sys/kern/kern_thread.c

   1 /*-
   2  * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>.
   3  *  All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice(s), this list of conditions and the following disclaimer as
  10  *    the first lines of this file unmodified other than the possible
  11  *    addition of one or more copyright notices.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice(s), this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  17  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  18  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  19  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  20  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  21  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  22  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  23  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  26  * DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/systm.h>
  34 #include <sys/kernel.h>
  35 #include <sys/lock.h>
  36 #include <sys/mutex.h>
  37 #include <sys/proc.h>
  38 #include <sys/resourcevar.h>
  39 #include <sys/smp.h>
  40 #include <sys/sysctl.h>
  41 #include <sys/sched.h>
  42 #include <sys/sleepqueue.h>
  43 #include <sys/selinfo.h>
  44 #include <sys/turnstile.h>
  45 #include <sys/ktr.h>
  46 #include <sys/umtx.h>
  47 #include <sys/cpuset.h>
  48
  49 #include <security/audit/audit.h>
  50
  51 #include <vm/vm.h>
  52 #include <vm/vm_extern.h>
  53 #include <vm/uma.h>
  54 #include <sys/eventhandler.h>
  55
  56 /*
  57  * thread related storage.
  58  */
  59 static uma_zone_t thread_zone;
  60
  61 SYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation");
  62
  63 int max_threads_per_proc = 1500;
  64 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_per_proc, CTLFLAG_RW,
  65         &max_threads_per_proc, 0, "Limit on threads per proc");
  66
  67 int max_threads_hits;
  68 SYSCTL_INT(_kern_threads, OID_AUTO, max_threads_hits, CTLFLAG_RD,
  69         &max_threads_hits, 0, "");
  70
  71 #ifdef KSE
  72 int virtual_cpu;
  73
  74 #endif
  75 TAILQ_HEAD(, thread) zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads);
  76 static struct mtx zombie_lock;
  77 MTX_SYSINIT(zombie_lock, &zombie_lock, "zombie lock", MTX_SPIN);
  78
  79 static void thread_zombie(struct thread *);
  80
  81 #ifdef KSE
  82 static int
  83 sysctl_kse_virtual_cpu(SYSCTL_HANDLER_ARGS)
  84 {
  85         int error, new_val;
  86         int def_val;
  87
  88         def_val = mp_ncpus;
  89         if (virtual_cpu == 0)
  90                 new_val = def_val;
  91         else
  92                 new_val = virtual_cpu;
  93         error = sysctl_handle_int(oidp, &new_val, 0, req);
  94         if (error != 0 || req->newptr == NULL)
  95                 return (error);
  96         if (new_val < 0)
  97                 return (EINVAL);
  98         virtual_cpu = new_val;
  99         return (0);
 100 }
 101
 102 /* DEBUG ONLY */
 103 SYSCTL_PROC(_kern_threads, OID_AUTO, virtual_cpu, CTLTYPE_INT|CTLFLAG_RW,
 104         0, sizeof(virtual_cpu), sysctl_kse_virtual_cpu, "I",
 105         "debug virtual cpus");
 106 #endif
 107
 108 struct mtx tid_lock;
 109 static struct unrhdr *tid_unrhdr;
 110
 111 /*
 112  * Prepare a thread for use.
 113  */
 114 static int
 115 thread_ctor(void *mem, int size, void *arg, int flags)
 116 {
 117         struct thread   *td;
 118
 119         td = (struct thread *)mem;
 120         td->td_state = TDS_INACTIVE;
 121         td->td_oncpu = NOCPU;
 122
 123         td->td_tid = alloc_unr(tid_unrhdr);
 124         td->td_syscalls = 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         return (0);
 196 }
 197
 198 /*
 199  * Tear down type-stable parts of a thread (just before being discarded).
