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_compat.h" 39 #include "opt_ktrace.h" 40 41 #include <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/sysproto.h> 44 #include <sys/kernel.h> 45 #include <sys/malloc.h> 46 #include <sys/proc.h> 47 #include <sys/ktrace.h> 48 #include <sys/pioctl.h> 49 #include <sys/tty.h> 50 #include <sys/wait.h> 51 #include <sys/vnode.h> 52 #include <sys/resourcevar.h> 53 #include <sys/signalvar.h> 54 #include <sys/taskqueue.h> 55 #include <sys/ptrace.h> 56 #include <sys/acct.h> /* for acct_process() function prototype */ 57 #include <sys/filedesc.h> 58 #include <sys/shm.h> 59 #include <sys/sem.h> 60 #include <sys/jail.h> 61 #include <sys/kern_syscall.h> 62 #include <sys/unistd.h> 63 #include <sys/eventhandler.h> 64 #include <sys/dsched.h> 65 66 #include <vm/vm.h> 67 #include <vm/vm_param.h> 68 #include <sys/lock.h> 69 #include <vm/pmap.h> 70 #include <vm/vm_map.h> 71 #include <vm/vm_extern.h> 72 #include <sys/user.h> 73 74 #include <sys/refcount.h> 75 #include <sys/thread2.h> 76 #include <sys/sysref2.h> 77 #include <sys/mplock2.h> 78 79 #include <machine/vmm.h> 80 81 static void reaplwps(void *context, int dummy); 82 static void reaplwp(struct lwp *lp); 83 static void killlwps(struct lwp *lp); 84 85 static MALLOC_DEFINE(M_ATEXIT, "atexit", "atexit callback"); 86 87 /* 88 * callout list for things to do at exit time 89 */ 90 struct exitlist { 91 exitlist_fn function; 92 TAILQ_ENTRY(exitlist) next; 93 }; 94 95 TAILQ_HEAD(exit_list_head, exitlist); 96 static struct exit_list_head exit_list = TAILQ_HEAD_INITIALIZER(exit_list); 97 98 /* 99 * LWP reaper data 100 */ 101 static struct task *deadlwp_task[MAXCPU]; 102 static struct lwplist deadlwp_list[MAXCPU]; 103 static struct lwkt_token deadlwp_token[MAXCPU]; 104 105 /* 106 * exit -- 107 * Death of process. 108 * 109 * SYS_EXIT_ARGS(int rval) 110 */ 111 int 112 sys_exit(struct exit_args *uap) 113 { 114 exit1(W_EXITCODE(uap->rval, 0)); 115 /* NOTREACHED */ 116 } 117 118 /* 119 * Extended exit -- 120 * Death of a lwp or process with optional bells and whistles. 121 */ 122 int 123 sys_extexit(struct extexit_args *uap) 124 { 125 struct proc *p = curproc; 126 int action, who; 127 int error; 128 129 action = EXTEXIT_ACTION(uap->how); 130 who = EXTEXIT_WHO(uap->how); 131 132 /* Check parameters before we might perform some action */ 133 switch (who) { 134 case EXTEXIT_PROC: 135 case EXTEXIT_LWP: 136 break; 137 default: 138 return (EINVAL); 139 } 140 141 switch (action) { 142 case EXTEXIT_SIMPLE: 143 break; 144 case EXTEXIT_SETINT: 145 error = copyout(&uap->status, uap->addr, sizeof(uap->status)); 146 if (error) 147 return (error); 148 break; 149 default: 150 return (EINVAL); 151 } 152 153 lwkt_gettoken(&p->p_token); 154 155 switch (who) { 156 case EXTEXIT_LWP: 157 /* 158 * Be sure only to perform a simple lwp exit if there is at 159 * least one more lwp in the proc, which will call exit1() 160 * later, otherwise the proc will be an UNDEAD and not even a 161 * SZOMB! 