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 "opt_witness.h"
  30 #include "opt_kdtrace.h"
  31 #include "opt_hwpmc_hooks.h"
  32
  33 #include <sys/cdefs.h>
  34 __FBSDID("$FreeBSD$");
  35
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/kernel.h>
  39 #include <sys/lock.h>
  40 #include <sys/mutex.h>
  41 #include <sys/proc.h>
  42 #include <sys/rangelock.h>
  43 #include <sys/resourcevar.h>
  44 #include <sys/sdt.h>
  45 #include <sys/smp.h>
  46 #include <sys/sched.h>
  47 #include <sys/sleepqueue.h>
  48 #include <sys/selinfo.h>
  49 #include <sys/turnstile.h>
  50 #include <sys/ktr.h>
  51 #include <sys/rwlock.h>
  52 #include <sys/umtx.h>
  53 #include <sys/cpuset.h>
  54 #ifdef  HWPMC_HOOKS
  55 #include <sys/pmckern.h>
  56 #endif
  57
  58 #include <security/audit/audit.h>
  59
  60 #include <vm/vm.h>
  61 #include <vm/vm_extern.h>
  62 #include <vm/uma.h>
  63 #include <sys/eventhandler.h>
  64
  65 SDT_PROVIDER_DECLARE(proc);
  66 SDT_PROBE_DEFINE(proc, , , lwp_exit, lwp-exit);
  67
  68
  69 /*
  70  * thread related storage.
  71  */
  72 static uma_zone_t thread_zone;
  73
  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 #define TID_BUFFER_SIZE 1024
  81
  82 struct mtx tid_lock;
  83 static struct unrhdr *tid_unrhdr;
  84 static lwpid_t tid_buffer[TID_BUFFER_SIZE];
  85 static int tid_head, tid_tail;
  86 static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash");
  87
  88 struct  tidhashhead *tidhashtbl;
  89 u_long  tidhash;
  90 struct  rwlock tidhash_lock;
  91
  92 static lwpid_t
  93 tid_alloc(void)
  94 {
  95         lwpid_t tid;
  96
  97         tid = alloc_unr(tid_unrhdr);
  98         if (tid != -1)
  99                 return (tid);
 100         mtx_lock(&tid_lock);
 101         if (tid_head == tid_tail) {
 102                 mtx_unlock(&tid_lock);
 103                 return (-1);
 104         }
 105         tid = tid_buffer[tid_head];
 106         tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
 107         mtx_unlock(&tid_lock);
 108         return (tid);
 109 }
 110
 111 static void
 112 tid_free(lwpid_t tid)
 113 {
 114         lwpid_t tmp_tid = -1;
 115
 116         mtx_lock(&tid_lock);
 117         if ((tid_tail + 1) % TID_BUFFER_SIZE == tid_head) {
 118                 tmp_tid = tid_buffer[tid_head];
 119                 tid_head = (tid_head + 1) % TID_BUFFER_SIZE;
 120         }
 121         tid_buffer[tid_tail] = tid;
 122         tid_tail = (tid_tail + 1) % TID_BUFFER_SIZE;
 123         mtx_unlock(&tid_lock);
 124         if (tmp_tid != -1)
 125                 free_unr(tid_unrhdr, tmp_tid);
 126 }
 127
 128 /*
 129  * Prepare a thread for use.
 130  */
 131 static int
 132 thread_ctor(void *mem, int size, void *arg, int flags)
 133 {
 134         struct thread   *td;
 135
 136         td = (struct thread *)mem;
 137         td->td_state = TDS_INACTIVE;
 138         td->td_oncpu = NOCPU;
 139
 140         td->td_tid = tid_alloc();
 141
 142         /*
 143          * Note that td_critnest begins life as 1 because the thread is not
 144          * running and is thereby implicitly waiting to be on the receiving
 145          * end of a context switch.
 146          */
 147         td->td_critnest = 1;
 148         td->td_lend_user_pri = PRI_MAX;
 149         EVENTHANDLER_INVOKE(thread_ctor, td);
 150 #ifdef AUDIT
 151         audit_thread_alloc(td);
 152 #endif
 153         umtx_thread_alloc(td);
 154         return (0);
 155 }
 156
 157 /*
 158  * Reclaim a thread after use.
