1 /* 2 * Copyright (c) 1982, 1986, 1989, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)kern_exit.c 8.7 (Berkeley) 2/12/94 35 * $FreeBSD: src/sys/kern/kern_exit.c,v 1.92.2.11 2003/01/13 22:51:16 dillon Exp $ 36 */ 37 38 #include "opt_ktrace.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/sysproto.h> 43 #include <sys/kernel.h> 44 #include <sys/malloc.h> 45 #include <sys/proc.h> 46 #include <sys/ktrace.h> 47 #include <sys/pioctl.h> 48 #include <sys/tty.h> 49 #include <sys/wait.h> 50 #include <sys/vnode.h> 51 #include <sys/resourcevar.h> 52 #include <sys/signalvar.h> 53 #include <sys/taskqueue.h> 54 #include <sys/ptrace.h> 55 #include <sys/acct.h> /* for acct_process() function prototype */ 56 #include <sys/filedesc.h> 57 #include <sys/shm.h> 58 #include <sys/sem.h> 59 #include <sys/jail.h> 60 #include <sys/kern_syscall.h> 61 #include <sys/unistd.h> 62 #include <sys/eventhandler.h> 63 #include <sys/dsched.h> 64 65 #include <vm/vm.h> 66 #include <vm/vm_param.h> 67 #include <sys/lock.h> 68 #include <vm/pmap.h> 69 #include <vm/vm_map.h> 70 #include <vm/vm_extern.h> 71 #include <sys/user.h> 72 73 #include <sys/refcount.h> 74 #include <sys/thread2.h> 75 #include <sys/spinlock2.h> 76 #include <sys/mplock2.h> 77 78 #include <machine/vmm.h> 79 80 static void reaplwps(void *context, int dummy); 81 static void reaplwp(struct lwp *lp); 82 static void killlwps(struct lwp *lp); 83 84 static MALLOC_DEFINE(M_ATEXIT, "atexit", "atexit callback"); 85 86 /* 87 * callout list for things to do at exit time 88 */ 89 struct exitlist { 90 exitlist_fn function; 91 TAILQ_ENTRY(exitlist) next; 92 }; 93 94 TAILQ_HEAD(exit_list_head, exitlist); 95 static struct exit_list_head exit_list = TAILQ_HEAD_INITIALIZER(exit_list); 96 97 /* 98 * LWP reaper data 99 */ 100 static struct task *deadlwp_task[MAXCPU]; 101 static struct lwplist deadlwp_list[MAXCPU]; 102 static struct lwkt_token deadlwp_token[MAXCPU]; 103 104 /* 105 * exit -- 106 * Death of process. 107 * 108 * SYS_EXIT_ARGS(int rval) 109 */ 110 int 111 sys_exit(struct exit_args *uap) 112 { 113 exit1(W_EXITCODE(uap->rval, 0)); 114 /* NOTREACHED */ 115 } 116 117 /* 118 * Extended exit -- 119 * Death of a lwp or process with optional bells and whistles. 120 */ 121 int 122 sys_extexit(struct extexit_args *uap) 123 { 124 struct proc *p = curproc; 125 int action, who; 126 int error; 127 128 action = EXTEXIT_ACTION(uap->how); 129 who = EXTEXIT_WHO(uap->how); 130 131 /* Check parameters before we might perform some action */ 132 switch (who) { 133 case EXTEXIT_PROC: 134 case EXTEXIT_LWP: 135 break; 136 default: 137 return (EINVAL); 138 } 139 140 switch (action) { 141 case EXTEXIT_SIMPLE: 142 break; 143 case EXTEXIT_SETINT: 144 error = copyout(&uap->status, uap->addr, sizeof(uap->status)); 145 if (error) 146 return (error); 147 break; 148 default: 149 return (EINVAL); 150 } 151 152 lwkt_gettoken(&p->p_token); 153 154 switch (who) { 155 case EXTEXIT_LWP: 156 /* 157 * Be sure only to perform a simple lwp exit if there is at 158 * least one more lwp in the proc, which will call exit1() 159 * later, otherwise the proc will be an UNDEAD and not even a 160 * SZOMB! 161 */ 162 if (p->p_nthreads > 1) { 163 lwp_exit(0, NULL); /* called w/ p_token held */ 164 /* NOT REACHED */ 165 } 166 /* else last lwp in proc: do the real thing */ 167 /* FALLTHROUGH */ 168 default: /* to help gcc */ 169 case EXTEXIT_PROC: 170 lwkt_reltoken(&p->p_token); 171 exit1(W_EXITCODE(uap->status, 0)); 172 /* NOTREACHED */ 173 } 174 175 /* NOTREACHED */ 176 lwkt_reltoken(&p->p_token); /* safety */ 177 } 178 179 /* 180 * Kill all lwps associated with the current process except the 181 * current lwp. Return an error if we race another thread trying to 182 * do the same thing and lose the race. 183 * 184 * If forexec is non-zero the current thread and process flags are 185 * cleaned up so they can be reused. 186 * 187 * Caller must hold curproc->p_token 188 */ 189 int 190 killalllwps(int forexec) 191 { 192 struct lwp *lp = curthread->td_lwp; 193 struct proc *p = lp->lwp_proc; 194 int fakestop; 195 196 /* 197 * Interlock against P_WEXIT. Only one of the process's thread 198 * is allowed to do the master exit. 199 */ 200 if (p->p_flags & P_WEXIT) 201 return (EALREADY); 202 p->p_flags |= P_WEXIT; 203 204 /* 205 * Set temporary stopped state in case we are racing a coredump. 206 * Otherwise the coredump may hang forever. 