162 */ 163 if (p->p_nthreads > 1) { 164 lwp_exit(0, NULL); /* called w/ p_token held */ 165 /* NOT REACHED */ 166 } 167 /* else last lwp in proc: do the real thing */ 168 /* FALLTHROUGH */ 169 default: /* to help gcc */ 170 case EXTEXIT_PROC: 171 lwkt_reltoken(&p->p_token); 172 exit1(W_EXITCODE(uap->status, 0)); 173 /* NOTREACHED */ 174 } 175 176 /* NOTREACHED */ 177 lwkt_reltoken(&p->p_token); /* safety */ 178 } 179 180 /* 181 * Kill all lwps associated with the current process except the 182 * current lwp. Return an error if we race another thread trying to 183 * do the same thing and lose the race. 184 * 185 * If forexec is non-zero the current thread and process flags are 186 * cleaned up so they can be reused. 187 * 188 * Caller must hold curproc->p_token 189 */ 190 int 191 killalllwps(int forexec) 192 { 193 struct lwp *lp = curthread->td_lwp; 194 struct proc *p = lp->lwp_proc; 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 * Interlock with LWP_MP_WEXIT and kill any remaining LWPs 206 */ 207 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 208 if (p->p_nthreads > 1) 209 killlwps(lp); 210 211 /* 212 * If doing this for an exec, clean up the remaining thread 213 * (us) for continuing operation after all the other threads 214 * have been killed. 215 */ 216 if (forexec) { 217 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 218 p->p_flags &= ~P_WEXIT; 219 } 220 return(0); 221 } 222 223 /* 224 * Kill all LWPs except the current one. Do not try to signal 225 * LWPs which have exited on their own or have already been 226 * signaled. 227 */ 228 static void 229 killlwps(struct lwp *lp) 230 { 231 struct proc *p = lp->lwp_proc; 232 struct lwp *tlp; 233 234 /* 235 * Kill the remaining LWPs. We must send the signal before setting 236 * LWP_MP_WEXIT. The setting of WEXIT is optional but helps reduce 237 * races. tlp must be held across the call as it might block and 238 * allow the target lwp to rip itself out from under our loop. 239 */ 240 FOREACH_LWP_IN_PROC(tlp, p) { 241 LWPHOLD(tlp); 242 lwkt_gettoken(&tlp->lwp_token); 243 if ((tlp->lwp_mpflags & LWP_MP_WEXIT) == 0) { 244 lwpsignal(p, tlp, SIGKILL); 245 atomic_set_int(&tlp->lwp_mpflags, LWP_MP_WEXIT); 246 } 247 lwkt_reltoken(&tlp->lwp_token); 248 LWPRELE(tlp); 249 } 250 251 /* 252 * Wait for everything to clear out. 253 */ 254 while (p->p_nthreads > 1) 255 tsleep(&p->p_nthreads, 0, "killlwps", 0); 256 } 257 258 /* 259 * Exit: deallocate address space and other resources, change proc state 260 * to zombie, and unlink proc from allproc and parent's lists. Save exit 261 * status and rusage for wait(). Check for child processes and orphan them. 262 */ 263 void 264 exit1(int rv) 265 { 266 struct thread *td = curthread; 267 struct proc *p = td->td_proc; 268 struct lwp *lp = td->td_lwp; 269 struct proc *q; 270 struct proc *pp; 271 struct proc *reproc; 272 struct sysreaper *reap; 273 struct vmspace *vm; 274 struct vnode *vtmp; 275 struct exitlist *ep; 276 int error; 277 278 lwkt_gettoken(&p->p_token); 279 280 if (p->p_pid == 1) { 281 kprintf("init died (signal %d, exit %d)\n", 282 WTERMSIG(rv), WEXITSTATUS(rv)); 283 panic("Going nowhere without my init!"); 284 } 285 varsymset_clean(&p->p_varsymset); 286 lockuninit(&p->p_varsymset.vx_lock); 287 288 /* 289 * Kill all lwps associated with the current process, return an 290 * error if we race another thread trying to do the same thing 291 * and lose the race. 292 */ 293 error = killalllwps(0); 294 if (error) { 295 lwp_exit(0, NULL); 296 /* NOT REACHED */ 297 } 298 299 /* are we a task leader? */ 300 if (p == p->p_leader) { 301 