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/sysref2.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_sigparent = SIGCHLD; 555 556 /* 557 * Traced processes are killed 558 * since their existence means someone is screwing up. 559 */ 560 if (q->p_flags & P_TRACED) { 561 q->p_flags &= ~P_TRACED; 562 ksignal(q, SIGKILL); 563 } 564 lwkt_reltoken(&q->p_token); 565 PRELE(q); 566 } 567 lwkt_reltoken(&reproc->p_token); 568 wakeup(reproc); 569 } 570 571 /* 572 * Save exit status and final rusage info, adding in child rusage 573 * info and self times. 574 */ 575 calcru_proc(p, &p->p_ru); 576 ruadd(&p->p_ru, &p->p_cru); 577 578 /* 579 * notify interested parties of our demise. 580 */ 581 KNOTE(&p->p_klist, NOTE_EXIT); 582 583 /* 584 * Notify parent that we're gone. If parent has the PS_NOCLDWAIT 585 * flag set, or if the handler is set to SIG_IGN, notify the reaper 586 * instead (it will handle this situation). 587 * 588 * NOTE: The reaper can still be the parent process. 589 * 590 * (must reload pp) 591 */ 592 if (p->p_pptr->p_sigacts->ps_flag & (PS_NOCLDWAIT | PS_CLDSIGIGN)) { 593 if (reproc == NULL) 594 reproc = reaper_get(reap); 595 proc_reparent(p, reproc); 596 } 597 if (reproc) 598 PRELE(reproc); 599 if (reap) 600 reaper_drop(reap); 601 602 /* 603 * Signal (possibly new) parent. 604 */ 605 pp = p->p_pptr; 606 PHOLD(pp); 607 if (p->p_sigparent && pp != initproc) { 608 int sig = p->p_sigparent; 609 610 if (sig != SIGUSR1 && sig != SIGCHLD) 611 sig = SIGCHLD; 612 ksignal(pp, sig); 613 } else { 614 ksignal(pp, SIGCHLD); 615 } 616 p->p_flags &= ~P_TRACED; 617 PRELE(pp); 618 619 /* 620 * cpu_exit is responsible for clearing curproc, since 621 * it is heavily integrated with the thread/switching sequence. 622 * 623 * Other substructures are freed from wait(). 624 */ 625 plimit_free(p); 626 627 /* 628 * Finally, call machine-dependent code to release as many of the 629 * lwp's resources as we can and halt execution of this thread. 630 * 631 * pp is a wild pointer now but still the correct wakeup() target. 632 * lwp_exit() only uses it to send the wakeup() signal to the likely 633 * parent. Any reparenting race that occurs will get a signal 634 * automatically and not be an issue. 635 */ 636 lwp_exit(1, pp); 637 } 638 639 /* 640 * Eventually called by every exiting LWP 641 * 642 * p->p_token must be held. mplock may be held and will be released. 643 */ 644 void 645 lwp_exit(int masterexit, void *waddr) 646 { 647 struct thread *td = curthread; 648 struct lwp *lp = td->td_lwp; 649 struct proc *p = lp->lwp_proc; 650 int dowake = 0; 651 652 /* 653 * Release the current user process designation on the process so 654 * the userland scheduler can work in someone else. 655 */ 656 p->p_usched->release_curproc(lp); 657 658 /* 659 * lwp_exit() may be called without setting LWP_MP_WEXIT, so 660 * make sure it is set here. 661 */ 662 ASSERT_LWKT_TOKEN_HELD(&p->p_token); 663 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 664 665 /* 666 * Clean up any virtualization 667 */ 668 if (lp->lwp_vkernel) 669 vkernel_lwp_exit(lp); 670 671 if (td->td_vmm) 672 vmm_vmdestroy(); 673 674 /* 675 * Clean up select/poll support 676 */ 677 kqueue_terminate(&lp->lwp_kqueue); 678 679 /* 680 * Clean up any syscall-cached ucred 681 */ 682 if (td->td_ucred) { 683 crfree(td->td_ucred); 684 td->td_ucred = NULL; 685 } 686 687 /* 688 * Nobody actually wakes us when the lock 689 * count reaches zero, so just wait one tick. 690 */ 691 while (lp->lwp_lock > 0) 692 tsleep(lp, 0, "lwpexit", 1); 693 694 /* Hand down resource usage to our proc */ 695 ruadd(&p->p_ru, &lp->lwp_ru); 696 697 /* 698 * If we don't hold the process until the LWP is reaped wait*() 699 * may try to dispose of its vmspace before all the LWPs have 700 * actually terminated. 