207 */ 208 if (lp->lwp_mpflags & LWP_MP_WSTOP) { 209 fakestop = 0; 210 } else { 211 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 212 ++p->p_nstopped; 213 fakestop = 1; 214 wakeup(&p->p_nstopped); 215 } 216 217 /* 218 * Interlock with LWP_MP_WEXIT and kill any remaining LWPs 219 */ 220 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 221 if (p->p_nthreads > 1) 222 killlwps(lp); 223 224 /* 225 * Undo temporary stopped state 226 */ 227 if (fakestop) { 228 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 229 --p->p_nstopped; 230 } 231 232 /* 233 * If doing this for an exec, clean up the remaining thread 234 * (us) for continuing operation after all the other threads 235 * have been killed. 236 */ 237 if (forexec) { 238 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 239 p->p_flags &= ~P_WEXIT; 240 } 241 return(0); 242 } 243 244 /* 245 * Kill all LWPs except the current one. Do not try to signal 246 * LWPs which have exited on their own or have already been 247 * signaled. 248 */ 249 static void 250 killlwps(struct lwp *lp) 251 { 252 struct proc *p = lp->lwp_proc; 253 struct lwp *tlp; 254 255 /* 256 * Kill the remaining LWPs. We must send the signal before setting 257 * LWP_MP_WEXIT. The setting of WEXIT is optional but helps reduce 258 * races. tlp must be held across the call as it might block and 259 * allow the target lwp to rip itself out from under our loop. 260 */ 261 FOREACH_LWP_IN_PROC(tlp, p) { 262 LWPHOLD(tlp); 263 lwkt_gettoken(&tlp->lwp_token); 264 if ((tlp->lwp_mpflags & LWP_MP_WEXIT) == 0) { 265 atomic_set_int(&tlp->lwp_mpflags, LWP_MP_WEXIT); 266 lwpsignal(p, tlp, SIGKILL); 267 } 268 lwkt_reltoken(&tlp->lwp_token); 269 LWPRELE(tlp); 270 } 271 272 /* 273 * Wait for everything to clear out. Also make sure any tstop()s 274 * are signalled (we are holding p_token for the interlock). 275 */ 276 wakeup(p); 277 while (p->p_nthreads > 1) 278 tsleep(&p->p_nthreads, 0, "killlwps", 0); 279 } 280 281 /* 282 * Exit: deallocate address space and other resources, change proc state 283 * to zombie, and unlink proc from allproc and parent's lists. Save exit 284 * status and rusage for wait(). Check for child processes and orphan them. 285 */ 286 void 287 exit1(int rv) 288 { 289 struct thread *td = curthread; 290 struct proc *p = td->td_proc; 291 struct lwp *lp = td->td_lwp; 292 struct proc *q; 293 struct proc *pp; 294 struct proc *reproc; 295 struct sysreaper *reap; 296 struct vmspace *vm; 297 struct vnode *vtmp; 298 struct exitlist *ep; 299 int error; 300 301 lwkt_gettoken(&p->p_token); 302 303 if (p->p_pid == 1) { 304 kprintf("init died (signal %d, exit %d)\n", 305 WTERMSIG(rv), WEXITSTATUS(rv)); 306 panic("Going nowhere without my init!"); 307 } 308 varsymset_clean(&p->p_varsymset); 309 lockuninit(&p->p_varsymset.vx_lock); 310 311 /* 312 * Kill all lwps associated with the current process, return an 313 * error if we race another thread trying to do the same thing 314 * and lose the race. 315 */ 316 error = killalllwps(0); 317 if (error) { 318 lwp_exit(0, NULL); 319 /* NOT REACHED */ 320 } 321 322 /* are we a task leader? */ 323 if (p == p->p_leader) { 324 struct kill_args killArgs; 325 killArgs.signum = SIGKILL; 326 q = p->p_peers; 327 while(q) { 328 killArgs.pid = q->p_pid; 329 /* 330 * The interface for kill is better 331 * than the internal signal 332 */ 333 sys_kill(&killArgs); 334 q = q->p_peers; 335 } 336 while (p->p_peers) 337 tsleep((caddr_t)p, 0, "exit1", 0); 338 } 339 340 #ifdef PGINPROF 341 vmsizmon(); 342 #endif 343 STOPEVENT(p, S_EXIT, rv); 344 p->p_flags |= P_POSTEXIT; /* stop procfs stepping */ 345 346 /* 347 * Check if any loadable modules need anything done at process exit. 348 * e.g. SYSV IPC stuff 349 * XXX what if one of these generates an error? 350 */ 351 p->p_xstat = rv; 352 353 /* 354 * XXX: imho, the eventhandler stuff is much cleaner than this. 355 * Maybe we should move everything to use eventhandler. 356 */ 357 TAILQ_FOREACH(ep, &exit_list, next) 358 (*ep->function)(td); 359 360 if (p->p_flags & P_PROFIL) 361 stopprofclock(p); 362 363 SIGEMPTYSET(p->p_siglist); 364 SIGEMPTYSET(lp->lwp_siglist); 365 if (timevalisset(&p->p_realtimer.it_value)) 366 callout_stop_sync(&p->p_ithandle); 367 368 /* 369 * Reset any sigio structures pointing to us as a result of 370 * F_SETOWN with our pid. 371 */ 372 funsetownlst(&p->p_sigiolst); 373 374 /* 375 * Close open files and release open-file table. 376 * This may block! 