 159  */
 160 static void
 161 thread_dtor(void *mem, int size, void *arg)
 162 {
 163         struct thread *td;
 164
 165         td = (struct thread *)mem;
 166
 167 #ifdef INVARIANTS
 168         /* Verify that this thread is in a safe state to free. */
 169         switch (td->td_state) {
 170         case TDS_INHIBITED:
 171         case TDS_RUNNING:
 172         case TDS_CAN_RUN:
 173         case TDS_RUNQ:
 174                 /*
 175                  * We must never unlink a thread that is in one of
 176                  * these states, because it is currently active.
 177                  */
 178                 panic("bad state for thread unlinking");
 179                 /* NOTREACHED */
 180         case TDS_INACTIVE:
 181                 break;
 182         default:
 183                 panic("bad thread state");
 184                 /* NOTREACHED */
 185         }
 186 #endif
 187 #ifdef AUDIT
 188         audit_thread_free(td);
 189 #endif
 190         /* Free all OSD associated to this thread. */
 191         osd_thread_exit(td);
 192
 193         EVENTHANDLER_INVOKE(thread_dtor, td);
 194         tid_free(td->td_tid);
 195 }
 196
 197 /*
 198  * Initialize type-stable parts of a thread (when newly created).
 199  */
 200 static int
 201 thread_init(void *mem, int size, int flags)
 202 {
 203         struct thread *td;
 204
 205         td = (struct thread *)mem;
 206
 207         td->td_sleepqueue = sleepq_alloc();
 208         td->td_turnstile = turnstile_alloc();
 209         td->td_rlqe = NULL;
 210         td->td_vp_reserv = 0;
 211         EVENTHANDLER_INVOKE(thread_init, td);
 212         td->td_sched = (struct td_sched *)&td[1];
 213         umtx_thread_init(td);
 214         td->td_kstack = 0;
 215         return (0);
 216 }
 217
 218 /*
 219  * Tear down type-stable parts of a thread (just before being discarded).
 220  */
 221 static void
 222 thread_fini(void *mem, int size)
 223 {
 224         struct thread *td;
 225
 226         td = (struct thread *)mem;
 227         EVENTHANDLER_INVOKE(thread_fini, td);
 228         rlqentry_free(td->td_rlqe);
 229         turnstile_free(td->td_turnstile);
 230         sleepq_free(td->td_sleepqueue);
 231         umtx_thread_fini(td);
 232         seltdfini(td);
 233 }
 234
 235 /*
 236  * For a newly created process,
 237  * link up all the structures and its initial threads etc.
 238  * called from:
 239  * {arch}/{arch}/machdep.c   ia64_init(), init386() etc.
 240  * proc_dtor() (should go away)
 241  * proc_init()
 242  */
 243 void
 244 proc_linkup0(struct proc *p, struct thread *td)
 245 {
 246         TAILQ_INIT(&p->p_threads);           /* all threads in proc */
 247         proc_linkup(p, td);
 248 }
 249
 250 void
 251 proc_linkup(struct proc *p, struct thread *td)
 252 {
 253
 254         sigqueue_init(&p->p_sigqueue, p);
 255         p->p_ksi = ksiginfo_alloc(1);
 256         if (p->p_ksi != NULL) {
 257                 /* XXX p_ksi may be null if ksiginfo zone is not ready */
 258                 p->p_ksi->ksi_flags = KSI_EXT | KSI_INS;
 259         }
 260         LIST_INIT(&p->p_mqnotifier);
 261         p->p_numthreads = 0;
 262         thread_link(td, p);
 263 }
 264
 265 /*
 266  * Initialize global thread allocation resources.
 267  */
 268 void
 269 threadinit(void)
 270 {
 271
 272         mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF);
 273
 274         /*
 275          * pid_max cannot be greater than PID_MAX.
 276          * leave one number for thread0.
 277          */
 278         tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
 279
 280         thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
 281             thread_ctor, thread_dtor, thread_init, thread_fini,
 282             16 - 1, 0);
 283         tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash);
 284         rw_init(&tidhash_lock, "tidhash");
 285 }
 286
 287 /*
 288  * Place an unused thread on the zombie list.
 289  * Use the slpq as that must be unused by now.
 290  */
 291 void
 292 thread_zombie(struct thread *td)
 293 {
 294         mtx_lock_spin(&zombie_lock);
 295         TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
 296         mtx_unlock_spin(&zombie_lock);
 297 }
 298
 299 /*
 300  * Release a thread that has exited after cpu_throw().
