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 #include <sys/cpuset.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         td->td_incruntime = 0;
 125
 126         /*
 127          * Note that td_critnest begins life as 1 because the thread is not
 128          * running and is thereby implicitly waiting to be on the receiving
 129          * end of a context switch.
 130          */
 131         td->td_critnest = 1;
 132         EVENTHANDLER_INVOKE(thread_ctor, td);
 133 #ifdef AUDIT
 134         audit_thread_alloc(td);
 135 #endif
 136         umtx_thread_alloc(td);
 137         return (0);
 138 }
 139
 140 /*
 141  * Reclaim a thread after use.
 142  */
 143 static void
 144 thread_dtor(void *mem, int size, void *arg)
 145 {
 146         struct thread *td;
 147
 148         td = (struct thread *)mem;
 149
 150 #ifdef INVARIANTS
 151         /* Verify that this thread is in a safe state to free. */
 152         switch (td->td_state) {
 153         case TDS_INHIBITED:
 154         case TDS_RUNNING:
 155         case TDS_CAN_RUN:
 156         case TDS_RUNQ:
 157                 /*
 158                  * We must never unlink a thread that is in one of
 159                  * these states, because it is currently active.
 160                  */
 161                 panic("bad state for thread unlinking");
 162                 /* NOTREACHED */
 163         case TDS_INACTIVE:
 164                 break;
 165         default:
 166                 panic("bad thread state");
 167                 /* NOTREACHED */
 168         }
 169 #endif
 170 #ifdef AUDIT
 171         audit_thread_free(td);
 172 #endif
 173         EVENTHANDLER_INVOKE(thread_dtor, td);
 174         free_unr(tid_unrhdr, td->td_tid);
 175         sched_newthread(td);
 176 }
 177
 178 /*
 179  * Initialize type-stable parts of a thread (when newly created).
 180  */
 181 static int
 182 thread_init(void *mem, int size, int flags)
 183 {
 184         struct thread *td;
 185
 186         td = (struct thread *)mem;
 187
 188         td->td_sleepqueue = sleepq_alloc();
 189         td->td_turnstile = turnstile_alloc();
 190         EVENTHANDLER_INVOKE(thread_init, td);
 191         td->td_sched = (struct td_sched *)&td[1];
 192         sched_newthread(td);
 193         umtx_thread_init(td);
 194         td->td_kstack = 0;
 195         td->td_fpop = NULL;
 196         return (0);
 197 }
 198
 199 /*
 200  * Tear down type-stable parts of a thread (just before being discarded).
 201  */
 202 static void
 203 thread_fini(void *mem, int size)
 204 {
 205         struct thread *td;
 206
 207         td = (struct thread *)mem;
 208         EVENTHANDLER_INVOKE(thread_fini, td);
 209         turnstile_free(td->td_turnstile);
 210         sleepq_free(td->td_sleepqueue);
 211         umtx_thread_fini(td);
 212 }
 213
 214 /*
 215  * For a newly created process,
 216  * link up all the structures and its initial threads etc.
 217  * called from:
 218  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
 219  * proc_dtor() (should go away)
 220  * proc_init()
 221  */
 222 void
 223 proc_linkup0(struct proc *p, struct thread *td)
 224 {
 225         TAILQ_INIT(&p->p_threads);           /* all threads in proc */
 226         proc_linkup(p, td);
 227 }
 228
 229 void
 230 proc_linkup(struct proc *p, struct thread *td)
 231 {
 232
 233 #ifdef KSE
 234         TAILQ_INIT(&p->p_upcalls);           /* upcall list */
 235 #endif
 236         sigqueue_init(&p->p_sigqueue, p);
 237         p->p_ksi = ksiginfo_alloc(1);
 238         if (p->p_ksi != NULL) {
 239                 /* XXX p_ksi may be null if ksiginfo zone is not ready */
 240                 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
 241         }
 242         LIST_INIT(&p->p_mqnotifier);
 243         p->p_numthreads = 0;
 244         thread_link(td, p);
 245 }
 246
 247 /*
 248  * Initialize global thread allocation resources.