701 */ 702 PHOLD(p); 703 704 /* 705 * Do any remaining work that might block on us. We should be 706 * coded such that further blocking is ok after decrementing 707 * p_nthreads but don't take the chance. 708 */ 709 dsched_exit_thread(td); 710 biosched_done(curthread); 711 712 /* 713 * We have to use the reaper for all the LWPs except the one doing 714 * the master exit. The LWP doing the master exit can just be 715 * left on p_lwps and the process reaper will deal with it 716 * synchronously, which is much faster. 717 * 718 * Wakeup anyone waiting on p_nthreads to drop to 1 or 0. 719 * 720 * The process is left held until the reaper calls lwp_dispose() on 721 * the lp (after calling lwp_wait()). 722 */ 723 if (masterexit == 0) { 724 int cpu = mycpuid; 725 726 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 727 --p->p_nthreads; 728 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 729 dowake = 1; 730 lwkt_gettoken(&deadlwp_token[cpu]); 731 LIST_INSERT_HEAD(&deadlwp_list[cpu], lp, u.lwp_reap_entry); 732 taskqueue_enqueue(taskqueue_thread[cpu], deadlwp_task[cpu]); 733 lwkt_reltoken(&deadlwp_token[cpu]); 734 } else { 735 --p->p_nthreads; 736 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 737 dowake = 1; 738 } 739 740 /* 741 * We no longer need p_token. 742 * 743 * Tell the userland scheduler that we are going away 744 */ 745 lwkt_reltoken(&p->p_token); 746 p->p_usched->heuristic_exiting(lp, p); 747 748 /* 749 * Issue late wakeups after releasing our token to give us a chance 750 * to deschedule and switch away before another cpu in a wait*() 751 * reaps us. This is done as late as possible to reduce contention. 752 */ 753 if (dowake) 754 wakeup(&p->p_nthreads); 755 if (waddr) 756 wakeup(waddr); 757 758 cpu_lwp_exit(); 759 } 760 761 /* 762 * Wait until a lwp is completely dead. The final interlock in this drama 763 * is when TDF_EXITING is set in cpu_thread_exit() just before the final 764 * switchout. 765 * 766 * At the point TDF_EXITING is set a complete exit is accomplished when 767 * TDF_RUNNING and TDF_PREEMPT_LOCK are both clear. td_mpflags has two 768 * post-switch interlock flags that can be used to wait for the TDF_ 769 * flags to clear. 770 * 771 * Returns non-zero on success, and zero if the caller needs to retry 772 * the lwp_wait(). 773 */ 774 static int 775 lwp_wait(struct lwp *lp) 776 { 777 struct thread *td = lp->lwp_thread; 778 u_int mpflags; 779 780 KKASSERT(lwkt_preempted_proc() != lp); 781 782 /* 783 * This bit of code uses the thread destruction interlock 784 * managed by lwkt_switch_return() to wait for the lwp's 785 * thread to completely disengage. 786 * 787 * It is possible for us to race another cpu core so we 788 * have to do this correctly. 789 */ 790 for (;;) { 791 mpflags = td->td_mpflags; 792 cpu_ccfence(); 793 if (mpflags & TDF_MP_EXITSIG) 794 break; 795 tsleep_interlock(td, 0); 796 if (atomic_cmpset_int(&td->td_mpflags, mpflags, 797 mpflags | TDF_MP_EXITWAIT)) { 798 tsleep(td, PINTERLOCKED, "lwpxt", 0); 799 } 800 } 801 802 /* 803 * We've already waited for the core exit but there can still 804 * be other refs from e.g. process scans and such. 805 */ 806 if (lp->lwp_lock > 0) { 807 tsleep(lp, 0, "lwpwait1", 1); 808 return(0); 809 } 810 if (td->td_refs) { 811 tsleep(td, 0, "lwpwait2", 1); 812 return(0); 813 } 814 815 /* 816 * Now that we have the thread destruction interlock these flags 817 * really should already be cleaned up, keep a check for safety. 