 301  */
 302 void
 303 thread_stash(struct thread *td)
 304 {
 305         atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
 306         thread_zombie(td);
 307 }
 308
 309 /*
 310  * Reap zombie resources.
 311  */
 312 void
 313 thread_reap(void)
 314 {
 315         struct thread *td_first, *td_next;
 316
 317         /*
 318          * Don't even bother to lock if none at this instant,
 319          * we really don't care about the next instant..
 320          */
 321         if (!TAILQ_EMPTY(&zombie_threads)) {
 322                 mtx_lock_spin(&zombie_lock);
 323                 td_first = TAILQ_FIRST(&zombie_threads);
 324                 if (td_first)
 325                         TAILQ_INIT(&zombie_threads);
 326                 mtx_unlock_spin(&zombie_lock);
 327                 while (td_first) {
 328                         td_next = TAILQ_NEXT(td_first, td_slpq);
 329                         if (td_first->td_ucred)
 330                                 crfree(td_first->td_ucred);
 331                         thread_free(td_first);
 332                         td_first = td_next;
 333                 }
 334         }
 335 }
 336
 337 /*
 338  * Allocate a thread.
 339  */
 340 struct thread *
 341 thread_alloc(int pages)
 342 {
 343         struct thread *td;
 344
 345         thread_reap(); /* check if any zombies to get */
 346
 347         td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
 348         KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
 349         if (!vm_thread_new(td, pages)) {
 350                 uma_zfree(thread_zone, td);
 351                 return (NULL);
 352         }
 353         cpu_thread_alloc(td);
 354         return (td);
 355 }
 356
 357 int
 358 thread_alloc_stack(struct thread *td, int pages)
 359 {
 360
 361         KASSERT(td->td_kstack == 0,
 362             ("thread_alloc_stack called on a thread with kstack"));
 363         if (!vm_thread_new(td, pages))
 364                 return (0);
 365         cpu_thread_alloc(td);
 366         return (1);
 367 }
 368
 369 /*
 370  * Deallocate a thread.
 371  */
 372 void
 373 thread_free(struct thread *td)
 374 {
 375
 376         lock_profile_thread_exit(td);
 377         if (td->td_cpuset)
 378                 cpuset_rel(td->td_cpuset);
 379         td->td_cpuset = NULL;
 380         cpu_thread_free(td);
 381         if (td->td_kstack != 0)
 382                 vm_thread_dispose(td);
 383         uma_zfree(thread_zone, td);
 384 }
 385
 386 /*
 387  * Discard the current thread and exit from its context.
 388  * Always called with scheduler locked.
 389  *
 390  * Because we can't free a thread while we're operating under its context,
 391  * push the current thread into our CPU's deadthread holder. This means
 392  * we needn't worry about someone else grabbing our context before we
 393  * do a cpu_throw().
 394  */
 395 void
 396 thread_exit(void)
 397 {
 398         uint64_t runtime, new_switchtime;
 399         struct thread *td;
 400         struct thread *td2;
 401         struct proc *p;
 402         int wakeup_swapper;
 403
 404         td = curthread;
 405         p = td->td_proc;
 406
 407         PROC_SLOCK_ASSERT(p, MA_OWNED);
 408         mtx_assert(&Giant, MA_NOTOWNED);
 409
 410         PROC_LOCK_ASSERT(p, MA_OWNED);
 411         KASSERT(p != NULL, ("thread exiting without a process"));
 412         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 413             (long)p->p_pid, td->td_name);
 414         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 415
 416 #ifdef AUDIT
 417         AUDIT_SYSCALL_EXIT(0, td);
 418 #endif
 419         umtx_thread_exit(td);
 420         /*
 421          * drop FPU & debug register state storage, or any other
 422          * architecture specific resources that
 423          * would not be on a new untouched process.
 424          */
 425         cpu_thread_exit(td);    /* XXXSMP */
 426
 427         /*
 428          * The last thread is left attached to the process
 429          * So that the whole bundle gets recycled. Skip
 430          * all this stuff if we never had threads.
 431          * EXIT clears all sign of other threads when
 432          * it goes to single threading, so the last thread always
 433          * takes the short path.