struct kill_args killArgs; 302 killArgs.signum = SIGKILL; 303 q = p->p_peers; 304 while(q) { 305 killArgs.pid = q->p_pid; 306 /* 307 * The interface for kill is better 308 * than the internal signal 309 */ 310 sys_kill(&killArgs); 311 q = q->p_peers; 312 } 313 while (p->p_peers) 314 tsleep((caddr_t)p, 0, "exit1", 0); 315 } 316 317 #ifdef PGINPROF 318 vmsizmon(); 319 #endif 320 STOPEVENT(p, S_EXIT, rv); 321 p->p_flags |= P_POSTEXIT; /* stop procfs stepping */ 322 323 /* 324 * Check if any loadable modules need anything done at process exit. 325 * e.g. SYSV IPC stuff 326 * XXX what if one of these generates an error? 327 */ 328 p->p_xstat = rv; 329 EVENTHANDLER_INVOKE(process_exit, p); 330 331 /* 332 * XXX: imho, the eventhandler stuff is much cleaner than this. 333 * Maybe we should move everything to use eventhandler. 334 */ 335 TAILQ_FOREACH(ep, &exit_list, next) 336 (*ep->function)(td); 337 338 if (p->p_flags & P_PROFIL) 339 stopprofclock(p); 340 341 SIGEMPTYSET(p->p_siglist); 342 SIGEMPTYSET(lp->lwp_siglist); 343 if (timevalisset(&p->p_realtimer.it_value)) 344 callout_stop_sync(&p->p_ithandle); 345 346 /* 347 * Reset any sigio structures pointing to us as a result of 348 * F_SETOWN with our pid. 349 */ 350 funsetownlst(&p->p_sigiolst); 351 352 /* 353 * Close open files and release open-file table. 354 * This may block! 355 */ 356 fdfree(p, NULL); 357 358 if (p->p_leader->p_peers) { 359 q = p->p_leader; 360 while(q->p_peers != p) 361 q = q->p_peers; 362 q->p_peers = p->p_peers; 363 wakeup((caddr_t)p->p_leader); 364 } 365 366 /* 367 * XXX Shutdown SYSV semaphores 368 */ 369 semexit(p); 370 371 KKASSERT(p->p_numposixlocks == 0); 372 373 /* The next two chunks should probably be moved to vmspace_exit. */ 374 vm = p->p_vmspace; 375 376 /* 377 * Clean up data related to virtual kernel operation. Clean up 378 * any vkernel context related to the current lwp now so we can 379 * destroy p_vkernel. 380 */ 381 if (p->p_vkernel) { 382 vkernel_lwp_exit(lp); 383 vkernel_exit(p); 384 } 385 386 /* 387 * Release the user portion of address space. The exitbump prevents 388 * the vmspace from being completely eradicated (using holdcnt). 389 * This releases references to vnodes, which could cause I/O if the 390 * file has been unlinked. We need to do this early enough that 391 * we can still sleep. 392 * 393 * We can't free the entire vmspace as the kernel stack may be mapped 394 * within that space also. 395 * 396 * Processes sharing the same vmspace may exit in one order, and 397 * get cleaned up by vmspace_exit() in a different order. The 398 * last exiting process to reach this point releases as much of 399 * the environment as it can, and the last process cleaned up 400 * by vmspace_exit() (which decrements exitingcnt) cleans up the 401 * remainder. 402 */ 403 vmspace_relexit(vm); 404 405 if (SESS_LEADER(p)) { 406 struct session *sp = p->p_session; 407 408 if (sp->s_ttyvp) { 409 /* 410 * We are the controlling process. Signal the 411 * foreground process group, drain the controlling 412 * terminal, and revoke access to the controlling 413 * terminal. 414 * 415 * NOTE: while waiting for the process group to exit 416 * it is possible that one of the processes in the 417 * group will revoke the tty, so the ttyclosesession() 418 * function will re-check sp->s_ttyvp. 