 249  */
 250 void
 251 threadinit(void)
 252 {
 253
 254         mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
 255         tid_unrhdr = new_unrhdr(PID_MAX + 1, 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         if (td->td_cpuset)
 346                 cpuset_rel(td->td_cpuset);
 347         td->td_cpuset = NULL;
 348         cpu_thread_free(td);
 349         if (td->td_altkstack != 0)
 350                 vm_thread_dispose_altkstack(td);
 351         if (td->td_kstack != 0)
 352                 vm_thread_dispose(td);
 353         uma_zfree(thread_zone, td);
 354 }
 355
 356 /*
 357  * Discard the current thread and exit from its context.
 358  * Always called with scheduler locked.
 359  *
 360  * Because we can't free a thread while we're operating under its context,
 361  * push the current thread into our CPU's deadthread holder. This means
 362  * we needn't worry about someone else grabbing our context before we
 363  * do a cpu_throw().  This may not be needed now as we are under schedlock.
 364  * Maybe we can just do a thread_stash() as thr_exit1 does.
 365  */
 366 /*  XXX
 367  * libthr expects its thread exit to return for the last
 368  * thread, meaning that the program is back to non-threaded
 369  * mode I guess. Because we do this (cpu_throw) unconditionally
 370  * here, they have their own version of it. (thr_exit1())
 371  * that doesn't do it all if this was the last thread.
 372  * It is also called from thread_suspend_check().
 373  * Of course in the end, they end up coming here through exit1
 374  * anyhow..  After fixing 'thr' to play by the rules we should be able
 375  * to merge these two functions together.
 376  *
 377  * called from:
 378  * exit1()
 379  * kse_exit()
 380  * thr_exit()
 381  * ifdef KSE
 382  * thread_user_enter()
 383  * thread_userret()
 384  * endif
 385  * thread_suspend_check()
 386  */
 387 void
 388 thread_exit(void)
 389 {
 390         uint64_t new_switchtime;
 391         struct thread *td;
 392         struct thread *td2;
 393         struct proc *p;
 394         int wakeup_swapper;
 395
 396         td = curthread;
 397         p = td->td_proc;
 398
 399         PROC_SLOCK_ASSERT(p, MA_OWNED);
 400         mtx_assert(&Giant, MA_NOTOWNED);
 401
 402         PROC_LOCK_ASSERT(p, MA_OWNED);
 403         KASSERT(p != NULL, ("thread exiting without a process"));
 404         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 405             (long)p->p_pid, p->p_comm);
 406         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 407
 408 #ifdef AUDIT
 409         AUDIT_SYSCALL_EXIT(0, td);
 410 #endif
 411
 412 #ifdef KSE
 413         if (td->td_standin != NULL) {
 414                 /*
 415                  * Note that we don't need to free the cred here as it
 416                  * is done in thread_reap().
 417                  */
 418                 thread_zombie(td->td_standin);
 419                 td->td_standin = NULL;
 420         }
 421 #endif
 422
 423         umtx_thread_exit(td);
 424
 425         /*
 426          * drop FPU & debug register state storage, or any other
 427          * architecture specific resources that
 428          * would not be on a new untouched process.
 429          */
 430         cpu_thread_exit(td);    /* XXXSMP */
 431
 432         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 433         new_switchtime = cpu_ticks();
 434         p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
 435         PCPU_SET(switchtime, new_switchtime);
 436         PCPU_SET(switchticks, ticks);
 437         PCPU_INC(cnt.v_swtch);
 438         /* Save our resource usage in our process. */
 439         td->td_ru.ru_nvcsw++;
 440         rucollect(&p->p_ru, &td->td_ru);
 441         /*
 442          * The last thread is left attached to the process
 443          * So that the whole bundle gets recycled. Skip
 444          * all this stuff if we never had threads.
 445          * EXIT clears all sign of other threads when
 446          * it goes to single threading, so the last thread always
 447          * takes the short path.