 434          */
 435         if (p->p_flag & P_HADTHREADS) {
 436                 if (p->p_numthreads > 1) {
 437                         thread_unlink(td);
 438                         td2 = FIRST_THREAD_IN_PROC(p);
 439                         sched_exit_thread(td2, td);
 440
 441                         /*
 442                          * The test below is NOT true if we are the
 443                          * sole exiting thread. P_STOPPED_SINGLE is unset
 444                          * in exit1() after it is the only survivor.
 445                          */
 446                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 447                                 if (p->p_numthreads == p->p_suspcount) {
 448                                         thread_lock(p->p_singlethread);
 449                                         wakeup_swapper = thread_unsuspend_one(
 450                                                 p->p_singlethread);
 451                                         thread_unlock(p->p_singlethread);
 452                                         if (wakeup_swapper)
 453                                                 kick_proc0();
 454                                 }
 455                         }
 456
 457                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 458                         PCPU_SET(deadthread, td);
 459                 } else {
 460                         /*
 461                          * The last thread is exiting.. but not through exit()
 462                          */
 463                         panic ("thread_exit: Last thread exiting on its own");
 464                 }
 465         }
 466 #ifdef  HWPMC_HOOKS
 467         /*
 468          * If this thread is part of a process that is being tracked by hwpmc(4),
 469          * inform the module of the thread's impending exit.
 470          */
 471         if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 472                 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
 473 #endif
 474         PROC_UNLOCK(p);
 475
 476         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 477         new_switchtime = cpu_ticks();
 478         runtime = new_switchtime - PCPU_GET(switchtime);
 479         td->td_runtime += runtime;
 480         td->td_incruntime += runtime;
 481         PCPU_SET(switchtime, new_switchtime);
 482         PCPU_SET(switchticks, ticks);
 483         PCPU_INC(cnt.v_swtch);
 484
 485         /* Save our resource usage in our process. */
 486         td->td_ru.ru_nvcsw++;
 487         ruxagg(p, td);
 488         rucollect(&p->p_ru, &td->td_ru);
 489
 490         thread_lock(td);
 491         PROC_SUNLOCK(p);
 492         td->td_state = TDS_INACTIVE;
 493 #ifdef WITNESS
 494         witness_thread_exit(td);
 495 #endif
 496         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 497         sched_throw(td);
 498         panic("I'm a teapot!");
 499         /* NOTREACHED */
 500 }
 501
 502 /*
 503  * Do any thread specific cleanups that may be needed in wait()
 504  * called with Giant, proc and schedlock not held.
 505  */
 506 void
 507 thread_wait(struct proc *p)
 508 {
 509         struct thread *td;
 510
 511         mtx_assert(&Giant, MA_NOTOWNED);
 512         KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
 513         td = FIRST_THREAD_IN_PROC(p);
 514         /* Lock the last thread so we spin until it exits cpu_throw(). */
 515         thread_lock(td);
 516         thread_unlock(td);
 517         /* Wait for any remaining threads to exit cpu_throw(). */
 518         while (p->p_exitthreads)
 519                 sched_relinquish(curthread);
 520         lock_profile_thread_exit(td);
 521         cpuset_rel(td->td_cpuset);
 522         td->td_cpuset = NULL;
 523         cpu_thread_clean(td);
 524         crfree(td->td_ucred);
 525         thread_reap();  /* check for zombie threads etc. */
 526 }
 527
 528 /*
 529  * Link a thread to a process.
 530  * set up anything that needs to be initialized for it to
 531  * be used by the process.
 532  */
 533 void
 534 thread_link(struct thread *td, struct proc *p)
 535 {
 536
 537         /*
 538          * XXX This can't be enabled because it's called for proc0 before
 539          * its lock has been created.
 540          * PROC_LOCK_ASSERT(p, MA_OWNED);
 541          */
 542         td->td_state    = TDS_INACTIVE;
 543         td->td_proc     = p;
 544         td->td_flags    = TDF_INMEM;
 545
 546         LIST_INIT(&td->td_contested);
 547         LIST_INIT(&td->td_lprof[0]);
 548         LIST_INIT(&td->td_lprof[1]);
 549         sigqueue_init(&td->td_sigqueue, p);
 550         callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
 551         TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
 552         p->p_numthreads++;
 553 }
 554
 555 /*
 556  * Convert a process with one thread to an unthreaded process.