 200  */
 201 static void
 202 thread_fini(void *mem, int size)
 203 {
 204         struct thread *td;
 205
 206         td = (struct thread *)mem;
 207         EVENTHANDLER_INVOKE(thread_fini, td);
 208         turnstile_free(td->td_turnstile);
 209         sleepq_free(td->td_sleepqueue);
 210         umtx_thread_fini(td);
 211         seltdfini(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         /* leave one number for thread0 */
 256         tid_unrhdr = new_unrhdr(PID_MAX + 2, INT_MAX, &tid_lock);
 257
 258         thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(),
 259             thread_ctor, thread_dtor, thread_init, thread_fini,
 260             16 - 1, 0);
 261 #ifdef KSE
 262         kseinit();      /* set up kse specific stuff  e.g. upcall zone*/
 263 #endif
 264 }
 265
 266 /*
 267  * Place an unused thread on the zombie list.
 268  * Use the slpq as that must be unused by now.
 269  */
 270 void
 271 thread_zombie(struct thread *td)
 272 {
 273         mtx_lock_spin(&zombie_lock);
 274         TAILQ_INSERT_HEAD(&zombie_threads, td, td_slpq);
 275         mtx_unlock_spin(&zombie_lock);
 276 }
 277
 278 /*
 279  * Release a thread that has exited after cpu_throw().
 280  */
 281 void
 282 thread_stash(struct thread *td)
 283 {
 284         atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1);
 285         thread_zombie(td);
 286 }
 287
 288 /*
 289  * Reap zombie kse resource.
 290  */
 291 void
 292 thread_reap(void)
 293 {
 294         struct thread *td_first, *td_next;
 295
 296         /*
 297          * Don't even bother to lock if none at this instant,
 298          * we really don't care about the next instant..
 299          */
 300         if (!TAILQ_EMPTY(&zombie_threads)) {
 301                 mtx_lock_spin(&zombie_lock);
 302                 td_first = TAILQ_FIRST(&zombie_threads);
 303                 if (td_first)
 304                         TAILQ_INIT(&zombie_threads);
 305                 mtx_unlock_spin(&zombie_lock);
 306                 while (td_first) {
 307                         td_next = TAILQ_NEXT(td_first, td_slpq);
 308                         if (td_first->td_ucred)
 309                                 crfree(td_first->td_ucred);
 310                         thread_free(td_first);
 311                         td_first = td_next;
 312                 }
 313         }
 314 #ifdef KSE
 315         upcall_reap();
 316 #endif
 317 }
 318
 319 /*
 320  * Allocate a thread.
 321  */
 322 struct thread *
 323 thread_alloc(void)
 324 {
 325         struct thread *td;
 326
 327         thread_reap(); /* check if any zombies to get */
 328
 329         td = (struct thread *)uma_zalloc(thread_zone, M_WAITOK);
 330         KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack"));
 331         if (!vm_thread_new(td, 0)) {
 332                 uma_zfree(thread_zone, td);
 333                 return (NULL);
 334         }
 335         cpu_thread_alloc(td);
 336         return (td);
 337 }
 338
 339
 340 /*
 341  * Deallocate a thread.
 342  */
 343 void
 344 thread_free(struct thread *td)
 345 {
 346 #ifdef KSE
 347         if (td->td_cpuset != NULL)
 348                 cpuset_rel(td->td_cpuset);
 349 #else
 350         cpuset_rel(td->td_cpuset);
 351 #endif
 352         td->td_cpuset = NULL;
 353         cpu_thread_free(td);
 354         if (td->td_altkstack != 0)
 355                 vm_thread_dispose_altkstack(td);
 356         if (td->td_kstack != 0)
 357                 vm_thread_dispose(td);
 358         uma_zfree(thread_zone, td);
 359 }
 360
 361 /*
 362  * Discard the current thread and exit from its context.
 363  * Always called with scheduler locked.
 364  *
 365  * Because we can't free a thread while we're operating under its context,
 366  * push the current thread into our CPU's deadthread holder. This means
 367  * we needn't worry about someone else grabbing our context before we
 368  * do a cpu_throw().  This may not be needed now as we are under schedlock.