 448          */
 449         if (p->p_flag & P_HADTHREADS) {
 450                 if (p->p_numthreads > 1) {
 451                         thread_lock(td);
 452 #ifdef KSE
 453                         kse_unlink(td);
 454 #else
 455                         thread_unlink(td);
 456 #endif
 457                         thread_unlock(td);
 458                         td2 = FIRST_THREAD_IN_PROC(p);
 459                         sched_exit_thread(td2, td);
 460
 461                         /*
 462                          * The test below is NOT true if we are the
 463                          * sole exiting thread. P_STOPPED_SNGL is unset
 464                          * in exit1() after it is the only survivor.
 465                          */
 466                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 467                                 if (p->p_numthreads == p->p_suspcount) {
 468                                         thread_lock(p->p_singlethread);
 469                                         wakeup_swapper = thread_unsuspend_one(
 470                                                 p->p_singlethread);
 471                                         thread_unlock(p->p_singlethread);
 472                                         if (wakeup_swapper)
 473                                                 kick_proc0();
 474                                 }
 475                         }
 476
 477                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 478                         PCPU_SET(deadthread, td);
 479                 } else {
 480                         /*
 481                          * The last thread is exiting.. but not through exit()
 482                          * what should we do?
 483                          * Theoretically this can't happen
 484                          * exit1() - clears threading flags before coming here
 485                          * kse_exit() - treats last thread specially
 486                          * thr_exit() - treats last thread specially
 487                          * ifdef KSE
 488                          * thread_user_enter() - only if more exist
 489                          * thread_userret() - only if more exist
 490                          * endif
 491                          * thread_suspend_check() - only if more exist
 492                          */
 493                         panic ("thread_exit: Last thread exiting on its own");
 494                 }
 495         }
 496         PROC_UNLOCK(p);
 497         thread_lock(td);
 498         /* Save our tick information with both the thread and proc locked */
 499         ruxagg(&p->p_rux, td);
 500         PROC_SUNLOCK(p);
 501         td->td_state = TDS_INACTIVE;
 502         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 503         sched_throw(td);
 504         panic("I'm a teapot!");
 505         /* NOTREACHED */
 506 }
 507
 508 /*
 509  * Do any thread specific cleanups that may be needed in wait()
 510  * called with Giant, proc and schedlock not held.
 511  */
 512 void
 513 thread_wait(struct proc *p)
 514 {
 515         struct thread *td;
 516
 517         mtx_assert(&Giant, MA_NOTOWNED);
 518         KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
 519         td = FIRST_THREAD_IN_PROC(p);
 520 #ifdef KSE
 521         if (td->td_standin != NULL) {
 522                 if (td->td_standin->td_ucred != NULL) {
 523                         crfree(td->td_standin->td_ucred);
 524                         td->td_standin->td_ucred = NULL;
 525                 }
 526                 thread_free(td->td_standin);
 527                 td->td_standin = NULL;
 528         }
 529 #endif
 530         /* Lock the last thread so we spin until it exits cpu_throw(). */
 531         thread_lock(td);
 532         thread_unlock(td);
 533         /* Wait for any remaining threads to exit cpu_throw(). */
 534         while (p->p_exitthreads)
 535                 sched_relinquish(curthread);
 536         cpuset_rel(td->td_cpuset);
 537         td->td_cpuset = NULL;
 538         cpu_thread_clean(td);
 539         crfree(td->td_ucred);
 540         thread_reap();  /* check for zombie threads etc. */
 541 }
 542
 543 /*
 544  * Link a thread to a process.
 545  * set up anything that needs to be initialized for it to
 546  * be used by the process.
 547  *
 548  * Note that we do not link to the proc's ucred here.
 549  * The thread is linked as if running but no KSE assigned.
 550  * Called from:
 551  *  proc_linkup()
 552  *  thread_schedule_upcall()
 553  *  thr_create()
 554  */
 555 void
 556 thread_link(struct thread *td, struct proc *p)
 557 {
 558
 559         /*
 560          * XXX This can't be enabled because it's called for proc0 before
 561          * it's spinlock has been created.