377 */ 378 fdfree(p, NULL); 379 380 if (p->p_leader->p_peers) { 381 q = p->p_leader; 382 while(q->p_peers != p) 383 q = q->p_peers; 384 q->p_peers = p->p_peers; 385 wakeup((caddr_t)p->p_leader); 386 } 387 388 /* 389 * XXX Shutdown SYSV semaphores 390 */ 391 semexit(p); 392 393 KKASSERT(p->p_numposixlocks == 0); 394 395 /* The next two chunks should probably be moved to vmspace_exit. */ 396 vm = p->p_vmspace; 397 398 /* 399 * Clean up data related to virtual kernel operation. Clean up 400 * any vkernel context related to the current lwp now so we can 401 * destroy p_vkernel. 402 */ 403 if (p->p_vkernel) { 404 vkernel_lwp_exit(lp); 405 vkernel_exit(p); 406 } 407 408 /* 409 * Release the user portion of address space. The exitbump prevents 410 * the vmspace from being completely eradicated (using holdcnt). 411 * This releases references to vnodes, which could cause I/O if the 412 * file has been unlinked. We need to do this early enough that 413 * we can still sleep. 414 * 415 * We can't free the entire vmspace as the kernel stack may be mapped 416 * within that space also. 417 * 418 * Processes sharing the same vmspace may exit in one order, and 419 * get cleaned up by vmspace_exit() in a different order. The 420 * last exiting process to reach this point releases as much of 421 * the environment as it can, and the last process cleaned up 422 * by vmspace_exit() (which decrements exitingcnt) cleans up the 423 * remainder. 424 * 425 * NOTE: Releasing p_token around this call is helpful if the 426 * vmspace had a huge RSS. Otherwise some other process 427 * trying to do an allproc or other scan (like 'ps') may 428 * stall for a long time. 429 */ 430 lwkt_reltoken(&p->p_token); 431 vmspace_relexit(vm); 432 lwkt_gettoken(&p->p_token); 433 434 if (SESS_LEADER(p)) { 435 struct session *sp = p->p_session; 436 437 if (sp->s_ttyvp) { 438 /* 439 * We are the controlling process. Signal the 440 * foreground process group, drain the controlling 441 * terminal, and revoke access to the controlling 442 * terminal. 443 * 444 * NOTE: while waiting for the process group to exit 445 * it is possible that one of the processes in the 446 * group will revoke the tty, so the ttyclosesession() 447 * function will re-check sp->s_ttyvp. 448 */ 449 if (sp->s_ttyp && (sp->s_ttyp->t_session == sp)) { 450 if (sp->s_ttyp->t_pgrp) 451 pgsignal(sp->s_ttyp->t_pgrp, SIGHUP, 1); 452 ttywait(sp->s_ttyp); 453 ttyclosesession(sp, 1); /* also revoke */ 454 } 455 /* 456 * Release the tty. If someone has it open via 457 * /dev/tty then close it (since they no longer can 458 * once we've NULL'd it out). 459 */ 460 ttyclosesession(sp, 0); 461 462 /* 463 * s_ttyp is not zero'd; we use this to indicate 464 * that the session once had a controlling terminal. 465 * (for logging and informational purposes) 466 */ 467 } 468 sp->s_leader = NULL; 469 } 470 fixjobc(p, p->p_pgrp, 0); 471 (void)acct_process(p); 472 #ifdef KTRACE 473 /* 474 * release trace file 475 */ 476 if (p->p_tracenode) 477 ktrdestroy(&p->p_tracenode); 478 p->p_traceflag = 0; 479 #endif 480 /* 481 * Release reference to text vnode 482 */ 483 if ((vtmp = p->p_textvp) != NULL) { 484 p->p_textvp = NULL; 485 vrele(vtmp); 486 } 487 488 /* Release namecache handle to text file */ 489 if (p->p_textnch.ncp) 490 cache_drop(&p->p_textnch); 491 492 /* 493 * We have to handle PPWAIT here or proc_move_allproc_zombie() 494 * will block on the PHOLD() the parent is doing. 495 * 496 * We are using the flag as an interlock so an atomic op is 497 * necessary to synchronize with the parent's cpu. 498 */ 499 if (p->p_flags & P_PPWAIT) { 500 if (p->p_pptr && p->p_pptr->p_upmap) 501 atomic_add_int(&p->p_pptr->p_upmap->invfork, -1); 502 atomic_clear_int(&p->p_flags, P_PPWAIT); 503 wakeup(p->p_pptr); 504 } 505 506 /* 507 * Move the process to the zombie list. This will block 508 * until the process p_lock count reaches 0. The process will 509 * not be reaped until TDF_EXITING is set by cpu_thread_exit(), 510 * which is called from cpu_proc_exit(). 511 * 512 * Interlock against waiters using p_waitgen. We increment 513 * p_waitgen after completing the move of our process to the 514 * zombie list. 515 * 516 * WARNING: pp becomes stale when we block, clear it now as a 517 * reminder. 518 */ 519 proc_move_allproc_zombie(p); 520 pp = p->p_pptr; 521 atomic_add_long(&pp->p_waitgen, 1); 522 pp = NULL; 523 524 /* 525 * release controlled reaper for exit if we own it and return the 526 * remaining reaper (the one for us), which we will drop after we 527 * are done. 528 */ 529 reap = reaper_exit(p); 530 531 /* 532 * Reparent all of this process's children to the init process or 533 * to the designated reaper. We must hold the reaper's p_token in 534 * order to safely mess with p_children. 