 369  * Maybe we can just do a thread_stash() as thr_exit1 does.
 370  */
 371 /*  XXX
 372  * libthr expects its thread exit to return for the last
 373  * thread, meaning that the program is back to non-threaded
 374  * mode I guess. Because we do this (cpu_throw) unconditionally
 375  * here, they have their own version of it. (thr_exit1())
 376  * that doesn't do it all if this was the last thread.
 377  * It is also called from thread_suspend_check().
 378  * Of course in the end, they end up coming here through exit1
 379  * anyhow..  After fixing 'thr' to play by the rules we should be able
 380  * to merge these two functions together.
 381  *
 382  * called from:
 383  * exit1()
 384  * kse_exit()
 385  * thr_exit()
 386  * ifdef KSE
 387  * thread_user_enter()
 388  * thread_userret()
 389  * endif
 390  * thread_suspend_check()
 391  */
 392 void
 393 thread_exit(void)
 394 {
 395         uint64_t new_switchtime;
 396         struct thread *td;
 397         struct thread *td2;
 398         struct proc *p;
 399
 400         td = curthread;
 401         p = td->td_proc;
 402
 403         PROC_SLOCK_ASSERT(p, MA_OWNED);
 404         mtx_assert(&Giant, MA_NOTOWNED);
 405
 406         PROC_LOCK_ASSERT(p, MA_OWNED);
 407         KASSERT(p != NULL, ("thread exiting without a process"));
 408         CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td,
 409             (long)p->p_pid, td->td_name);
 410         KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending"));
 411
 412 #ifdef AUDIT
 413         AUDIT_SYSCALL_EXIT(0, td);
 414 #endif
 415
 416 #ifdef KSE
 417         if (td->td_standin != NULL) {
 418                 /*
 419                  * Note that we don't need to free the cred here as it
 420                  * is done in thread_reap().
 421                  */
 422                 thread_zombie(td->td_standin);
 423                 td->td_standin = NULL;
 424         }
 425 #endif
 426
 427         umtx_thread_exit(td);
 428
 429         /*
 430          * drop FPU & debug register state storage, or any other
 431          * architecture specific resources that
 432          * would not be on a new untouched process.
 433          */
 434         cpu_thread_exit(td);    /* XXXSMP */
 435
 436         /* Do the same timestamp bookkeeping that mi_switch() would do. */
 437         new_switchtime = cpu_ticks();
 438         p->p_rux.rux_runtime += (new_switchtime - PCPU_GET(switchtime));
 439         PCPU_SET(switchtime, new_switchtime);
 440         PCPU_SET(switchticks, ticks);
 441         PCPU_INC(cnt.v_swtch);
 442         /* Save our resource usage in our process. */
 443         td->td_ru.ru_nvcsw++;
 444         rucollect(&p->p_ru, &td->td_ru);
 445         /*
 446          * The last thread is left attached to the process
 447          * So that the whole bundle gets recycled. Skip
 448          * all this stuff if we never had threads.
 449          * EXIT clears all sign of other threads when
 450          * it goes to single threading, so the last thread always
 451          * takes the short path.
 452          */
 453         if (p->p_flag & P_HADTHREADS) {
 454                 if (p->p_numthreads > 1) {
 455                         thread_lock(td);
 456 #ifdef KSE
 457                         kse_unlink(td);
 458 #else
 459                         thread_unlink(td);
 460 #endif
 461                         thread_unlock(td);
 462                         td2 = FIRST_THREAD_IN_PROC(p);
 463                         sched_exit_thread(td2, td);
 464
 465                         /*
 466                          * The test below is NOT true if we are the
 467                          * sole exiting thread. P_STOPPED_SNGL is unset
 468                          * in exit1() after it is the only survivor.