818 * 819 * We can't rip its stack out from under it until TDF_EXITING is 820 * set and both TDF_RUNNING and TDF_PREEMPT_LOCK are clear. 821 * TDF_PREEMPT_LOCK must be checked because TDF_RUNNING 822 * will be cleared temporarily if a thread gets preempted. 823 */ 824 while ((td->td_flags & (TDF_RUNNING | 825 TDF_RUNQ | 826 TDF_PREEMPT_LOCK | 827 TDF_EXITING)) != TDF_EXITING) { 828 tsleep(lp, 0, "lwpwait3", 1); 829 return (0); 830 } 831 832 KASSERT((td->td_flags & (TDF_RUNQ|TDF_TSLEEPQ)) == 0, 833 ("lwp_wait: td %p (%s) still on run or sleep queue", 834 td, td->td_comm)); 835 return (1); 836 } 837 838 /* 839 * Release the resources associated with a lwp. 840 * The lwp must be completely dead. 841 */ 842 void 843 lwp_dispose(struct lwp *lp) 844 { 845 struct thread *td = lp->lwp_thread; 846 847 KKASSERT(lwkt_preempted_proc() != lp); 848 KKASSERT(lp->lwp_lock == 0); 849 KKASSERT(td->td_refs == 0); 850 KKASSERT((td->td_flags & (TDF_RUNNING | 851 TDF_RUNQ | 852 TDF_PREEMPT_LOCK | 853 TDF_EXITING)) == TDF_EXITING); 854 855 PRELE(lp->lwp_proc); 856 lp->lwp_proc = NULL; 857 if (td != NULL) { 858 td->td_proc = NULL; 859 td->td_lwp = NULL; 860 lp->lwp_thread = NULL; 861 lwkt_free_thread(td); 862 } 863 kfree(lp, M_LWP); 864 } 865 866 int 867 sys_wait4(struct wait_args *uap) 868 { 869 struct rusage rusage; 870 int error, status; 871 872 error = kern_wait(uap->pid, (uap->status ? &status : NULL), 873 uap->options, (uap->rusage ? &rusage : NULL), 874 &uap->sysmsg_result); 875 876 if (error == 0 && uap->status) 877 error = copyout(&status, uap->status, sizeof(*uap->status)); 878 if (error == 0 && uap->rusage) 879 error = copyout(&rusage, uap->rusage, sizeof(*uap->rusage)); 880 return (error); 881 } 882 883 /* 884 * wait1() 885 * 886 * wait_args(int pid, int *status, int options, struct rusage *rusage) 887 */ 888 int 889 kern_wait(pid_t pid, int *status, int options, struct rusage *rusage, int *res) 890 { 891 struct thread *td = curthread; 892 struct lwp *lp; 893 struct proc *q = td->td_proc; 894 struct proc *p, *t; 895 struct pargs *pa; 896 struct sigacts *ps; 897 int nfound, error; 898 long waitgen; 899 900 if (pid == 0) 901 pid = -q->p_pgid; 902 if (options &~ (WUNTRACED|WNOHANG|WCONTINUED|WLINUXCLONE)) 903 return (EINVAL); 904 905 /* 906 * Protect the q->p_children list 907 */ 908 lwkt_gettoken(&q->p_token); 909 loop: 910 /* 911 * All sorts of things can change due to blocking so we have to loop 912 * all the way back up here. 913 * 914 * The problem is that if a process group is stopped and the parent 915 * is doing a wait*(..., WUNTRACED, ...), it will see the STOP 916 * of the child and then stop itself when it tries to return from the 917 * system call. When the process group is resumed the parent will 918 * then get the STOP status even though the child has now resumed 919 * (a followup wait*() will get the CONT status). 920 * 921 * Previously the CONT would overwrite the STOP because the tstop 922 * was handled within tsleep(), and the parent would only see 923 * the CONT when both are stopped and continued together. This little 924 * two-line hack restores this effect. 925 */ 926 if (STOPLWP(q, td->td_lwp)) 927 tstop(); 928 929 nfound = 0; 930 931 /* 932 * Loop on children. 933 * 934 * NOTE: We don't want to break q's p_token in the loop for the 935 * case where no children are found or we risk breaking the 936 * interlock between child and parent. 