535 * 536 * We already hold p->p_token (to remove the children from our list). 537 */ 538 reproc = NULL; 539 q = LIST_FIRST(&p->p_children); 540 if (q) { 541 reproc = reaper_get(reap); 542 lwkt_gettoken(&reproc->p_token); 543 while ((q = LIST_FIRST(&p->p_children)) != NULL) { 544 PHOLD(q); 545 lwkt_gettoken(&q->p_token); 546 if (q != LIST_FIRST(&p->p_children)) { 547 lwkt_reltoken(&q->p_token); 548 PRELE(q); 549 continue; 550 } 551 LIST_REMOVE(q, p_sibling); 552 LIST_INSERT_HEAD(&reproc->p_children, q, p_sibling); 553 q->p_pptr = reproc; 554 q->p_ppid = reproc->p_pid; 555 q->p_sigparent = SIGCHLD; 556 557 /* 558 * Traced processes are killed 559 * since their existence means someone is screwing up. 560 */ 561 if (q->p_flags & P_TRACED) { 562 q->p_flags &= ~P_TRACED; 563 ksignal(q, SIGKILL); 564 } 565 lwkt_reltoken(&q->p_token); 566 PRELE(q); 567 } 568 lwkt_reltoken(&reproc->p_token); 569 wakeup(reproc); 570 } 571 572 /* 573 * Save exit status and final rusage info, adding in child rusage 574 * info and self times. 575 */ 576 calcru_proc(p, &p->p_ru); 577 ruadd(&p->p_ru, &p->p_cru); 578 579 /* 580 * notify interested parties of our demise. 581 */ 582 KNOTE(&p->p_klist, NOTE_EXIT); 583 584 /* 585 * Notify parent that we're gone. If parent has the PS_NOCLDWAIT 586 * flag set, or if the handler is set to SIG_IGN, notify the reaper 587 * instead (it will handle this situation). 588 * 589 * NOTE: The reaper can still be the parent process. 590 * 591 * (must reload pp) 592 */ 593 if (p->p_pptr->p_sigacts->ps_flag & (PS_NOCLDWAIT | PS_CLDSIGIGN)) { 594 if (reproc == NULL) 595 reproc = reaper_get(reap); 596 proc_reparent(p, reproc); 597 } 598 if (reproc) 599 PRELE(reproc); 600 if (reap) 601 reaper_drop(reap); 602 603 /* 604 * Signal (possibly new) parent. 605 */ 606 pp = p->p_pptr; 607 PHOLD(pp); 608 if (p->p_sigparent && pp != initproc) { 609 int sig = p->p_sigparent; 610 611 if (sig != SIGUSR1 && sig != SIGCHLD) 612 sig = SIGCHLD; 613 ksignal(pp, sig); 614 } else { 615 ksignal(pp, SIGCHLD); 616 } 617 p->p_flags &= ~P_TRACED; 618 PRELE(pp); 619 620 /* 621 * cpu_exit is responsible for clearing curproc, since 622 * it is heavily integrated with the thread/switching sequence. 623 * 624 * Other substructures are freed from wait(). 625 */ 626 if (p->p_limit) { 627 struct plimit *rlimit; 628 629 rlimit = p->p_limit; 630 p->p_limit = NULL; 631 plimit_free(rlimit); 632 } 633 634 /* 635 * Finally, call machine-dependent code to release as many of the 636 * lwp's resources as we can and halt execution of this thread. 637 * 638 * pp is a wild pointer now but still the correct wakeup() target. 639 * lwp_exit() only uses it to send the wakeup() signal to the likely 640 * parent. Any reparenting race that occurs will get a signal 641 * automatically and not be an issue. 642 */ 643 lwp_exit(1, pp); 644 } 645 646 /* 647 * Eventually called by every exiting LWP 648 * 649 * p->p_token must be held. mplock may be held and will be released. 650 */ 651 void 652 lwp_exit(int masterexit, void *waddr) 653 { 654 struct thread *td = curthread; 655 struct lwp *lp = td->td_lwp; 656 struct proc *p = lp->lwp_proc; 657 int dowake = 0; 658 659 /* 660 * Release the current user process designation on the process so 661 * the userland scheduler can work in someone else. 662 */ 663 p->p_usched->release_curproc(lp); 664 665 /* 666 * lwp_exit() may be called without setting LWP_MP_WEXIT, so 667 * make sure it is set here. 668 */ 669 ASSERT_LWKT_TOKEN_HELD(&p->p_token); 670 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 671 672 /* 673 * Clean up any virtualization 674 */ 675 if (lp->lwp_vkernel) 676 vkernel_lwp_exit(lp); 677 678 if (td->td_vmm) 679 vmm_vmdestroy(); 680 681 /* 682 * Clean up select/poll support 683 */ 684 kqueue_terminate(&lp->lwp_kqueue); 685 686 /* 687 * Clean up any syscall-cached ucred or rlimit. 688 */ 689 if (td->td_ucred) { 690 crfree(td->td_ucred); 691 td->td_ucred = NULL; 692 } 693 if (td->td_limit) { 694 struct plimit *rlimit; 695 696 rlimit = td->td_limit; 697 td->td_limit = NULL; 698 plimit_free(rlimit); 699 } 700 701 /* 702 * Nobody actually wakes us when the lock 703 * count reaches zero, so just wait one tick. 704 */ 705 while (lp->lwp_lock > 0) 706 tsleep(lp, 0, "lwpexit", 1); 707 708 /* Hand down resource usage to our proc */ 709 ruadd(&p->p_ru, &lp->lwp_ru); 710 711 /* 712 * If we don't hold the process until the LWP is reaped wait*() 713 * may try to dispose of its vmspace before all the LWPs have 714 * actually terminated. 