 562          * PROC_SLOCK_ASSERT(p, MA_OWNED);
 563          */
 564         td->td_state    = TDS_INACTIVE;
 565         td->td_proc     = p;
 566         td->td_flags    = TDF_INMEM;
 567
 568         LIST_INIT(&td->td_contested);
 569         sigqueue_init(&td->td_sigqueue, p);
 570         callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
 571         TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
 572         p->p_numthreads++;
 573 }
 574
 575 /*
 576  * Convert a process with one thread to an unthreaded process.
 577  * Called from:
 578  *  thread_single(exit)  (called from execve and exit)
 579  *  kse_exit()          XXX may need cleaning up wrt KSE stuff
 580  */
 581 void
 582 thread_unthread(struct thread *td)
 583 {
 584         struct proc *p = td->td_proc;
 585
 586         KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
 587 #ifdef KSE
 588         thread_lock(td);
 589         upcall_remove(td);
 590         thread_unlock(td);
 591         p->p_flag &= ~(P_SA|P_HADTHREADS);
 592         td->td_mailbox = NULL;
 593         td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
 594         if (td->td_standin != NULL) {
 595                 thread_zombie(td->td_standin);
 596                 td->td_standin = NULL;
 597         }
 598 #else
 599         p->p_flag &= ~P_HADTHREADS;
 600 #endif
 601 }
 602
 603 /*
 604  * Called from:
 605  *  thread_exit()
 606  */
 607 void
 608 thread_unlink(struct thread *td)
 609 {
 610         struct proc *p = td->td_proc;
 611
 612         PROC_SLOCK_ASSERT(p, MA_OWNED);
 613         TAILQ_REMOVE(&p->p_threads, td, td_plist);
 614         p->p_numthreads--;
 615         /* could clear a few other things here */
 616         /* Must  NOT clear links to proc! */
 617 }
 618
 619 /*
 620  * Enforce single-threading.
 621  *
 622  * Returns 1 if the caller must abort (another thread is waiting to
 623  * exit the process or similar). Process is locked!
 624  * Returns 0 when you are successfully the only thread running.
 625  * A process has successfully single threaded in the suspend mode when
 626  * There are no threads in user mode. Threads in the kernel must be
 627  * allowed to continue until they get to the user boundary. They may even
 628  * copy out their return values and data before suspending. They may however be
 629  * accelerated in reaching the user boundary as we will wake up
 630  * any sleeping threads that are interruptable. (PCATCH).
 631  */
 632 int
 633 thread_single(int mode)
 634 {
 635         struct thread *td;
 636         struct thread *td2;
 637         struct proc *p;
 638         int remaining, wakeup_swapper;
 639
 640         td = curthread;
 641         p = td->td_proc;
 642         mtx_assert(&Giant, MA_NOTOWNED);
 643         PROC_LOCK_ASSERT(p, MA_OWNED);
 644         KASSERT((td != NULL), ("curthread is NULL"));
 645
 646         if ((p->p_flag & P_HADTHREADS) == 0)
 647                 return (0);
 648
 649         /* Is someone already single threading? */
 650         if (p->p_singlethread != NULL && p->p_singlethread != td)
 651                 return (1);
 652
 653         if (mode == SINGLE_EXIT) {
 654                 p->p_flag |= P_SINGLE_EXIT;
 655                 p->p_flag &= ~P_SINGLE_BOUNDARY;
 656         } else {
 657                 p->p_flag &= ~P_SINGLE_EXIT;
 658                 if (mode == SINGLE_BOUNDARY)
 659                         p->p_flag |= P_SINGLE_BOUNDARY;
 660                 else
 661                         p->p_flag &= ~P_SINGLE_BOUNDARY;
 662         }
 663         p->p_flag |= P_STOPPED_SINGLE;
 664         PROC_SLOCK(p);
 665         p->p_singlethread = td;
 666         if (mode == SINGLE_EXIT)
 667                 remaining = p->p_numthreads;
 668         else if (mode == SINGLE_BOUNDARY)
 669                 remaining = p->p_numthreads - p->p_boundary_count;
 670         else
 671                 remaining = p->p_numthreads - p->p_suspcount;
 672         while (remaining != 1) {
 673                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
 674                         goto stopme;
 675                 wakeup_swapper = 0;
 676                 FOREACH_THREAD_IN_PROC(p, td2) {
 677                         if (td2 == td)
 678                                 continue;
 679                         thread_lock(td2);
 680                         td2->td_flags |= TDF_ASTPENDING;
 681                         if (TD_IS_INHIBITED(td2)) {
 682                                 switch (mode) {
 683                                 case SINGLE_EXIT:
 684                                         if (td->td_flags & TDF_DBSUSPEND)
 685                                                 td->td_flags &= ~TDF_DBSUSPEND;