 469                          */
 470                         if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 471                                 if (p->p_numthreads == p->p_suspcount) {
 472                                         thread_lock(p->p_singlethread);
 473                                         thread_unsuspend_one(p->p_singlethread);
 474                                         thread_unlock(p->p_singlethread);
 475                                 }
 476                         }
 477
 478                         atomic_add_int(&td->td_proc->p_exitthreads, 1);
 479                         PCPU_SET(deadthread, td);
 480                 } else {
 481                         /*
 482                          * The last thread is exiting.. but not through exit()
 483                          * what should we do?
 484                          * Theoretically this can't happen
 485                          * exit1() - clears threading flags before coming here
 486                          * kse_exit() - treats last thread specially
 487                          * thr_exit() - treats last thread specially
 488                          * ifdef KSE
 489                          * thread_user_enter() - only if more exist
 490                          * thread_userret() - only if more exist
 491                          * endif
 492                          * thread_suspend_check() - only if more exist
 493                          */
 494                         panic ("thread_exit: Last thread exiting on its own");
 495                 }
 496         }
 497         PROC_UNLOCK(p);
 498         thread_lock(td);
 499         /* Save our tick information with both the thread and proc locked */
 500         ruxagg(&p->p_rux, td);
 501         PROC_SUNLOCK(p);
 502         td->td_state = TDS_INACTIVE;
 503         CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td);
 504         sched_throw(td);
 505         panic("I'm a teapot!");
 506         /* NOTREACHED */
 507 }
 508
 509 /*
 510  * Do any thread specific cleanups that may be needed in wait()
 511  * called with Giant, proc and schedlock not held.
 512  */
 513 void
 514 thread_wait(struct proc *p)
 515 {
 516         struct thread *td;
 517
 518         mtx_assert(&Giant, MA_NOTOWNED);
 519         KASSERT((p->p_numthreads == 1), ("Multiple threads in wait1()"));
 520         td = FIRST_THREAD_IN_PROC(p);
 521 #ifdef KSE
 522         if (td->td_standin != NULL) {
 523                 if (td->td_standin->td_ucred != NULL) {
 524                         crfree(td->td_standin->td_ucred);
 525                         td->td_standin->td_ucred = NULL;
 526                 }
 527                 thread_free(td->td_standin);
 528                 td->td_standin = NULL;
 529         }
 530 #endif
 531         /* Lock the last thread so we spin until it exits cpu_throw(). */
 532         thread_lock(td);
 533         thread_unlock(td);
 534         /* Wait for any remaining threads to exit cpu_throw(). */
 535         while (p->p_exitthreads)
 536                 sched_relinquish(curthread);
 537         cpuset_rel(td->td_cpuset);
 538         td->td_cpuset = NULL;
 539         cpu_thread_clean(td);
 540         crfree(td->td_ucred);
 541         thread_reap();  /* check for zombie threads etc. */
 542 }
 543
 544 /*
 545  * Link a thread to a process.
 546  * set up anything that needs to be initialized for it to
 547  * be used by the process.
 548  *
 549  * Note that we do not link to the proc's ucred here.
 550  * The thread is linked as if running but no KSE assigned.
 551  * Called from:
 552  *  proc_linkup()
 553  *  thread_schedule_upcall()
 554  *  thr_create()
 555  */
 556 void
 557 thread_link(struct thread *td, struct proc *p)
 558 {
 559
 560         /*
 561          * XXX This can't be enabled because it's called for proc0 before
 562          * it's spinlock has been created.
 563          * PROC_SLOCK_ASSERT(p, MA_OWNED);
 564          */
 565         td->td_state    = TDS_INACTIVE;
 566         td->td_proc     = p;
 567         td->td_flags    = TDF_INMEM;
 568
 569         LIST_INIT(&td->td_contested);
 570         LIST_INIT(&td->td_lprof[0]);
 571         LIST_INIT(&td->td_lprof[1]);
 572         sigqueue_init(&td->td_sigqueue, p);
 573         callout_init(&td->td_slpcallout, CALLOUT_MPSAFE);
 574         TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist);
 575         p->p_numthreads++;
 576 }
 577
 578 /*
 579  * Convert a process with one thread to an unthreaded process.