715 */ 716 PHOLD(p); 717 718 /* 719 * Do any remaining work that might block on us. We should be 720 * coded such that further blocking is ok after decrementing 721 * p_nthreads but don't take the chance. 722 */ 723 dsched_exit_thread(td); 724 biosched_done(curthread); 725 726 /* 727 * We have to use the reaper for all the LWPs except the one doing 728 * the master exit. The LWP doing the master exit can just be 729 * left on p_lwps and the process reaper will deal with it 730 * synchronously, which is much faster. 731 * 732 * Wakeup anyone waiting on p_nthreads to drop to 1 or 0. 733 * 734 * The process is left held until the reaper calls lwp_dispose() on 735 * the lp (after calling lwp_wait()). 736 */ 737 if (masterexit == 0) { 738 int cpu = mycpuid; 739 740 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 741 --p->p_nthreads; 742 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 743 dowake = 1; 744 lwkt_gettoken(&deadlwp_token[cpu]); 745 LIST_INSERT_HEAD(&deadlwp_list[cpu], lp, u.lwp_reap_entry); 746 taskqueue_enqueue(taskqueue_thread[cpu], deadlwp_task[cpu]); 747 lwkt_reltoken(&deadlwp_token[cpu]); 748 } else { 749 --p->p_nthreads; 750 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 751 dowake = 1; 752 } 753 754 /* 755 * We no longer need p_token. 756 * 757 * Tell the userland scheduler that we are going away 758 */ 759 lwkt_reltoken(&p->p_token); 760 p->p_usched->heuristic_exiting(lp, p); 761 762 /* 763 * Issue late wakeups after releasing our token to give us a chance 764 * to deschedule and switch away before another cpu in a wait*() 765 * reaps us. This is done as late as possible to reduce contention. 766 */ 767 if (dowake) 768 wakeup(&p->p_nthreads); 769 if (waddr) 770 wakeup(waddr); 771 772 cpu_lwp_exit(); 773 } 774 775 /* 776 * Wait until a lwp is completely dead. The final interlock in this drama 777 * is when TDF_EXITING is set in cpu_thread_exit() just before the final 778 * switchout. 779 * 780 * At the point TDF_EXITING is set a complete exit is accomplished when 781 * TDF_RUNNING and TDF_PREEMPT_LOCK are both clear. td_mpflags has two 782 * post-switch interlock flags that can be used to wait for the TDF_ 783 * flags to clear. 784 * 785 * Returns non-zero on success, and zero if the caller needs to retry 786 * the lwp_wait(). 787 */ 788 static int 789 lwp_wait(struct lwp *lp) 790 { 791 struct thread *td = lp->lwp_thread; 792 u_int mpflags; 793 794 KKASSERT(lwkt_preempted_proc() != lp); 795 796 /* 797 * This bit of code uses the thread destruction interlock 798 * managed by lwkt_switch_return() to wait for the lwp's 799 * thread to completely disengage. 800 * 801 * It is possible for us to race another cpu core so we 802 * have to do this correctly. 803 */ 804 for (;;) { 805 mpflags = td->td_mpflags; 806 cpu_ccfence(); 807 if (mpflags & TDF_MP_EXITSIG) 808 break; 809 tsleep_interlock(td, 0); 810 if (atomic_cmpset_int(&td->td_mpflags, mpflags, 811 mpflags | TDF_MP_EXITWAIT)) { 812 tsleep(td, PINTERLOCKED, "lwpxt", 0); 813 } 814 } 815 816 /* 817 * We've already waited for the core exit but there can still 818 * be other refs from e.g. process scans and such. 819 */ 820 if (lp->lwp_lock > 0) { 821 tsleep(lp, 0, "lwpwait1", 1); 822 return(0); 823 } 824 if (td->td_refs) { 825 tsleep(td, 0, "lwpwait2", 1); 826 return(0); 827 } 828 829 /* 830 * Now that we have the thread destruction interlock these flags 831 * really should already be cleaned up, keep a check for safety. 832 * 833 * We can't rip its stack out from under it until TDF_EXITING is 834 * set and both TDF_RUNNING and TDF_PREEMPT_LOCK are clear. 835 * TDF_PREEMPT_LOCK must be checked because TDF_RUNNING 836 * will be cleared temporarily if a thread gets preempted. 837 */ 838 while ((td->td_flags & (TDF_RUNNING | 839 TDF_RUNQ | 840 TDF_PREEMPT_LOCK | 841 TDF_EXITING)) != TDF_EXITING) { 842 tsleep(lp, 0, "lwpwait3", 1); 843 return (0); 844 } 845 846 KASSERT((td->td_flags & (TDF_RUNQ|TDF_TSLEEPQ)) == 0, 847 ("lwp_wait: td %p (%s) still on run or sleep queue", 848 td, td->td_comm)); 849 return (1); 850 } 851 852 /* 853 * Release the resources associated with a lwp. 854 * The lwp must be completely dead. 855 */ 856 void 857 lwp_dispose(struct lwp *lp) 858 { 859 struct thread *td = lp->lwp_thread; 860 861 KKASSERT(lwkt_preempted_proc() != lp); 862 KKASSERT(lp->lwp_lock == 0); 863 KKASSERT(td->td_refs == 0); 864 KKASSERT((td->td_flags & (TDF_RUNNING | 865 TDF_RUNQ | 866 TDF_PREEMPT_LOCK | 