 686                                         if (TD_IS_SUSPENDED(td2))
 687                                                 wakeup_swapper |=
 688                                                     thread_unsuspend_one(td2);
 689                                         if (TD_ON_SLEEPQ(td2) &&
 690                                             (td2->td_flags & TDF_SINTR))
 691                                                 wakeup_swapper |=
 692                                                     sleepq_abort(td2, EINTR);
 693                                         break;
 694                                 case SINGLE_BOUNDARY:
 695                                         if (TD_IS_SUSPENDED(td2) &&
 696                                             !(td2->td_flags & TDF_BOUNDARY))
 697                                                 wakeup_swapper |=
 698                                                     thread_unsuspend_one(td2);
 699                                         if (TD_ON_SLEEPQ(td2) &&
 700                                             (td2->td_flags & TDF_SINTR))
 701                                                 wakeup_swapper |=
 702                                                     sleepq_abort(td2, ERESTART);
 703                                         break;
 704                                 default:
 705                                         if (TD_IS_SUSPENDED(td2)) {
 706                                                 thread_unlock(td2);
 707                                                 continue;
 708                                         }
 709                                         /*
 710                                          * maybe other inhibited states too?
 711                                          */
 712                                         if ((td2->td_flags & TDF_SINTR) &&
 713                                             (td2->td_inhibitors &
 714                                             (TDI_SLEEPING | TDI_SWAPPED)))
 715                                                 thread_suspend_one(td2);
 716                                         break;
 717                                 }
 718                         }
 719 #ifdef SMP
 720                         else if (TD_IS_RUNNING(td2) && td != td2) {
 721                                 forward_signal(td2);
 722                         }
 723 #endif
 724                         thread_unlock(td2);
 725                 }
 726                 if (wakeup_swapper)
 727                         kick_proc0();
 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                 /*
 736                  * Maybe we suspended some threads.. was it enough?
 737                  */
 738                 if (remaining == 1)
 739                         break;
 740
 741 stopme:
 742                 /*
 743                  * Wake us up when everyone else has suspended.
 744                  * In the mean time we suspend as well.
 745                  */
 746                 thread_suspend_switch(td);
 747                 if (mode == SINGLE_EXIT)
 748                         remaining = p->p_numthreads;
 749                 else if (mode == SINGLE_BOUNDARY)
 750                         remaining = p->p_numthreads - p->p_boundary_count;
 751                 else
 752                         remaining = p->p_numthreads - p->p_suspcount;
 753         }
 754         if (mode == SINGLE_EXIT) {
 755                 /*
 756                  * We have gotten rid of all the other threads and we
 757                  * are about to either exit or exec. In either case,
 758                  * we try our utmost  to revert to being a non-threaded
 759                  * process.
 760                  */
 761                 p->p_singlethread = NULL;
 762                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
 763                 thread_unthread(td);
 764         }
 765         PROC_SUNLOCK(p);
 766         return (0);
 767 }
 768
 769 /*
 770  * Called in from locations that can safely check to see
 771  * whether we have to suspend or at least throttle for a
 772  * single-thread event (e.g. fork).
 773  *
 774  * Such locations include userret().
 775  * If the "return_instead" argument is non zero, the thread must be able to
 776  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
 777  *
 778  * The 'return_instead' argument tells the function if it may do a
 779  * thread_exit() or suspend, or whether the caller must abort and back
 780  * out instead.
 781  *
 782  * If the thread that set the single_threading request has set the
 783  * P_SINGLE_EXIT bit in the process flags then this call will never return
 784  * if 'return_instead' is false, but will exit.