 580  * Called from:
 581  *  thread_single(exit)  (called from execve and exit)
 582  *  kse_exit()          XXX may need cleaning up wrt KSE stuff
 583  */
 584 void
 585 thread_unthread(struct thread *td)
 586 {
 587         struct proc *p = td->td_proc;
 588
 589         KASSERT((p->p_numthreads == 1), ("Unthreading with >1 threads"));
 590 #ifdef KSE
 591         thread_lock(td);
 592         upcall_remove(td);
 593         thread_unlock(td);
 594         p->p_flag &= ~(P_SA|P_HADTHREADS);
 595         td->td_mailbox = NULL;
 596         td->td_pflags &= ~(TDP_SA | TDP_CAN_UNBIND);
 597         if (td->td_standin != NULL) {
 598                 thread_zombie(td->td_standin);
 599                 td->td_standin = NULL;
 600         }
 601 #else
 602         p->p_flag &= ~P_HADTHREADS;
 603 #endif
 604 }
 605
 606 /*
 607  * Called from:
 608  *  thread_exit()
 609  */
 610 void
 611 thread_unlink(struct thread *td)
 612 {
 613         struct proc *p = td->td_proc;
 614
 615         PROC_SLOCK_ASSERT(p, MA_OWNED);
 616         TAILQ_REMOVE(&p->p_threads, td, td_plist);
 617         p->p_numthreads--;
 618         /* could clear a few other things here */
 619         /* Must  NOT clear links to proc! */
 620 }
 621
 622 /*
 623  * Enforce single-threading.
 624  *
 625  * Returns 1 if the caller must abort (another thread is waiting to
 626  * exit the process or similar). Process is locked!
 627  * Returns 0 when you are successfully the only thread running.
 628  * A process has successfully single threaded in the suspend mode when
 629  * There are no threads in user mode. Threads in the kernel must be
 630  * allowed to continue until they get to the user boundary. They may even
 631  * copy out their return values and data before suspending. They may however be
 632  * accelerated in reaching the user boundary as we will wake up
 633  * any sleeping threads that are interruptable. (PCATCH).
 634  */
 635 int
 636 thread_single(int mode)
 637 {
 638         struct thread *td;
 639         struct thread *td2;
 640         struct proc *p;
 641         int remaining;
 642
 643         td = curthread;
 644         p = td->td_proc;
 645         mtx_assert(&Giant, MA_NOTOWNED);
 646         PROC_LOCK_ASSERT(p, MA_OWNED);
 647         KASSERT((td != NULL), ("curthread is NULL"));
 648
 649         if ((p->p_flag & P_HADTHREADS) == 0)
 650                 return (0);
 651
 652         /* Is someone already single threading? */
 653         if (p->p_singlethread != NULL && p->p_singlethread != td)
 654                 return (1);
 655
 656         if (mode == SINGLE_EXIT) {
 657                 p->p_flag |= P_SINGLE_EXIT;
 658                 p->p_flag &= ~P_SINGLE_BOUNDARY;
 659         } else {
 660                 p->p_flag &= ~P_SINGLE_EXIT;
 661                 if (mode == SINGLE_BOUNDARY)
 662                         p->p_flag |= P_SINGLE_BOUNDARY;
 663                 else
 664                         p->p_flag &= ~P_SINGLE_BOUNDARY;
 665         }
 666         p->p_flag |= P_STOPPED_SINGLE;
 667         PROC_SLOCK(p);
 668         p->p_singlethread = td;
 669         if (mode == SINGLE_EXIT)
 670                 remaining = p->p_numthreads;
 671         else if (mode == SINGLE_BOUNDARY)
 672                 remaining = p->p_numthreads - p->p_boundary_count;
 673         else
 674                 remaining = p->p_numthreads - p->p_suspcount;
 675         while (remaining != 1) {
 676                 if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE)
 677                         goto stopme;
 678                 FOREACH_THREAD_IN_PROC(p, td2) {
 679                         if (td2 == td)
 680                                 continue;
 681                         thread_lock(td2);
 682                         td2->td_flags |= TDF_ASTPENDING;
 683                         if (TD_IS_INHIBITED(td2)) {
 684                                 switch (mode) {
 685                                 case SINGLE_EXIT:
 686                                         if (td->td_flags & TDF_DBSUSPEND)
 687                                                 td->td_flags &= ~TDF_DBSUSPEND;
 688                                         if (TD_IS_SUSPENDED(td2))