867 TDF_EXITING)) == TDF_EXITING); 868 869 PRELE(lp->lwp_proc); 870 lp->lwp_proc = NULL; 871 if (td != NULL) { 872 td->td_proc = NULL; 873 td->td_lwp = NULL; 874 lp->lwp_thread = NULL; 875 lwkt_free_thread(td); 876 } 877 kfree(lp, M_LWP); 878 } 879 880 int 881 sys_wait4(struct wait_args *uap) 882 { 883 struct rusage rusage; 884 int error, status; 885 886 error = kern_wait(uap->pid, (uap->status ? &status : NULL), 887 uap->options, (uap->rusage ? &rusage : NULL), 888 &uap->sysmsg_result); 889 890 if (error == 0 && uap->status) 891 error = copyout(&status, uap->status, sizeof(*uap->status)); 892 if (error == 0 && uap->rusage) 893 error = copyout(&rusage, uap->rusage, sizeof(*uap->rusage)); 894 return (error); 895 } 896 897 /* 898 * wait1() 899 * 900 * wait_args(int pid, int *status, int options, struct rusage *rusage) 901 */ 902 int 903 kern_wait(pid_t pid, int *status, int options, struct rusage *rusage, int *res) 904 { 905 struct thread *td = curthread; 906 struct lwp *lp; 907 struct proc *q = td->td_proc; 908 struct proc *p, *t; 909 struct ucred *cr; 910 struct pargs *pa; 911 struct sigacts *ps; 912 int nfound, error; 913 long waitgen; 914 915 if (pid == 0) 916 pid = -q->p_pgid; 917 if (options &~ (WUNTRACED|WNOHANG|WCONTINUED|WLINUXCLONE)) 918 return (EINVAL); 919 920 /* 921 * Protect the q->p_children list 922 */ 923 lwkt_gettoken(&q->p_token); 924 loop: 925 /* 926 * All sorts of things can change due to blocking so we have to loop 927 * all the way back up here. 928 * 929 * The problem is that if a process group is stopped and the parent 930 * is doing a wait*(..., WUNTRACED, ...), it will see the STOP 931 * of the child and then stop itself when it tries to return from the 932 * system call. When the process group is resumed the parent will 933 * then get the STOP status even though the child has now resumed 934 * (a followup wait*() will get the CONT status). 935 * 936 * Previously the CONT would overwrite the STOP because the tstop 937 * was handled within tsleep(), and the parent would only see 938 * the CONT when both are stopped and continued together. This little 939 * two-line hack restores this effect. 940 */ 941 if (STOPLWP(q, td->td_lwp)) 942 tstop(); 943 944 nfound = 0; 945 946 /* 947 * Loop on children. 948 * 949 * NOTE: We don't want to break q's p_token in the loop for the 950 * case where no children are found or we risk breaking the 951 * interlock between child and parent. 952 */ 953 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 954 LIST_FOREACH(p, &q->p_children, p_sibling) { 955 if (pid != WAIT_ANY && 956 p->p_pid != pid && p->p_pgid != -pid) { 957 continue; 958 } 959 960 /* 961 * This special case handles a kthread spawned by linux_clone 962 * (see linux_misc.c). The linux_wait4 and linux_waitpid 963 * functions need to be able to distinguish between waiting 964 * on a process and waiting on a thread. It is a thread if 965 * p_sigparent is not SIGCHLD, and the WLINUXCLONE option 966 * signifies we want to wait for threads and not processes. 967 */ 968 if ((p->p_sigparent != SIGCHLD) ^ 969 ((options & WLINUXCLONE) != 0)) { 970 continue; 971 } 972 973 nfound++; 974 if (p->p_stat == SZOMB) { 975 /* 976 * We may go into SZOMB with threads still present. 977 * We must wait for them to exit before we can reap 978 * the master thread, otherwise we may race reaping 979 * non-master threads. 980 * 981 * Only this routine can remove a process from 982 * the zombie list and destroy it, use PACQUIREZOMB() 983 * to serialize us and loop if it blocks (interlocked 984 * by the parent's q->p_token). 985 * 986 * WARNING! (p) can be invalid when PHOLDZOMB(p) 987 * returns non-zero. Be sure not to 988 * mess with it. 989 */ 990 if (PHOLDZOMB(p)) 991 goto loop; 992 lwkt_gettoken(&p->p_token); 993 if (p->p_pptr != q) { 994 lwkt_reltoken(&p->p_token); 995 PRELEZOMB(p); 996 goto loop; 997 } 998 while (p->p_nthreads > 0) { 999 tsleep(&p->p_nthreads, 0, "lwpzomb", hz); 1000 } 1001 1002 /* 1003 * Reap any LWPs left in p->p_lwps. This is usually 1004 * just the last LWP. This must be done before 1005 * we loop on p_lock since the lwps hold a ref on 1006 * it as a vmspace interlock. 1007 * 1008 * Once that is accomplished p_nthreads had better 1009 * be zero. 1010 */ 1011 while ((lp = RB_ROOT(&p->p_lwp_tree)) != NULL) { 1012 /* 1013 * Make sure no one is using this lwp, before 1014 * it is removed from the tree. If we didn't 1015 * wait it here, lwp tree iteration with 1016 * blocking operation would be broken. 