937 */ 938 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 939 LIST_FOREACH(p, &q->p_children, p_sibling) { 940 if (pid != WAIT_ANY && 941 p->p_pid != pid && p->p_pgid != -pid) { 942 continue; 943 } 944 945 /* 946 * This special case handles a kthread spawned by linux_clone 947 * (see linux_misc.c). The linux_wait4 and linux_waitpid 948 * functions need to be able to distinguish between waiting 949 * on a process and waiting on a thread. It is a thread if 950 * p_sigparent is not SIGCHLD, and the WLINUXCLONE option 951 * signifies we want to wait for threads and not processes. 952 */ 953 if ((p->p_sigparent != SIGCHLD) ^ 954 ((options & WLINUXCLONE) != 0)) { 955 continue; 956 } 957 958 nfound++; 959 if (p->p_stat == SZOMB) { 960 /* 961 * We may go into SZOMB with threads still present. 962 * We must wait for them to exit before we can reap 963 * the master thread, otherwise we may race reaping 964 * non-master threads. 965 * 966 * Only this routine can remove a process from 967 * the zombie list and destroy it, use PACQUIREZOMB() 968 * to serialize us and loop if it blocks (interlocked 969 * by the parent's q->p_token). 970 * 971 * WARNING! (p) can be invalid when PHOLDZOMB(p) 972 * returns non-zero. Be sure not to 973 * mess with it. 974 */ 975 if (PHOLDZOMB(p)) 976 goto loop; 977 lwkt_gettoken(&p->p_token); 978 if (p->p_pptr != q) { 979 lwkt_reltoken(&p->p_token); 980 PRELEZOMB(p); 981 goto loop; 982 } 983 while (p->p_nthreads > 0) { 984 tsleep(&p->p_nthreads, 0, "lwpzomb", hz); 985 } 986 987 /* 988 * Reap any LWPs left in p->p_lwps. This is usually 989 * just the last LWP. This must be done before 990 * we loop on p_lock since the lwps hold a ref on 991 * it as a vmspace interlock. 992 * 993 * Once that is accomplished p_nthreads had better 994 * be zero. 995 */ 996 while ((lp = RB_ROOT(&p->p_lwp_tree)) != NULL) { 997 /* 998 * Make sure no one is using this lwp, before 999 * it is removed from the tree. If we didn't 1000 * wait it here, lwp tree iteration with 1001 * blocking operation would be broken. 1002 */ 1003 while (lp->lwp_lock > 0) 1004 tsleep(lp, 0, "zomblwp", 1); 1005 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 1006 reaplwp(lp); 1007 } 1008 KKASSERT(p->p_nthreads == 0); 1009 1010 /* 1011 * Don't do anything really bad until all references 1012 * to the process go away. This may include other 1013 * LWPs which are still in the process of being 1014 * reaped. We can't just pull the rug out from under 1015 * them because they may still be using the VM space. 1016 * 1017 * Certain kernel facilities such as /proc will also 1018 * put a hold on the process for short periods of 1019 * time. 1020 */ 1021 PRELE(p); 1022 PSTALL(p, "reap3", 0); 1023 1024 /* Take care of our return values. */ 1025 *res = p->p_pid; 1026 1027 if (status) 1028 *status = p->p_xstat; 1029 if (rusage) 1030 *rusage = p->p_ru; 1031 1032 /* 1033 * If we got the child via a ptrace 'attach', 1034 * we need to give it back to the old parent. 1035 */ 1036 if (p->p_oppid && (t = pfind(p->p_oppid)) != NULL) { 1037 PHOLD(p); 1038 p->p_oppid = 0; 1039 proc_reparent(p, t); 1040 ksignal(t, SIGCHLD); 1041 wakeup((caddr_t)t); 1042 error = 0; 1043 PRELE(t); 1044 lwkt_reltoken(&p->p_token); 1045 PRELEZOMB(p); 1046 goto done; 1047 } 1048 1049 /* 1050 * Unlink the proc from its process group so that 1051 * the following operations won't lead to an 1052 * inconsistent state for processes running down 1053 * the zombie list. 1054 */ 1055 proc_remove_zombie(p); 1056 proc_userunmap(p); 1057 lwkt_reltoken(&p->p_token); 1058 leavepgrp(p); 1059 1060 p->p_xstat = 0; 1061 ruadd(&q->p_cru, &p->p_ru); 1062 1063 /* 1064 * Decrement the count of procs running with this uid. 1065 */ 1066 chgproccnt(p->p_ucred->cr_ruidinfo, -1, 0); 1067 1068 /* 1069 * Free up credentials. 