OWNED);
 861         PROC_SLOCK_ASSERT(p, MA_OWNED);
 862         /*
 863          * We implement thread_suspend_one in stages here to avoid
 864          * dropping the proc lock while the thread lock is owned.
 865          */
 866         thread_stopped(p);
 867         p->p_suspcount++;
 868         PROC_UNLOCK(p);
 869         thread_lock(td);
 870         sched_sleep(td);
 871         TD_SET_SUSPENDED(td);
 872         PROC_SUNLOCK(p);
 873         DROP_GIANT();
 874         mi_switch(SW_VOL, NULL);
 875         thread_unlock(td);
 876         PICKUP_GIANT();
 877         PROC_LOCK(p);
 878         PROC_SLOCK(p);
 879 }
 880
 881 void
 882 thread_suspend_one(struct thread *td)
 883 {
 884         struct proc *p = td->td_proc;
 885
 886         PROC_SLOCK_ASSERT(p, MA_OWNED);
 887         THREAD_LOCK_ASSERT(td, MA_OWNED);
 888         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 889         p->p_suspcount++;
 890         sched_sleep(td);
 891         TD_SET_SUSPENDED(td);
 892 }
 893
 894 void
 895 thread_unsuspend_one(struct thread *td)
 896 {
 897         struct proc *p = td->td_proc;
 898
 899         PROC_SLOCK_ASSERT(p, MA_OWNED);
 900         THREAD_LOCK_ASSERT(td, MA_OWNED);
 901         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
 902         TD_CLR_SUSPENDED(td);
 903         p->p_suspcount--;
 904         setrunnable(td);
 905 }
 906
 907 /*
 908  * Allow all threads blocked by single threading to continue running.
 909  */
 910 void
 911 thread_unsuspend(struct proc *p)
 912 {
 913         struct thread *td;
 914
 915         PROC_LOCK_ASSERT(p, MA_OWNED);
 916         PROC_SLOCK_ASSERT(p, MA_OWNED);
 917         if (!P_SHOULDSTOP(p)) {
 918                 FOREACH_THREAD_IN_PROC(p, td) {
 919                         thread_lock(td);
 920                         if (TD_IS_SUSPENDED(td)) {
 921                                 thread_unsuspend_one(td);
 922                         }
 923                         thread_unlock(td);
 924                 }
 925         } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
 926             (p->p_numthreads == p->p_suspcount)) {
 927                 /*
 928                  * Stopping everything also did the job for the single
 929                  * threading request. Now we've downgraded to single-threaded,
 930                  * let it continue.
 931                  */
 932                 thread_lock(p->p_singlethread);
 933                 thread_unsuspend_one(p->p_singlethread);
 934                 thread_unlock(p->p_singlethread);
 935         }
 936 }
 937
 938 /*
 939  * End the single threading mode..
 940  */
 941 void
 942 thread_single_end(void)
 943 {
 944         struct thread *td;
 945         struct proc *p;
 946
 947         td = curthread;
 948         p = td->td_proc;
 949         PROC_LOCK_ASSERT(p, MA_OWNED);
 950         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
 951         PROC_SLOCK(p);
 952         p->p_singlethread = NULL;
 953         /*
 954          * If there are other threads they mey now run,
 955          * unless of course there is a blanket 'stop order'
 956          * on the process. The single threader must be allowed
 957          * to continue however as this is a bad place to stop.
 958          */
 959         if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
 960                 FOREACH_THREAD_IN_PROC(p, td) {
 961                         thread_lock(td);
 962                         if (TD_IS_SUSPENDED(td)) {
 963                                 thread_unsuspend_one(td);
 964                         }
 965                         thread_unlock(td);
 966                 }
 967         }
 968         PROC_SUNLOCK(p);
 969 }
 970
 971 struct thread *
 972 thread_find(struct proc *p, lwpid_t tid)
 973 {
 974         struct thread *td;
 975
 976         PROC_LOCK_ASSERT(p, MA_OWNED);
 977         PROC_SLOCK(p);
 978         FOREACH_THREAD_IN_PROC(p, td) {
 979                 if (td->td_tid == tid)
 980                         break;
 981         }
 982         PROC_SUNLOCK(p);
 983         return (td);
 984 }