1017 */ 1018 while (lp->lwp_lock > 0) 1019 tsleep(lp, 0, "zomblwp", 1); 1020 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 1021 reaplwp(lp); 1022 } 1023 KKASSERT(p->p_nthreads == 0); 1024 1025 /* 1026 * Don't do anything really bad until all references 1027 * to the process go away. This may include other 1028 * LWPs which are still in the process of being 1029 * reaped. We can't just pull the rug out from under 1030 * them because they may still be using the VM space. 1031 * 1032 * Certain kernel facilities such as /proc will also 1033 * put a hold on the process for short periods of 1034 * time. 1035 */ 1036 PRELE(p); 1037 PSTALL(p, "reap3", 0); 1038 1039 /* Take care of our return values. */ 1040 *res = p->p_pid; 1041 1042 if (status) 1043 *status = p->p_xstat; 1044 if (rusage) 1045 *rusage = p->p_ru; 1046 1047 /* 1048 * If we got the child via a ptrace 'attach', 1049 * we need to give it back to the old parent. 1050 */ 1051 if (p->p_oppid && (t = pfind(p->p_oppid)) != NULL) { 1052 PHOLD(p); 1053 p->p_oppid = 0; 1054 proc_reparent(p, t); 1055 ksignal(t, SIGCHLD); 1056 wakeup((caddr_t)t); 1057 error = 0; 1058 PRELE(t); 1059 lwkt_reltoken(&p->p_token); 1060 PRELEZOMB(p); 1061 goto done; 1062 } 1063 1064 /* 1065 * Unlink the proc from its process group so that 1066 * the following operations won't lead to an 1067 * inconsistent state for processes running down 1068 * the zombie list. 1069 */ 1070 proc_remove_zombie(p); 1071 proc_userunmap(p); 1072 lwkt_reltoken(&p->p_token); 1073 leavepgrp(p); 1074 1075 p->p_xstat = 0; 1076 ruadd(&q->p_cru, &p->p_ru); 1077 1078 /* 1079 * Decrement the count of procs running with this uid. 1080 */ 1081 chgproccnt(p->p_ucred->cr_ruidinfo, -1, 0); 1082 1083 /* 1084 * Free up credentials. p_spin is required to 1085 * avoid races against allproc scans. 1086 */ 1087 spin_lock(&p->p_spin); 1088 cr = p->p_ucred; 1089 p->p_ucred = NULL; 1090 spin_unlock(&p->p_spin); 1091 crfree(cr); 1092 1093 /* 1094 * Remove unused arguments 1095 */ 1096 pa = p->p_args; 1097 p->p_args = NULL; 1098 if (pa && refcount_release(&pa->ar_ref)) { 1099 kfree(pa, M_PARGS); 1100 pa = NULL; 1101 } 1102 1103 ps = p->p_sigacts; 1104 p->p_sigacts = NULL; 1105 if (ps && refcount_release(&ps->ps_refcnt)) { 1106 kfree(ps, M_SUBPROC); 1107 ps = NULL; 1108 } 1109 1110 /* 1111 * Our exitingcount was incremented when the process 1112 * became a zombie, now that the process has been 1113 * removed from (almost) all lists we should be able 1114 * to safely destroy its vmspace. Wait for any current 1115 * holders to go away (so the vmspace remains stable), 1116 * then scrap it. 1117 * 1118 * NOTE: Releasing the parent process (q) p_token 1119 * across the vmspace_exitfree() call is 1120 * important here to reduce stalls on 1121 * interactions with (q) (such as 1122 * fork/exec/wait or 'ps'). 1123 */ 1124 PSTALL(p, "reap4", 0); 1125 lwkt_reltoken(&q->p_token); 1126 vmspace_exitfree(p); 1127 lwkt_gettoken(&q->p_token); 1128 PSTALL(p, "reap5", 0); 1129 1130 /* 1131 * NOTE: We have to officially release ZOMB in order 1132 * to ensure that a racing thread in kern_wait() 1133 * which blocked on ZOMB is woken up. 1134 */ 1135 PHOLD(p); 1136 PRELEZOMB(p); 1137 kfree(p, M_PROC); 1138 atomic_add_int(&nprocs, -1); 1139 error = 0; 1140 goto done; 1141 } 1142 if ((p->p_stat == SSTOP || p->p_stat == SCORE) && 1143 (p->p_flags & P_WAITED) == 0 && 1144 ((p->p_flags & P_TRACED) || (options & WUNTRACED))) { 1145 PHOLD(p); 1146 lwkt_gettoken(&p->p_token); 1147 if (p->p_pptr != q) { 1148 lwkt_reltoken(&p->p_token); 1149 PRELE(p); 1150 goto loop; 1151 } 1152 if ((p->p_stat != SSTOP && p->p_stat != SCORE) || 1153 (p->p_flags & P_WAITED) != 0 || 1154 ((p->p_flags & P_TRACED) == 0 && 1155 (options & WUNTRACED) == 0)) { 1156 lwkt_reltoken(&p->p_token); 1157 PRELE(p); 1158 goto loop; 1159 } 1160 1161 p->p_flags |= P_WAITED; 1162 1163 *res = p->p_pid; 1164 if (status) 1165 *status = W_STOPCODE(p->p_xstat); 1166 /* Zero rusage so we get something consistent. */ 1167 if (rusage) 1168 bzero(rusage, sizeof(*rusage)); 1169 error = 0; 1170 lwkt_reltoken(&p->p_token); 1171 PRELE(p); 1172 goto done; 1173 } 1174 if ((options & WCONTINUED) && (p->p_flags & P_CONTINUED)) { 1175 PHOLD(p); 1176 lwkt_gettoken(&p->p_token); 1177 if (p->p_pptr != q) { 1178 lwkt_reltoken(&p->p_token); 1179 PRELE(p); 1180 goto loop; 1181 } 1182 if ((p->p_flags & P_CONTINUED) == 0) { 1183 lwkt_reltoken(&p->p_token); 1184 PRELE(p); 1185 goto loop; 1186 } 1187 1188 *res = p->p_pid; 1189 p->p_flags &= ~P_CONTINUED; 1190 1191 if (status) 1192 *status = SIGCONT; 1193 error = 0; 1194 lwkt_reltoken(&p->p_token); 1195 PRELE(p); 1196 goto done; 1197 } 1198 } 1199 if (nfound == 0) { 1200 error = ECHILD; 1201 goto done; 1202 } 1203 if (options & WNOHANG) { 1204 *res = 0; 1205 error = 0; 1206 goto done; 1207 } 1208 1209 /* 1210 * Wait for signal - interlocked using q->p_waitgen. 