 785  *
 786  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
 787  *---------------+--------------------+---------------------
 788  *       0       | returns 0          |   returns 0 or 1
 789  *               | when ST ends       |   immediatly
 790  *---------------+--------------------+---------------------
 791  *       1       | thread exits       |   returns 1
 792  *               |                    |  immediatly
 793  * 0 = thread_exit() or suspension ok,
 794  * other = return error instead of stopping the thread.
 795  *
 796  * While a full suspension is under effect, even a single threading
 797  * thread would be suspended if it made this call (but it shouldn't).
 798  * This call should only be made from places where
 799  * thread_exit() would be safe as that may be the outcome unless
 800  * return_instead is set.
 801  */
 802 int
 803 thread_suspend_check(int return_instead)
 804 {
 805         struct thread *td;
 806         struct proc *p;
 807         int wakeup_swapper;
 808
 809         td = curthread;
 810         p = td->td_proc;
 811         mtx_assert(&Giant, MA_NOTOWNED);
 812         PROC_LOCK_ASSERT(p, MA_OWNED);
 813         while (P_SHOULDSTOP(p) ||
 814               ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
 815                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 816                         KASSERT(p->p_singlethread != NULL,
 817                             ("singlethread not set"));
 818                         /*
 819                          * The only suspension in action is a
 820                          * single-threading. Single threader need not stop.
 821                          * XXX Should be safe to access unlocked
 822                          * as it can only be set to be true by us.
 823                          */
 824                         if (p->p_singlethread == td)
 825                                 return (0);     /* Exempt from stopping. */
 826                 }
 827                 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
 828                         return (EINTR);
 829
 830                 /* Should we goto user boundary if we didn't come from there? */
 831                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
 832                     (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
 833                         return (ERESTART);
 834
 835                 /* If thread will exit, flush its pending signals */
 836                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 837                         sigqueue_flush(&td->td_sigqueue);
 838
 839                 PROC_SLOCK(p);
 840                 thread_stopped(p);
 841                 /*
 842                  * If the process is waiting for us to exit,
 843                  * this thread should just suicide.
 844                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
 845                  */
 846                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 847                         thread_exit();
 848                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 849                         if (p->p_numthreads == p->p_suspcount + 1) {
 850                                 thread_lock(p->p_singlethread);
 851                                 wakeup_swapper =
 852                                     thread_unsuspend_one(p->p_singlethread);
 853                                 thread_unlock(p->p_singlethread);
 854                                 if (wakeup_swapper)
 855                                         kick_proc0();
 856                         }
 857                 }
 858                 PROC_UNLOCK(p);
 859                 thread_lock(td);
 860                 /*
 861                  * When a thread suspends, it just
 862                  * gets taken off all queues.
 863                  */
 864                 thread_suspend_one(td);
 865                 if (return_instead == 0) {
 866                         p->p_boundary_count++;
 867                         td->td_flags |= TDF_BOUNDARY;
 868                 }
 869                 PROC_SUNLOCK(p);
 870                 mi_switch(SW_INVOL, NULL);
 871                 if (return_instead == 0)
 872                         td->td_flags &= ~TDF_BOUNDARY;
 873                 thread_unlock(td);
 874                 PROC_LOCK(p);
 875                 if (return_instead == 0)
 876                         p->p_boundary_count--;
 877         }
 878         return (0);
 879 }
 880
 881 void
 882 thread_suspend_switch(struct thread *td)
 883 {
 884         struct proc *p;
 885
 886         p = td->td_proc;
 887         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 888         PROC_LOCK_ASSERT(p, MA_OWNED);
 889         PROC_SLOCK_ASSERT(p, MA_OWNED);
 890         /*
 891          * We implement thread_suspend_one in stages here to avoid
 892          * dropping the proc lock while the thread lock is owned.