 689                                                 thread_unsuspend_one(td2);
 690                                         if (TD_ON_SLEEPQ(td2) &&
 691                                             (td2->td_flags & TDF_SINTR))
 692                                                 sleepq_abort(td2, EINTR);
 693                                         break;
 694                                 case SINGLE_BOUNDARY:
 695                                         break;
 696                                 default:
 697                                         if (TD_IS_SUSPENDED(td2)) {
 698                                                 thread_unlock(td2);
 699                                                 continue;
 700                                         }
 701                                         /*
 702                                          * maybe other inhibited states too?
 703                                          */
 704                                         if ((td2->td_flags & TDF_SINTR) &&
 705                                             (td2->td_inhibitors &
 706                                             (TDI_SLEEPING | TDI_SWAPPED)))
 707                                                 thread_suspend_one(td2);
 708                                         break;
 709                                 }
 710                         }
 711 #ifdef SMP
 712                         else if (TD_IS_RUNNING(td2) && td != td2) {
 713                                 forward_signal(td2);
 714                         }
 715 #endif
 716                         thread_unlock(td2);
 717                 }
 718                 if (mode == SINGLE_EXIT)
 719                         remaining = p->p_numthreads;
 720                 else if (mode == SINGLE_BOUNDARY)
 721                         remaining = p->p_numthreads - p->p_boundary_count;
 722                 else
 723                         remaining = p->p_numthreads - p->p_suspcount;
 724
 725                 /*
 726                  * Maybe we suspended some threads.. was it enough?
 727                  */
 728                 if (remaining == 1)
 729                         break;
 730
 731 stopme:
 732                 /*
 733                  * Wake us up when everyone else has suspended.
 734                  * In the mean time we suspend as well.
 735                  */
 736                 thread_suspend_switch(td);
 737                 if (mode == SINGLE_EXIT)
 738                         remaining = p->p_numthreads;
 739                 else if (mode == SINGLE_BOUNDARY)
 740                         remaining = p->p_numthreads - p->p_boundary_count;
 741                 else
 742                         remaining = p->p_numthreads - p->p_suspcount;
 743         }
 744         if (mode == SINGLE_EXIT) {
 745                 /*
 746                  * We have gotten rid of all the other threads and we
 747                  * are about to either exit or exec. In either case,
 748                  * we try our utmost  to revert to being a non-threaded
 749                  * process.
 750                  */
 751                 p->p_singlethread = NULL;
 752                 p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT);
 753                 thread_unthread(td);
 754         }
 755         PROC_SUNLOCK(p);
 756         return (0);
 757 }
 758
 759 /*
 760  * Called in from locations that can safely check to see
 761  * whether we have to suspend or at least throttle for a
 762  * single-thread event (e.g. fork).
 763  *
 764  * Such locations include userret().
 765  * If the "return_instead" argument is non zero, the thread must be able to
 766  * accept 0 (caller may continue), or 1 (caller must abort) as a result.
 767  *
 768  * The 'return_instead' argument tells the function if it may do a
 769  * thread_exit() or suspend, or whether the caller must abort and back
 770  * out instead.
 771  *
 772  * If the thread that set the single_threading request has set the
 773  * P_SINGLE_EXIT bit in the process flags then this call will never return
 774  * if 'return_instead' is false, but will exit.