 557  */
 558 void
 559 thread_unthread(struct thread *td)
 560 {
 561         struct proc *p = td->td_proc;
 562
 563         KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
 564         p->p_flag &= ~P_HADTHREADS;
 565 }
 566
 567 /*
 568  * Called from:
 569  *  thread_exit()
 570  */
 571 void
 572 thread_unlink(struct thread *td)
 573 {
 574         struct proc *p = td->td_proc;
 575
 576         PROC_LOCK_ASSERT(p, MA_OWNED);
 577         TAILQ_REMOVE(&p->p_threads, td, td_plist);
 578         p->p_numthreads--;
 579         /* could clear a few other things here */
 580         /* Must  NOT clear links to proc! */
 581 }
 582
 583 static int
 584 calc_remaining(struct proc *p, int mode)
 585 {
 586         int remaining;
 587
 588         PROC_LOCK_ASSERT(p, MA_OWNED);
 589         PROC_SLOCK_ASSERT(p, MA_OWNED);
 590         if (mode == SINGLE_EXIT)
 591                 remaining = p->p_numthreads;
 592         else if (mode == SINGLE_BOUNDARY)
 593                 remaining = p->p_numthreads - p->p_boundary_count;
 594         else if (mode == SINGLE_NO_EXIT)
 595                 remaining = p->p_numthreads - p->p_suspcount;
 596         else
 597                 panic("calc_remaining: wrong mode %d", mode);
 598         return (remaining);
 599 }
 600
 601 /*
 602  * Enforce single-threading.
 603  *
 604  * Returns 1 if the caller must abort (another thread is waiting to
 605  * exit the process or similar). Process is locked!
 606  * Returns 0 when you are successfully the only thread running.
 607  * A process has successfully single threaded in the suspend mode when
 608  * There are no threads in user mode. Threads in the kernel must be
 609  * allowed to continue until they get to the user boundary. They may even
 610  * copy out their return values and data before suspending. They may however be
 611  * accelerated in reaching the user boundary as we will wake up
 612  * any sleeping threads that are interruptable. (PCATCH).
 613  */
 614 int
 615 thread_single(int mode)
 616 {
 617         struct thread *td;
 618         struct thread *td2;
 619         struct proc *p;
 620         int remaining, wakeup_swapper;
 621
 622         td = curthread;
 623         p = td->td_proc;
 624         mtx_assert(&Giant, MA_NOTOWNED);
 625         PROC_LOCK_ASSERT(p, MA_OWNED);
 626         KASSERT((td != NULL), ("curthread is NULL"));
 627
 628         if ((p->p_flag & P_HADTHREADS) == 0)
 629                 return (0);
 630
 631         /* Is someone already single threading? */
 632         if (p->p_singlethread != NULL && p->p_singlethread != td)
 633                 return (1);
 634
 635         if (mode == SINGLE_EXIT) {
 636                 p->p_flag |= P_SINGLE_EXIT;
 637                 p->p_flag &= ~P_SINGLE_BOUNDARY;
 638         } else {
 639                 p->p_flag &= ~P_SINGLE_EXIT;
 640                 if (mode == SINGLE_BOUNDARY)
 641                         p->p_flag |= P_SINGLE_BOUNDARY;
 642                 else
 643                         p->p_flag &= ~P_SINGLE_BOUNDARY;
 644         }
 645         p->p_flag |= P_STOPPED_SINGLE;
 646         PROC_SLOCK(p);
 647         p->p_singlethread = td;
 648         remaining = calc_remaining(p, mode);
 649         while (remaining != 1) {
 650                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
 651                         goto stopme;
 652                 wakeup_swapper = 0;
 653                 FOREACH_THREAD_IN_PROC(p, td2) {
 654                         if (td2 == td)
 655                                 continue;
 656                         thread_lock(td2);
 657                         td2->td_flags |= TDF_ASTPENDING | TDF_NEEDSUSPCHK;
 658                         if (TD_IS_INHIBITED(td2)) {
 659                                 switch (mode) {
 660                                 case SINGLE_EXIT:
 661                                         if (TD_IS_SUSPENDED(td2))
 662                                                 wakeup_swapper |=
 663                                                     thread_unsuspend_one(td2);
 664                                         if (TD_ON_SLEEPQ(td2) &&
 665                                             (td2->td_flags & TDF_SINTR))
 666                                                 wakeup_swapper |=
 667                                                     sleepq_abort(td2, EINTR);
 668                                         break;
 669                                 case SINGLE_BOUNDARY:
 670                                         if (TD_IS_SUSPENDED(td2) &&
 671                                             !(td2->td_flags & TDF_BOUNDARY))
 672                                                 wakeup_swapper |=
 673                                                     thread_unsuspend_one(td2);
 674                                         if (TD_ON_SLEEPQ(td2) &&
 675                                             (td2->td_flags & TDF_SINTR))
 676                                                 wakeup_swapper |=