419 */ 420 if (sp->s_ttyp && (sp->s_ttyp->t_session == sp)) { 421 if (sp->s_ttyp->t_pgrp) 422 pgsignal(sp->s_ttyp->t_pgrp, SIGHUP, 1); 423 ttywait(sp->s_ttyp); 424 ttyclosesession(sp, 1); /* also revoke */ 425 } 426 /* 427 * Release the tty. If someone has it open via 428 * /dev/tty then close it (since they no longer can 429 * once we've NULL'd it out). 430 */ 431 ttyclosesession(sp, 0); 432 433 /* 434 * s_ttyp is not zero'd; we use this to indicate 435 * that the session once had a controlling terminal. 436 * (for logging and informational purposes) 437 */ 438 } 439 sp->s_leader = NULL; 440 } 441 fixjobc(p, p->p_pgrp, 0); 442 (void)acct_process(p); 443 #ifdef KTRACE 444 /* 445 * release trace file 446 */ 447 if (p->p_tracenode) 448 ktrdestroy(&p->p_tracenode); 449 p->p_traceflag = 0; 450 #endif 451 /* 452 * Release reference to text vnode 453 */ 454 if ((vtmp = p->p_textvp) != NULL) { 455 p->p_textvp = NULL; 456 vrele(vtmp); 457 } 458 459 /* Release namecache handle to text file */ 460 if (p->p_textnch.ncp) 461 cache_drop(&p->p_textnch); 462 463 /* 464 * We have to handle PPWAIT here or proc_move_allproc_zombie() 465 * will block on the PHOLD() the parent is doing. 466 * 467 * We are using the flag as an interlock so an atomic op is 468 * necessary to synchronize with the parent's cpu. 469 */ 470 if (p->p_flags & P_PPWAIT) { 471 if (p->p_pptr && p->p_pptr->p_upmap) 472 p->p_pptr->p_upmap->invfork = 0; 473 atomic_clear_int(&p->p_flags, P_PPWAIT); 474 wakeup(p->p_pptr); 475 } 476 477 /* 478 * Move the process to the zombie list. This will block 479 * until the process p_lock count reaches 0. The process will 480 * not be reaped until TDF_EXITING is set by cpu_thread_exit(), 481 * which is called from cpu_proc_exit(). 482 * 483 * Interlock against waiters using p_waitgen. We increment 484 * p_waitgen after completing the move of our process to the 485 * zombie list. 486 * 487 * WARNING: pp becomes stale when we block, clear it now as a 488 * reminder. 489 */ 490 proc_move_allproc_zombie(p); 491 pp = p->p_pptr; 492 atomic_add_long(&pp->p_waitgen, 1); 493 pp = NULL; 494 495 /* 496 * release controlled reaper for exit if we own it and return the 497 * remaining reaper (the one for us), which we will drop after we 498 * are done. 499 */ 500 reap = reaper_exit(p); 501 502 /* 503 * Reparent all of this process's children to the init process or 504 * to the designated reaper. We must hold the reaper's p_token in 505 * order to safely mess with p_children. 506 * 507 * We already hold p->p_token (to remove the children from our list). 508 */ 509 reproc = NULL; 510 q = LIST_FIRST(&p->p_children); 511 if (q) { 512 reproc = reaper_get(reap); 513 lwkt_gettoken(&reproc->p_token); 514 while ((q = LIST_FIRST(&p->p_children)) != NULL) { 515 PHOLD(q); 516 lwkt_gettoken(&q->p_token); 517 if (q != LIST_FIRST(&p->p_children)) { 518 lwkt_reltoken(&q->p_token); 519 PRELE(q); 520 continue; 521 } 522 LIST_REMOVE(q, p_sibling); 523 LIST_INSERT_HEAD(&reproc->p_children, q, p_sibling); 524 q->p_pptr = reproc; 525 q->p_sigparent = SIGCHLD; 526 527 /* 528 * Traced processes are killed 529 * since their existence means someone is screwing up. 530 */ 531 if (q->p_flags & P_TRACED) { 532 q->p_flags &= ~P_TRACED; 533 ksignal(q, SIGKILL); 534 } 535 lwkt_reltoken(&q->p_token); 536 PRELE(q); 537 } 538 lwkt_reltoken(&reproc->p_token); 539 wakeup(reproc); 540 } 541 542 /* 543 * Save exit status and final rusage info, adding in child rusage 544 * info and self times. 