 775  *
 776  * P_SINGLE_EXIT | return_instead == 0| return_instead != 0
 777  *---------------+--------------------+---------------------
 778  *       0       | returns 0          |   returns 0 or 1
 779  *               | when ST ends       |   immediatly
 780  *---------------+--------------------+---------------------
 781  *       1       | thread exits       |   returns 1
 782  *               |                    |  immediatly
 783  * 0 = thread_exit() or suspension ok,
 784  * other = return error instead of stopping the thread.
 785  *
 786  * While a full suspension is under effect, even a single threading
 787  * thread would be suspended if it made this call (but it shouldn't).
 788  * This call should only be made from places where
 789  * thread_exit() would be safe as that may be the outcome unless
 790  * return_instead is set.
 791  */
 792 int
 793 thread_suspend_check(int return_instead)
 794 {
 795         struct thread *td;
 796         struct proc *p;
 797
 798         td = curthread;
 799         p = td->td_proc;
 800         mtx_assert(&Giant, MA_NOTOWNED);
 801         PROC_LOCK_ASSERT(p, MA_OWNED);
 802         while (P_SHOULDSTOP(p) ||
 803               ((p->p_flag & P_TRACED) && (td->td_flags & TDF_DBSUSPEND))) {
 804                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 805                         KASSERT(p->p_singlethread != NULL,
 806                             ("singlethread not set"));
 807                         /*
 808                          * The only suspension in action is a
 809                          * single-threading. Single threader need not stop.
 810                          * XXX Should be safe to access unlocked
 811                          * as it can only be set to be true by us.
 812                          */
 813                         if (p->p_singlethread == td)
 814                                 return (0);     /* Exempt from stopping. */
 815                 }
 816                 if ((p->p_flag & P_SINGLE_EXIT) && return_instead)
 817                         return (EINTR);
 818
 819                 /* Should we goto user boundary if we didn't come from there? */
 820                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE &&
 821                     (p->p_flag & P_SINGLE_BOUNDARY) && return_instead)
 822                         return (ERESTART);
 823
 824                 /* If thread will exit, flush its pending signals */
 825                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 826                         sigqueue_flush(&td->td_sigqueue);
 827
 828                 PROC_SLOCK(p);
 829                 thread_stopped(p);
 830                 /*
 831                  * If the process is waiting for us to exit,
 832                  * this thread should just suicide.
 833                  * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE.
 834                  */
 835                 if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td))
 836                         thread_exit();
 837                 if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) {
 838                         if (p->p_numthreads == p->p_suspcount + 1) {
 839                                 thread_lock(p->p_singlethread);
 840                                 thread_unsuspend_one(p->p_singlethread);
 841                                 thread_unlock(p->p_singlethread);
 842                         }
 843                 }
 844                 PROC_UNLOCK(p);
 845                 thread_lock(td);
 846                 /*
 847                  * When a thread suspends, it just
 848                  * gets taken off all queues.
 849                  */
 850                 thread_suspend_one(td);
 851                 if (return_instead == 0) {
 852                         p->p_boundary_count++;
 853                         td->td_flags |= TDF_BOUNDARY;
 854                 }
 855                 PROC_SUNLOCK(p);
 856                 mi_switch(SW_INVOL, NULL);
 857                 if (return_instead == 0)
 858                         td->td_flags &= ~TDF_BOUNDARY;
 859                 thread_unlock(td);
 860                 PROC_LOCK(p);
 861                 if (return_instead == 0)
 862                         p->p_boundary_count--;
 863         }
 864         return (0);
 865 }
 866
 867 void
 868 thread_suspend_switch(struct thread *td)
 869 {
 870         struct proc *p;
 871
 872         p = td->td_proc;
 873         KASSERT(!TD_IS_SUSPENDED(td), ("already suspended"));
 874         PROC_LOCK_ASSERT(p, MA_OWNED);
 875         PROC_SLOCK_ASSERT(p, MA_OWNED);
 876         /*
 877          * We implement thread_suspend_one in stages here to avoid
 878          * dropping the proc lock while the thread lock is owned.