1070 */ 1071 crfree(p->p_ucred); 1072 p->p_ucred = NULL; 1073 1074 /* 1075 * Remove unused arguments 1076 */ 1077 pa = p->p_args; 1078 p->p_args = NULL; 1079 if (pa && refcount_release(&pa->ar_ref)) { 1080 kfree(pa, M_PARGS); 1081 pa = NULL; 1082 } 1083 1084 ps = p->p_sigacts; 1085 p->p_sigacts = NULL; 1086 if (ps && refcount_release(&ps->ps_refcnt)) { 1087 kfree(ps, M_SUBPROC); 1088 ps = NULL; 1089 } 1090 1091 /* 1092 * Our exitingcount was incremented when the process 1093 * became a zombie, now that the process has been 1094 * removed from (almost) all lists we should be able 1095 * to safely destroy its vmspace. Wait for any current 1096 * holders to go away (so the vmspace remains stable), 1097 * then scrap it. 1098 * 1099 * NOTE: Releasing the parent process (q) p_token 1100 * across the vmspace_exitfree() call is 1101 * important here to reduce stalls on 1102 * interactions with (q) (such as 1103 * fork/exec/wait or 'ps'). 1104 */ 1105 PSTALL(p, "reap4", 0); 1106 lwkt_reltoken(&q->p_token); 1107 vmspace_exitfree(p); 1108 lwkt_gettoken(&q->p_token); 1109 PSTALL(p, "reap5", 0); 1110 1111 /* 1112 * NOTE: We have to officially release ZOMB in order 1113 * to ensure that a racing thread in kern_wait() 1114 * which blocked on ZOMB is woken up. 1115 */ 1116 PHOLD(p); 1117 PRELEZOMB(p); 1118 kfree(p, M_PROC); 1119 atomic_add_int(&nprocs, -1); 1120 error = 0; 1121 goto done; 1122 } 1123 if ((p->p_stat == SSTOP || p->p_stat == SCORE) && 1124 (p->p_flags & P_WAITED) == 0 && 1125 ((p->p_flags & P_TRACED) || (options & WUNTRACED))) { 1126 PHOLD(p); 1127 lwkt_gettoken(&p->p_token); 1128 if (p->p_pptr != q) { 1129 lwkt_reltoken(&p->p_token); 1130 PRELE(p); 1131 goto loop; 1132 } 1133 if ((p->p_stat != SSTOP && p->p_stat != SCORE) || 1134 (p->p_flags & P_WAITED) != 0 || 1135 ((p->p_flags & P_TRACED) == 0 && 1136 (options & WUNTRACED) == 0)) { 1137 lwkt_reltoken(&p->p_token); 1138 PRELE(p); 1139 goto loop; 1140 } 1141 1142 p->p_flags |= P_WAITED; 1143 1144 *res = p->p_pid; 1145 if (status) 1146 *status = W_STOPCODE(p->p_xstat); 1147 /* Zero rusage so we get something consistent. */ 1148 if (rusage) 1149 bzero(rusage, sizeof(*rusage)); 1150 error = 0; 1151 lwkt_reltoken(&p->p_token); 1152 PRELE(p); 1153 goto done; 1154 } 1155 if ((options & WCONTINUED) && (p->p_flags & P_CONTINUED)) { 1156 PHOLD(p); 1157 lwkt_gettoken(&p->p_token); 1158 if (p->p_pptr != q) { 1159 lwkt_reltoken(&p->p_token); 1160 PRELE(p); 1161 goto loop; 1162 } 1163 if ((p->p_flags & P_CONTINUED) == 0) { 1164 lwkt_reltoken(&p->p_token); 1165 PRELE(p); 1166 goto loop; 1167 } 1168 1169 *res = p->p_pid; 1170 p->p_flags &= ~P_CONTINUED; 1171 1172 if (status) 1173 *status = SIGCONT; 1174 error = 0; 1175 lwkt_reltoken(&p->p_token); 1176 PRELE(p); 1177 goto done; 1178 } 1179 } 1180 if (nfound == 0) { 1181 error = ECHILD; 1182 goto done; 1183 } 1184 if (options & WNOHANG) { 1185 *res = 0; 1186 error = 0; 1187 goto done; 1188 } 1189 1190 /* 1191 * Wait for signal - interlocked using q->p_waitgen. 1192 */ 1193 error = 0; 1194 while ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1195 tsleep_interlock(q, PCATCH); 1196 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 1197 if ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1198 error = tsleep(q, PCATCH | PINTERLOCKED, "wait", 0); 1199 break; 1200 } 1201 } 1202 if (error) { 1203 done: 1204 lwkt_reltoken(&q->p_token); 1205 return (error); 1206 } 1207 goto loop; 1208 } 1209 1210 /* 1211 * Change child's parent process to parent. 