 677                                                     sleepq_abort(td2, ERESTART);
 678                                         break;
 679                                 case SINGLE_NO_EXIT:
 680                                         if (TD_IS_SUSPENDED(td2) &&
 681                                             !(td2->td_flags & TDF_BOUNDARY))
 682                                                 wakeup_swapper |=
 683                                                     thread_unsuspend_one(td2);
 684                                         if (TD_ON_SLEEPQ(td2) &&
 685                                             (td2->td_flags & TDF_SINTR))
 686                                                 wakeup_swapper |=
 687                                                     sleepq_abort(td2, ERESTART);
 688                                         break;
 689                                 default:
 690                                         break;
 691                                 }
 692                         }
 693 #ifdef SMP
 694                         else if (TD_IS_RUNNING(td2) && td != td2) {
 695                                 forward_signal(td2);
 696                         }
 697 #endif
 698                         thread_unlock(td2);
 699                 }
 700                 if (wakeup_swapper)
 701                         kick_proc0();
 702                 remaining = calc_remaining(p, mode);
 703
 704                 /*
 705                  * Maybe we suspended some threads.. was it enough?
 706                  */
 707                 if (remaining == 1)
 708                         break;
 709
 710 stopme:
 711                 /*
 712                  * Wake us up when everyone else has suspended.
 713                  * In the mean time we suspend as well.
 714                  */
 715                 thread_suspend_switch(td);
 716                 remaining = calc_remaining(p, mode);
 717         }
 718         if (mode == SINGLE_EXIT) {
 719                 /*
 720                  * We have gotten rid of all the other threads and we
 721                  * are about to either exit or exec. In either case,
 722                  * we try our utmost  to revert to being a non-threaded
 723                  * process.
 724                  */
 725                 p->p_singlethread = NULL;
 726                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
 727                 thread_unthread(td);
 728         }
 729         PROC_SUNLOCK(p);
 730         return (0);
 731 }
 732
 733 /*
 734  * Called in from locations that can safely check to see
 735  * whether we have to suspend or at least throttle for a
 736  * single-thread event (e.g. fork).
 737  *
 738  * Such locations include userret().
 739  * If the "return_instead" argument is non zero, the thread must be able to
 740  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
 741  *
 742  * The 'return_instead' argument tells the function if it may do a
 743  * thread_exit() or suspend, or whether the caller must abort and back
 744  * out instead.
 745  *
 746  * If the thread that set the single_threading request has set the
 747  * P_SINGLE_EXIT bit in the process flags then this call will never return
 748  * if 'return_instead' is false, but will exit.
 749  *
 750  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
 751  *---------------+--------------------+---------------------
 752  *       0       | returns 0          |   returns 0 or 1
 753  *               | when ST ends       |   immediatly
 754  *---------------+--------------------+---------------------
 755  *       1       | thread exits       |   returns 1
 756  *               |                    |  immediatly
 757  * 0 = thread_exit() or suspension ok,
 758  * other = return error instead of stopping the thread.
 759  *
 760  * While a full suspension is under effect, even a single threading
 761  * thread would be suspended if it made this call (but it shouldn't).
 762  * This call should only be made from places where
 763  * thread_exit() would be safe as that may be the outcome unless
 764  * return_instead is set.
 765  */
 766 int
 767 thread_suspend_check(int return_instead)
 768 {
 769         struct thread *td;
 770         struct proc *p;
 771         int wakeup_swapper;
 772
 773         td = curthread;
 774         p = td->td_proc;
 775         mtx_assert(&Giant, MA_NOTOWNED);
 776         PROC_LOCK_ASSERT(p, MA_OWNED);
 777         while (P_SHOULDSTOP(p) ||
 778               ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
 779                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 780                         KASSERT(p->p_singlethread != NULL,
 781                             ("singlethread not set"));
 782                         /*
 783                          * The only suspension in action is a
 784                          * single-threading. Single threader need not stop.