 879          */
 880         thread_stopped(p);
 881         p->p_suspcount++;
 882         PROC_UNLOCK(p);
 883         thread_lock(td);
 884         TD_SET_SUSPENDED(td);
 885         sched_sleep(td, 0);
 886         PROC_SUNLOCK(p);
 887         DROP_GIANT();
 888         mi_switch(SW_VOL, NULL);
 889         thread_unlock(td);
 890         PICKUP_GIANT();
 891         PROC_LOCK(p);
 892         PROC_SLOCK(p);
 893 }
 894
 895 void
 896 thread_suspend_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), ("already suspended"));
 903         p->p_suspcount++;
 904         TD_SET_SUSPENDED(td);
 905         sched_sleep(td, 0);
 906 }
 907
 908 void
 909 thread_unsuspend_one(struct thread *td)
 910 {
 911         struct proc *p = td->td_proc;
 912
 913         PROC_SLOCK_ASSERT(p, MA_OWNED);
 914         THREAD_LOCK_ASSERT(td, MA_OWNED);
 915         KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended"));
 916         TD_CLR_SUSPENDED(td);
 917         p->p_suspcount--;
 918         setrunnable(td);
 919 }
 920
 921 /*
 922  * Allow all threads blocked by single threading to continue running.
 923  */
 924 void
 925 thread_unsuspend(struct proc *p)
 926 {
 927         struct thread *td;
 928
 929         PROC_LOCK_ASSERT(p, MA_OWNED);
 930         PROC_SLOCK_ASSERT(p, MA_OWNED);
 931         if (!P_SHOULDSTOP(p)) {
 932                 FOREACH_THREAD_IN_PROC(p, td) {
 933                         thread_lock(td);
 934                         if (TD_IS_SUSPENDED(td)) {
 935                                 thread_unsuspend_one(td);
 936                         }
 937                         thread_unlock(td);
 938                 }
 939         } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) &&
 940             (p->p_numthreads == p->p_suspcount)) {
 941                 /*
 942                  * Stopping everything also did the job for the single
 943                  * threading request. Now we've downgraded to single-threaded,
 944                  * let it continue.
 945                  */
 946                 thread_lock(p->p_singlethread);
 947                 thread_unsuspend_one(p->p_singlethread);
 948                 thread_unlock(p->p_singlethread);
 949         }
 950 }
 951
 952 /*
 953  * End the single threading mode..
 954  */
 955 void
 956 thread_single_end(void)
 957 {
 958         struct thread *td;
 959         struct proc *p;
 960
 961         td = curthread;
 962         p = td->td_proc;
 963         PROC_LOCK_ASSERT(p, MA_OWNED);
 964         p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY);
 965         PROC_SLOCK(p);
 966         p->p_singlethread = NULL;
 967         /*
 968          * If there are other threads they mey now run,
 969          * unless of course there is a blanket 'stop order'
 970          * on the process. The single threader must be allowed
 971          * to continue however as this is a bad place to stop.
 972          */
 973         if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) {
 974                 FOREACH_THREAD_IN_PROC(p, td) {
 975                         thread_lock(td);
 976                         if (TD_IS_SUSPENDED(td)) {
 977                                 thread_unsuspend_one(td);
 978                         }
 979                         thread_unlock(td);
 980                 }
 981         }
 982         PROC_SUNLOCK(p);
 983 }
 984
 985 struct thread *
 986 thread_find(struct proc *p, lwpid_t tid)
 987 {
 988         struct thread *td;
 989
 990         PROC_LOCK_ASSERT(p, MA_OWNED);
 991         PROC_SLOCK(p);
 992         FOREACH_THREAD_IN_PROC(p, td) {
 993                 if (td->td_tid == tid)
 994                         break;
 995         }
 996         PROC_SUNLOCK(p);
 997         return (td);
 998 }