545 */ 546 calcru_proc(p, &p->p_ru); 547 ruadd(&p->p_ru, &p->p_cru); 548 549 /* 550 * notify interested parties of our demise. 551 */ 552 KNOTE(&p->p_klist, NOTE_EXIT); 553 554 /* 555 * Notify parent that we're gone. If parent has the PS_NOCLDWAIT 556 * flag set, or if the handler is set to SIG_IGN, notify the reaper 557 * instead (it will handle this situation). 558 * 559 * NOTE: The reaper can still be the parent process. 560 * 561 * (must reload pp) 562 */ 563 if (p->p_pptr->p_sigacts->ps_flag & (PS_NOCLDWAIT | PS_CLDSIGIGN)) { 564 if (reproc == NULL) 565 reproc = reaper_get(reap); 566 proc_reparent(p, reproc); 567 } 568 if (reproc) 569 PRELE(reproc); 570 if (reap) 571 reaper_drop(reap); 572 573 /* 574 * Signal (possibly new) parent. 575 */ 576 pp = p->p_pptr; 577 PHOLD(pp); 578 if (p->p_sigparent && pp != initproc) { 579 int sig = p->p_sigparent; 580 581 if (sig != SIGUSR1 && sig != SIGCHLD) 582 sig = SIGCHLD; 583 ksignal(pp, sig); 584 } else { 585 ksignal(pp, SIGCHLD); 586 } 587 p->p_flags &= ~P_TRACED; 588 PRELE(pp); 589 590 /* 591 * cpu_exit is responsible for clearing curproc, since 592 * it is heavily integrated with the thread/switching sequence. 593 * 594 * Other substructures are freed from wait(). 595 */ 596 plimit_free(p); 597 598 /* 599 * Finally, call machine-dependent code to release as many of the 600 * lwp's resources as we can and halt execution of this thread. 601 * 602 * pp is a wild pointer now but still the correct wakeup() target. 603 * lwp_exit() only uses it to send the wakeup() signal to the likely 604 * parent. Any reparenting race that occurs will get a signal 605 * automatically and not be an issue. 606 */ 607 lwp_exit(1, pp); 608 } 609 610 /* 611 * Eventually called by every exiting LWP 612 * 613 * p->p_token must be held. mplock may be held and will be released. 614 */ 615 void 616 lwp_exit(int masterexit, void *waddr) 617 { 618 struct thread *td = curthread; 619 struct lwp *lp = td->td_lwp; 620 struct proc *p = lp->lwp_proc; 621 int dowake = 0; 622 623 /* 624 * Release the current user process designation on the process so 625 * the userland scheduler can work in someone else. 626 */ 627 p->p_usched->release_curproc(lp); 628 629 /* 630 * lwp_exit() may be called without setting LWP_MP_WEXIT, so 631 * make sure it is set here. 632 */ 633 ASSERT_LWKT_TOKEN_HELD(&p->p_token); 634 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 635 636 /* 637 * Clean up any virtualization 638 */ 639 if (lp->lwp_vkernel) 640 vkernel_lwp_exit(lp); 641 642 if (td->td_vmm) 643 vmm_vmdestroy(); 644 645 /* 646 * Clean up select/poll support 647 */ 648 kqueue_terminate(&lp->lwp_kqueue); 649 650 /* 651 * Clean up any syscall-cached ucred 652 */ 653 if (td->td_ucred) { 654 crfree(td->td_ucred); 655 td->td_ucred = NULL; 656 } 657 658 /* 659 * Nobody actually wakes us when the lock 660 * count reaches zero, so just wait one tick. 661 */ 662 while (lp->lwp_lock > 0) 663 tsleep(lp, 0, "lwpexit", 1); 664 665 /* Hand down resource usage to our proc */ 666 ruadd(&p->p_ru, &lp->lwp_ru); 667 668 /* 669 * If we don't hold the process until the LWP is reaped wait*() 670 * may try to dispose of its vmspace before all the LWPs have 671 * actually terminated. 