 785                          * XXX Should be safe to access unlocked
 786                          * as it can only be set to be true by us.
 787                          */
 788                         if (p->p_singlethread == td)
 789                                 return (0);     /* Exempt from stopping. */
 790                 }
 791                 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
 792                         return (EINTR);
 793
 794                 /* Should we goto user boundary if we didn't come from there? */
 795                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
 796                     (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
 797                         return (ERESTART);
 798
 799                 /*
 800                  * If the process is waiting for us to exit,
 801                  * this thread should just suicide.
 802                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
 803                  */
 804                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) {
 805                         PROC_UNLOCK(p);
 806                         tidhash_remove(td);
 807                         PROC_LOCK(p);
 808                         tdsigcleanup(td);
 809                         PROC_SLOCK(p);
 810                         thread_stopped(p);
 811                         thread_exit();
 812                 }
 813
 814                 PROC_SLOCK(p);
 815                 thread_stopped(p);
 816                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 817                         if (p->p_numthreads == p->p_suspcount + 1) {
 818                                 thread_lock(p->p_singlethread);
 819                                 wakeup_swapper =
 820                                     thread_unsuspend_one(p->p_singlethread);
 821                                 thread_unlock(p->p_singlethread);
 822                                 if (wakeup_swapper)
 823                                         kick_proc0();
 824                         }
 825                 }
 826                 PROC_UNLOCK(p);
 827                 thread_lock(td);
 828                 /*
 829                  * When a thread suspends, it just
 830                  * gets taken off all queues.
 831                  */
 832                 thread_suspend_one(td);
 833                 if (return_instead == 0) {
 834                         p->p_boundary_count++;
 835                         td->td_flags |= TDF_BOUNDARY;
 836                 }
 837                 PROC_SUNLOCK(p);
 838                 mi_switch(SW_INVOL | SWT_SUSPEND, NULL);
 839                 if (return_instead == 0)
 840                         td->td_flags &= ~TDF_BOUNDARY;
 841                 thread_unlock(td);
 842                 PROC_LOCK(p);
 843                 if (return_instead == 0) {
 844                         PROC_SLOCK(p);
 845                         p->p_boundary_count--;
 846                         PROC_SUNLOCK(p);
 847                 }
 848         }
 849         return (0);
 850 }
 851
 852 void
 853 thread_suspend_switch(struct thread *td)
 854 {
 855         struct proc *p;
 856
 857         p = td->td_proc;
 858         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 859         PROC_LOCK_ASSERT(p, MA_OWNED);
 860         PROC_SLOCK_ASSERT(p, MA_OWNED);
 861         /*
 862          * We implement thread_suspend_one in stages here to avoid
 863          * dropping the proc lock while the thread lock is owned.
 864          */
 865         thread_stopped(p);
 866         p->p_suspcount++;
 867         PROC_UNLOCK(p);
 868         thread_lock(td);
 869         td->td_flags &= ~TDF_NEEDSUSPCHK;
 870         TD_SET_SUSPENDED(td);
 871         sched_sleep(td, 0);
 872         PROC_SUNLOCK(p);
 873         DROP_GIANT();
 874         mi_switch(SW_VOL | SWT_SUSPEND, 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         td->td_flags &= ~TDF_NEEDSUSPCHK;
 891         TD_SET_SUSPENDED(td);
 892         sched_sleep(td, 0);
 893 }
 894
 895 int
 896 thread_unsuspend_one(struct thread *td)
 897 {
 898         struct proc *p = td->td_proc;
 899
 900         PROC_SLOCK_ASSERT(p, MA_OWNED);
 901         THREAD_LOCK_ASSERT(td, MA_OWNED);
 902         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
 903         TD_CLR_SUSPENDED(td);
 904         p->p_suspcount--;
 905         return (setrunnable(td));
 906 }
 907
 908 /*
 909  * Allow all threads blocked by single threading to continue running.