 893          */
 894         thread_stopped(p);
 895         p->p_suspcount++;
 896         PROC_UNLOCK(p);
 897         thread_lock(td);
 898         sched_sleep(td);
 899         TD_SET_SUSPENDED(td);
 900         PROC_SUNLOCK(p);
 901         DROP_GIANT();
 902         mi_switch(SW_VOL, NULL);
 903         thread_unlock(td);
 904         PICKUP_GIANT();
 905         PROC_LOCK(p);
 906         PROC_SLOCK(p);
 907 }
 908
 909 void
 910 thread_suspend_one(struct thread *td)
 911 {
 912         struct proc *p = td->td_proc;
 913
 914         PROC_SLOCK_ASSERT(p, MA_OWNED);
 915         THREAD_LOCK_ASSERT(td, MA_OWNED);
 916         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 917         p->p_suspcount++;
 918         sched_sleep(td);
 919         TD_SET_SUSPENDED(td);
 920 }
 921
 922 int
 923 thread_unsuspend_one(struct thread *td)
 924 {
 925         struct proc *p = td->td_proc;
 926
 927         PROC_SLOCK_ASSERT(p, MA_OWNED);
 928         THREAD_LOCK_ASSERT(td, MA_OWNED);
 929         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
 930         TD_CLR_SUSPENDED(td);
 931         p->p_suspcount--;
 932         return (setrunnable(td));
 933 }
 934
 935 /*
 936  * Allow all threads blocked by single threading to continue running.
 937  */
 938 void
 939 thread_unsuspend(struct proc *p)
 940 {
 941         struct thread *td;
 942         int wakeup_swapper;
 943
 944         PROC_LOCK_ASSERT(p, MA_OWNED);
 945         PROC_SLOCK_ASSERT(p, MA_OWNED);
 946         wakeup_swapper = 0;
 947         if (!P_SHOULDSTOP(p)) {
 948                 FOREACH_THREAD_IN_PROC(p, td) {
 949                         thread_lock(td);
 950                         if (TD_IS_SUSPENDED(td)) {
 951                                 wakeup_swapper |= thread_unsuspend_one(td);
 952                         }
 953                         thread_unlock(td);
 954                 }
 955         } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
 956             (p->p_numthreads == p->p_suspcount)) {
 957                 /*
 958                  * Stopping everything also did the job for the single
 959                  * threading request. Now we've downgraded to single-threaded,
 960                  * let it continue.
 961                  */
 962                 thread_lock(p->p_singlethread);
 963                 wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
 964                 thread_unlock(p->p_singlethread);
 965         }
 966         if (wakeup_swapper)
 967                 kick_proc0();
 968 }
 969
 970 /*
 971  * End the single threading mode..
 972  */
 973 void
 974 thread_single_end(void)
 975 {
 976         struct thread *td;
 977         struct proc *p;
 978         int wakeup_swapper;
 979
 980         td = curthread;
 981         p = td->td_proc;
 982         PROC_LOCK_ASSERT(p, MA_OWNED);
 983         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
 984         PROC_SLOCK(p);
 985         p->p_singlethread = NULL;
 986         wakeup_swapper = 0;
 987         /*
 988          * If there are other threads they may now run,
 989          * unless of course there is a blanket 'stop order'
 990          * on the process. The single threader must be allowed
 991          * to continue however as this is a bad place to stop.
 992          */
 993         if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
 994                 FOREACH_THREAD_IN_PROC(p, td) {
 995                         thread_lock(td);
 996                         if (TD_IS_SUSPENDED(td)) {
 997                                 wakeup_swapper |= thread_unsuspend_one(td);
 998                         }
 999                         thread_unlock(td);
1000                 }
1001         }
1002         PROC_SUNLOCK(p);
1003         if (wakeup_swapper)
1004                 kick_proc0();
1005 }
1006
1007 struct thread *
1008 thread_find(struct proc *p, lwpid_t tid)
1009 {
1010         struct thread *td;
1011
1012         PROC_LOCK_ASSERT(p, MA_OWNED);
1013         PROC_SLOCK(p);
1014         FOREACH_THREAD_IN_PROC(p, td) {
1015                 if (td->td_tid == tid)
1016                         break;
1017         }
1018         PROC_SUNLOCK(p);
1019         return (td);
1020 }