1211 */ 1212 error = 0; 1213 while ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1214 tsleep_interlock(q, PCATCH); 1215 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 1216 if ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1217 error = tsleep(q, PCATCH | PINTERLOCKED, "wait", 0); 1218 break; 1219 } 1220 } 1221 if (error) { 1222 done: 1223 lwkt_reltoken(&q->p_token); 1224 return (error); 1225 } 1226 goto loop; 1227 } 1228 1229 /* 1230 * Change child's parent process to parent. 1231 * 1232 * p_children/p_sibling requires the parent's token, and 1233 * changing pptr requires the child's token, so we have to 1234 * get three tokens to do this operation. We also need to 1235 * hold pointers that might get ripped out from under us to 1236 * preserve structural integrity. 1237 * 1238 * It is possible to race another reparent or disconnect or other 1239 * similar operation. We must retry when this situation occurs. 1240 * Once we successfully reparent the process we no longer care 1241 * about any races. 1242 */ 1243 void 1244 proc_reparent(struct proc *child, struct proc *parent) 1245 { 1246 struct proc *opp; 1247 1248 PHOLD(parent); 1249 while ((opp = child->p_pptr) != parent) { 1250 PHOLD(opp); 1251 lwkt_gettoken(&opp->p_token); 1252 lwkt_gettoken(&child->p_token); 1253 lwkt_gettoken(&parent->p_token); 1254 if (child->p_pptr != opp) { 1255 lwkt_reltoken(&parent->p_token); 1256 lwkt_reltoken(&child->p_token); 1257 lwkt_reltoken(&opp->p_token); 1258 PRELE(opp); 1259 continue; 1260 } 1261 LIST_REMOVE(child, p_sibling); 1262 LIST_INSERT_HEAD(&parent->p_children, child, p_sibling); 1263 child->p_pptr = parent; 1264 child->p_ppid = parent->p_pid; 1265 lwkt_reltoken(&parent->p_token); 1266 lwkt_reltoken(&child->p_token); 1267 lwkt_reltoken(&opp->p_token); 1268 if (LIST_EMPTY(&opp->p_children)) 1269 wakeup(opp); 1270 PRELE(opp); 1271 break; 1272 } 1273 PRELE(parent); 1274 } 1275 1276 /* 1277 * The next two functions are to handle adding/deleting items on the 1278 * exit callout list 1279 * 1280 * at_exit(): 1281 * Take the arguments given and put them onto the exit callout list, 1282 * However first make sure that it's not already there. 1283 * returns 0 on success. 1284 */ 1285 1286 int 1287 at_exit(exitlist_fn function) 1288 { 1289 struct exitlist *ep; 1290 1291 #ifdef INVARIANTS 1292 /* Be noisy if the programmer has lost track of things */ 1293 if (rm_at_exit(function)) 1294 kprintf("WARNING: exit callout entry (%p) already present\n", 1295 function); 1296 #endif 1297 ep = kmalloc(sizeof(*ep), M_ATEXIT, M_NOWAIT); 1298 if (ep == NULL) 1299 return (ENOMEM); 1300 ep->function = function; 1301 TAILQ_INSERT_TAIL(&exit_list, ep, next); 1302 return (0); 1303 } 1304 1305 /* 1306 * Scan the exit callout list for the given item and remove it. 1307 * Returns the number of items removed (0 or 1) 1308 */ 1309 int 1310 rm_at_exit(exitlist_fn function) 1311 { 1312 struct exitlist *ep; 1313 1314 TAILQ_FOREACH(ep, &exit_list, next) { 1315 if (ep->function == function) { 1316 TAILQ_REMOVE(&exit_list, ep, next); 1317 kfree(ep, M_ATEXIT); 1318 return(1); 1319 } 1320 } 1321 return (0); 1322 } 1323 1324 /* 1325 * LWP reaper related code. 1326 */ 1327 static void 1328 reaplwps(void *context, int dummy) 1329 { 1330 struct lwplist *lwplist = context; 1331 struct lwp *lp; 1332 int cpu = mycpuid; 1333 1334 lwkt_gettoken(&deadlwp_token[cpu]); 1335 while ((lp = LIST_FIRST(lwplist))) { 1336 LIST_REMOVE(lp, u.lwp_reap_entry); 1337 reaplwp(lp); 1338 } 1339 lwkt_reltoken(&deadlwp_token[cpu]); 1340 } 1341 1342 static void 1343 reaplwp(struct lwp *lp) 1344 { 1345 while (lwp_wait(lp) == 0) 1346 ; 1347 lwp_dispose(lp); 1348 } 1349 1350 static void 1351 deadlwp_init(void) 1352 { 1353 int cpu; 1354 1355 for (cpu = 0; cpu < ncpus; cpu++) { 1356 lwkt_token_init(&deadlwp_token[cpu], "deadlwpl"); 1357 LIST_INIT(&deadlwp_list[cpu]); 1358 deadlwp_task[cpu] = kmalloc(sizeof(*deadlwp_task[cpu]), 1359 M_DEVBUF, M_WAITOK); 1360 TASK_INIT(deadlwp_task[cpu], 0, reaplwps, &deadlwp_list[cpu]); 1361 } 1362 } 1363 1364 SYSINIT(deadlwpinit, SI_SUB_CONFIGURE, SI_ORDER_ANY, deadlwp_init, NULL); 1365