 910  */
 911 void
 912 thread_unsuspend(struct proc *p)
 913 {
 914         struct thread *td;
 915         int wakeup_swapper;
 916
 917         PROC_LOCK_ASSERT(p, MA_OWNED);
 918         PROC_SLOCK_ASSERT(p, MA_OWNED);
 919         wakeup_swapper = 0;
 920         if (!P_SHOULDSTOP(p)) {
 921                 FOREACH_THREAD_IN_PROC(p, td) {
 922                         thread_lock(td);
 923                         if (TD_IS_SUSPENDED(td)) {
 924                                 wakeup_swapper |= thread_unsuspend_one(td);
 925                         }
 926                         thread_unlock(td);
 927                 }
 928         } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
 929             (p->p_numthreads == p->p_suspcount)) {
 930                 /*
 931                  * Stopping everything also did the job for the single
 932                  * threading request. Now we've downgraded to single-threaded,
 933                  * let it continue.
 934                  */
 935                 thread_lock(p->p_singlethread);
 936                 wakeup_swapper = thread_unsuspend_one(p->p_singlethread);
 937                 thread_unlock(p->p_singlethread);
 938         }
 939         if (wakeup_swapper)
 940                 kick_proc0();
 941 }
 942
 943 /*
 944  * End the single threading mode..
 945  */
 946 void
 947 thread_single_end(void)
 948 {
 949         struct thread *td;
 950         struct proc *p;
 951         int wakeup_swapper;
 952
 953         td = curthread;
 954         p = td->td_proc;
 955         PROC_LOCK_ASSERT(p, MA_OWNED);
 956         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
 957         PROC_SLOCK(p);
 958         p->p_singlethread = NULL;
 959         wakeup_swapper = 0;
 960         /*
 961          * If there are other threads they may now run,
 962          * unless of course there is a blanket 'stop order'
 963          * on the process. The single threader must be allowed
 964          * to continue however as this is a bad place to stop.
 965          */
 966         if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
 967                 FOREACH_THREAD_IN_PROC(p, td) {
 968                         thread_lock(td);
 969                         if (TD_IS_SUSPENDED(td)) {
 970                                 wakeup_swapper |= thread_unsuspend_one(td);
 971                         }
 972                         thread_unlock(td);
 973                 }
 974         }
 975         PROC_SUNLOCK(p);
 976         if (wakeup_swapper)
 977                 kick_proc0();
 978 }
 979
 980 struct thread *
 981 thread_find(struct proc *p, lwpid_t tid)
 982 {
 983         struct thread *td;
 984
 985         PROC_LOCK_ASSERT(p, MA_OWNED);
 986         FOREACH_THREAD_IN_PROC(p, td) {
 987                 if (td->td_tid == tid)
 988                         break;
 989         }
 990         return (td);
 991 }
 992
 993 /* Locate a thread by number; return with proc lock held. */
 994 struct thread *
 995 tdfind(lwpid_t tid, pid_t pid)
 996 {
 997 #define RUN_THRESH      16
 998         struct thread *td;
 999         int run = 0;
1000
1001         rw_rlock(&tidhash_lock);
1002         LIST_FOREACH(td, TIDHASH(tid), td_hash) {
1003                 if (td->td_tid == tid) {
1004                         if (pid != -1 && td->td_proc->p_pid != pid) {
1005                                 td = NULL;
1006                                 break;
1007                         }
1008                         PROC_LOCK(td->td_proc);
1009                         if (td->td_proc->p_state == PRS_NEW) {
1010                                 PROC_UNLOCK(td->td_proc);
1011                                 td = NULL;
1012                                 break;
1013                         }
1014                         if (run > RUN_THRESH) {
1015                                 if (rw_try_upgrade(&tidhash_lock)) {
1016                                         LIST_REMOVE(td, td_hash);
1017                                         LIST_INSERT_HEAD(TIDHASH(td->td_tid),
1018                                                 td, td_hash);
1019                                         rw_wunlock(&tidhash_lock);
1020                                         return (td);
1021                                 }
1022                         }
1023                         break;
1024                 }
1025                 run++;
1026         }
1027         rw_runlock(&tidhash_lock);
1028         return (td);
1029 }
1030
1031 void
1032 tidhash_add(struct thread *td)
1033 {
1034         rw_wlock(&tidhash_lock);
1035         LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash);
1036         rw_wunlock(&tidhash_lock);
1037 }
1038
1039 void
1040 tidhash_remove(struct thread *td)
1041 {
1042         rw_wlock(&tidhash_lock);
1043         LIST_REMOVE(td, td_hash);
1044         rw_wunlock(&tidhash_lock);
1045 }