672 */ 673 PHOLD(p); 674 675 /* 676 * Do any remaining work that might block on us. We should be 677 * coded such that further blocking is ok after decrementing 678 * p_nthreads but don't take the chance. 679 */ 680 dsched_exit_thread(td); 681 biosched_done(curthread); 682 683 /* 684 * We have to use the reaper for all the LWPs except the one doing 685 * the master exit. The LWP doing the master exit can just be 686 * left on p_lwps and the process reaper will deal with it 687 * synchronously, which is much faster. 688 * 689 * Wakeup anyone waiting on p_nthreads to drop to 1 or 0. 690 * 691 * The process is left held until the reaper calls lwp_dispose() on 692 * the lp (after calling lwp_wait()). 693 */ 694 if (masterexit == 0) { 695 int cpu = mycpuid; 696 697 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 698 --p->p_nthreads; 699 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 700 dowake = 1; 701 lwkt_gettoken(&deadlwp_token[cpu]); 702 LIST_INSERT_HEAD(&deadlwp_list[cpu], lp, u.lwp_reap_entry); 703 taskqueue_enqueue(taskqueue_thread[cpu], deadlwp_task[cpu]); 704 lwkt_reltoken(&deadlwp_token[cpu]); 705 } else { 706 --p->p_nthreads; 707 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 708 dowake = 1; 709 } 710 711 /* 712 * We no longer need p_token. 713 * 714 * Tell the userland scheduler that we are going away 715 */ 716 lwkt_reltoken(&p->p_token); 717 p->p_usched->heuristic_exiting(lp, p); 718 719 /* 720 * Issue late wakeups after releasing our token to give us a chance 721 * to deschedule and switch away before another cpu in a wait*() 722 * reaps us. This is done as late as possible to reduce contention. 723 */ 724 if (dowake) 725 wakeup(&p->p_nthreads); 726 if (waddr) 727 wakeup(waddr); 728 729 cpu_lwp_exit(); 730 } 731 732 /* 733 * Wait until a lwp is completely dead. The final interlock in this drama 734 * is when TDF_EXITING is set in cpu_thread_exit() just before the final 735 * switchout. 736 * 737 * At the point TDF_EXITING is set a complete exit is accomplished when 738 * TDF_RUNNING and TDF_PREEMPT_LOCK are both clear. td_mpflags has two 739 * post-switch interlock flags that can be used to wait for the TDF_ 740 * flags to clear. 741 * 742 * Returns non-zero on success, and zero if the caller needs to retry 743 * the lwp_wait(). 744 */ 745 static int 746 lwp_wait(struct lwp *lp) 747 { 748 struct thread *td = lp->lwp_thread; 749 u_int mpflags; 750 751 KKASSERT(lwkt_preempted_proc() != lp); 752 753 /* 754 * This bit of code uses the thread destruction interlock 755 * managed by lwkt_switch_return() to wait for the lwp's 756 * thread to completely disengage. 757 * 758 * It is possible for us to race another cpu core so we 759 * have to do this correctly. 760 */ 761 for (;;) { 762 mpflags = td->td_mpflags; 763 cpu_ccfence(); 764 if (mpflags & TDF_MP_EXITSIG) 765 break; 766 tsleep_interlock(td, 0); 767 if (atomic_cmpset_int(&td->td_mpflags, mpflags, 768 mpflags | TDF_MP_EXITWAIT)) { 769 tsleep(td, PINTERLOCKED, "lwpxt", 0); 770 } 771 } 772 773 /* 774 * We've already waited for the core exit but there can still 775 * be other refs from e.g. process scans and such. 776 */ 777 if (lp->lwp_lock > 0) { 778 tsleep(lp, 0, "lwpwait1", 1); 779 return(0); 780 } 781 if (td->td_refs) { 782 tsleep(td, 0, "lwpwait2", 1); 783 return(0); 784 } 785 786 /* 787 * Now that we have the thread destruction interlock these flags 788 * really should already be cleaned up, keep a check for safety. 