1212 * 1213 * p_children/p_sibling requires the parent's token, and 1214 * changing pptr requires the child's token, so we have to 1215 * get three tokens to do this operation. We also need to 1216 * hold pointers that might get ripped out from under us to 1217 * preserve structural integrity. 1218 * 1219 * It is possible to race another reparent or disconnect or other 1220 * similar operation. We must retry when this situation occurs. 1221 * Once we successfully reparent the process we no longer care 1222 * about any races. 1223 */ 1224 void 1225 proc_reparent(struct proc *child, struct proc *parent) 1226 { 1227 struct proc *opp; 1228 1229 PHOLD(parent); 1230 while ((opp = child->p_pptr) != parent) { 1231 PHOLD(opp); 1232 lwkt_gettoken(&opp->p_token); 1233 lwkt_gettoken(&child->p_token); 1234 lwkt_gettoken(&parent->p_token); 1235 if (child->p_pptr != opp) { 1236 lwkt_reltoken(&parent->p_token); 1237 lwkt_reltoken(&child->p_token); 1238 lwkt_reltoken(&opp->p_token); 1239 PRELE(opp); 1240 continue; 1241 } 1242 LIST_REMOVE(child, p_sibling); 1243 LIST_INSERT_HEAD(&parent->p_children, child, p_sibling); 1244 child->p_pptr = parent; 1245 lwkt_reltoken(&parent->p_token); 1246 lwkt_reltoken(&child->p_token); 1247 lwkt_reltoken(&opp->p_token); 1248 if (LIST_EMPTY(&opp->p_children)) 1249 wakeup(opp); 1250 PRELE(opp); 1251 break; 1252 } 1253 PRELE(parent); 1254 } 1255 1256 /* 1257 * The next two functions are to handle adding/deleting items on the 1258 * exit callout list 1259 * 1260 * at_exit(): 1261 * Take the arguments given and put them onto the exit callout list, 1262 * However first make sure that it's not already there. 1263 * returns 0 on success. 1264 */ 1265 1266 int 1267 at_exit(exitlist_fn function) 1268 { 1269 struct exitlist *ep; 1270 1271 #ifdef INVARIANTS 1272 /* Be noisy if the programmer has lost track of things */ 1273 if (rm_at_exit(function)) 1274 kprintf("WARNING: exit callout entry (%p) already present\n", 1275 function); 1276 #endif 1277 ep = kmalloc(sizeof(*ep), M_ATEXIT, M_NOWAIT); 1278 if (ep == NULL) 1279 return (ENOMEM); 1280 ep->function = function; 1281 TAILQ_INSERT_TAIL(&exit_list, ep, next); 1282 return (0); 1283 } 1284 1285 /* 1286 * Scan the exit callout list for the given item and remove it. 1287 * Returns the number of items removed (0 or 1) 1288 */ 1289 int 1290 rm_at_exit(exitlist_fn function) 1291 { 1292 struct exitlist *ep; 1293 1294 TAILQ_FOREACH(ep, &exit_list, next) { 1295 if (ep->function == function) { 1296 TAILQ_REMOVE(&exit_list, ep, next); 1297 kfree(ep, M_ATEXIT); 1298 return(1); 1299 } 1300 } 1301 return (0); 1302 } 1303 1304 /* 1305 * LWP reaper related code. 1306 */ 1307 static void 1308 reaplwps(void *context, int dummy) 1309 { 1310 struct lwplist *lwplist = context; 1311 struct lwp *lp; 1312 int cpu = mycpuid; 1313 1314 lwkt_gettoken(&deadlwp_token[cpu]); 1315 while ((lp = LIST_FIRST(lwplist))) { 1316 LIST_REMOVE(lp, u.lwp_reap_entry); 1317 reaplwp(lp); 1318 } 1319 lwkt_reltoken(&deadlwp_token[cpu]); 1320 } 1321 1322 static void 1323 reaplwp(struct lwp *lp) 1324 { 1325 while (lwp_wait(lp) == 0) 1326 ; 1327 lwp_dispose(lp); 1328 } 1329 1330 static void 1331 deadlwp_init(void) 1332 { 1333 int cpu; 1334 1335 for (cpu = 0; cpu < ncpus; cpu++) { 1336 lwkt_token_init(&deadlwp_token[cpu], "deadlwpl"); 1337 LIST_INIT(&deadlwp_list[cpu]); 1338 deadlwp_task[cpu] = kmalloc(sizeof(*deadlwp_task[cpu]), 1339 M_DEVBUF, M_WAITOK); 1340 TASK_INIT(deadlwp_task[cpu], 0, reaplwps, &deadlwp_list[cpu]); 1341 } 1342 } 1343 1344 SYSINIT(deadlwpinit, SI_SUB_CONFIGURE, SI_ORDER_ANY, deadlwp_init, NULL); 1345