789 * 790 * We can't rip its stack out from under it until TDF_EXITING is 791 * set and both TDF_RUNNING and TDF_PREEMPT_LOCK are clear. 792 * TDF_PREEMPT_LOCK must be checked because TDF_RUNNING 793 * will be cleared temporarily if a thread gets preempted. 794 */ 795 while ((td->td_flags & (TDF_RUNNING | 796 TDF_RUNQ | 797 TDF_PREEMPT_LOCK | 798 TDF_EXITING)) != TDF_EXITING) { 799 tsleep(lp, 0, "lwpwait3", 1); 800 return (0); 801 } 802 803 KASSERT((td->td_flags & (TDF_RUNQ|TDF_TSLEEPQ)) == 0, 804 ("lwp_wait: td %p (%s) still on run or sleep queue", 805 td, td->td_comm)); 806 return (1); 807 } 808 809 /* 810 * Release the resources associated with a lwp. 811 * The lwp must be completely dead. 812 */ 813 void 814 lwp_dispose(struct lwp *lp) 815 { 816 struct thread *td = lp->lwp_thread; 817 818 KKASSERT(lwkt_preempted_proc() != lp); 819 KKASSERT(td->td_refs == 0); 820 KKASSERT((td->td_flags & (TDF_RUNNING | 821 TDF_RUNQ | 822 TDF_PREEMPT_LOCK | 823 TDF_EXITING)) == TDF_EXITING); 824 825 PRELE(lp->lwp_proc); 826 lp->lwp_proc = NULL; 827 if (td != NULL) { 828 td->td_proc = NULL; 829 td->td_lwp = NULL; 830 lp->lwp_thread = NULL; 831 lwkt_free_thread(td); 832 } 833 kfree(lp, M_LWP); 834 } 835 836 int 837 sys_wait4(struct wait_args *uap) 838 { 839 struct rusage rusage; 840 int error, status; 841 842 error = kern_wait(uap->pid, (uap->status ? &status : NULL), 843 uap->options, (uap->rusage ? &rusage : NULL), 844 &uap->sysmsg_result); 845 846 if (error == 0 && uap->status) 847 error = copyout(&status, uap->status, sizeof(*uap->status)); 848 if (error == 0 && uap->rusage) 849 error = copyout(&rusage, uap->rusage, sizeof(*uap->rusage)); 850 return (error); 851 } 852 853 /* 854 * wait1() 855 * 856 * wait_args(int pid, int *status, int options, struct rusage *rusage) 857 */ 858 int 859 kern_wait(pid_t pid, int *status, int options, struct rusage *rusage, int *res) 860 { 861 struct thread *td = curthread; 862 struct lwp *lp; 863 struct proc *q = td->td_proc; 864 struct proc *p, *t; 865 struct pargs *pa; 866 struct sigacts *ps; 867 int nfound, error; 868 long waitgen; 869 870 if (pid == 0) 871 pid = -q->p_pgid; 872 if (options &~ (WUNTRACED|WNOHANG|WCONTINUED|WLINUXCLONE)) 873 return (EINVAL); 874 875 /* 876 * Protect the q->p_children list 877 */ 878 lwkt_gettoken(&q->p_token); 879 loop: 880 /* 881 * All sorts of things can change due to blocking so we have to loop 882 * all the way back up here. 883 * 884 * The problem is that if a process group is stopped and the parent 885 * is doing a wait*(..., WUNTRACED, ...), it will see the STOP 886 * of the child and then stop itself when it tries to return from the 887 * system call. When the process group is resumed the parent will 888 * then get the STOP status even though the child has now resumed 889 * (a followup wait*() will get the CONT status). 890 * 891 * Previously the CONT would overwrite the STOP because the tstop 892 * was handled within tsleep(), and the parent would only see 893 * the CONT when both are stopped and continued together. This little 894 * two-line hack restores this effect. 895 */ 896 while (q->p_stat == SSTOP) 897 tstop(); 898 899 nfound = 0; 900 901 /* 902 * Loop on children. 903 * 904 * NOTE: We don't want to break q's p_token in the loop for the 905 * case where no children are found or we risk breaking the 906 * interlock between child and parent. 907 */ 908 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 909 LIST_FOREACH(p, &q->p_children, p_sibling) { 910 if (pid != WAIT_ANY && 911 p->p_pid != pid && p->p_pgid != -pid) { 912 continue; 913 } 914 915 /* 916 * This special case handles a kthread spawned by linux_clone 917 * (see linux_misc.c). The linux_wait4 and linux_waitpid 918 * functions need to be able to distinguish between waiting 919 * on a process and waiting on a thread. It is a thread if 920 * p_sigparent is not SIGCHLD, and the WLINUXCLONE option 921 * signifies we want to wait for threads and not processes. 922 */ 923 if ((p->p_sigparent != SIGCHLD) ^ 924 ((options & WLINUXCLONE) != 0)) { 925 continue; 926 } 927 928 nfound++; 929 if (p->p_stat == SZOMB) { 930 /* 931 * We may go into SZOMB with threads still present. 932 * We must wait for them to exit before we can reap 933 * the master thread, otherwise we may race reaping 934 * non-master threads. 935 * 936 * Only this routine can remove a process from 937 * the zombie list and destroy it, use PACQUIREZOMB() 938 * to serialize us and loop if it blocks (interlocked 939 * by the parent's q->p_token). 940 * 941 * WARNING! (p) can be invalid when PHOLDZOMB(p) 942 * returns non-zero. Be sure not to 943 * mess with it. 944 */ 945 if (PHOLDZOMB(p)) 946 goto loop; 947 lwkt_gettoken(&p->p_token); 948 if (p->p_pptr != q) { 949 lwkt_reltoken(&p->p_token); 950 PRELEZOMB(p); 951 goto loop; 952 } 953 while (p->p_nthreads > 0) { 954 tsleep(&p->p_nthreads, 0, "lwpzomb", hz); 955 } 956 957 /* 958 * Reap any LWPs left in p->p_lwps. This is usually 959 * just the last LWP. This must be done before 960 * we loop on p_lock since the lwps hold a ref on 961 * it as a vmspace interlock. 962 * 963 * Once that is accomplished p_nthreads had better 964 * be zero. 965 */ 966 while ((lp = RB_ROOT(&p->p_lwp_tree)) != NULL) { 967 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 968 reaplwp(lp); 969 } 970 KKASSERT(p->p_nthreads == 0); 971 972 /* 973 * Don't do anything really bad until all references 974 * to the process go away. This may include other 975 * LWPs which are still in the process of being 976 * reaped. We can't just pull the rug out from under 977 * them because they may still be using the VM space. 978 * 979 * Certain kernel facilities such as /proc will also 980 * put a hold on the process for short periods of 981 * time. 982 */ 983 PRELE(p); 984 PSTALL(p, "reap3", 0); 985 986 /* Take care of our return values. */ 987 *res = p->p_pid; 988 989 if (status) 990 *status = p->p_xstat; 991 if (rusage) 992 *rusage = p->p_ru; 993 994 /* 995 * If we got the child via a ptrace 'attach', 996 * we need to give it back to the old parent. 997 */ 998 if (p->p_oppid && (t = pfind(p->p_oppid)) != NULL) { 999 PHOLD(p); 1000 p->p_oppid = 0; 1001 proc_reparent(p, t); 1002 ksignal(t, SIGCHLD); 1003 wakeup((caddr_t)t); 1004 error = 0; 1005 PRELE(t); 1006 lwkt_reltoken(&p->p_token); 1007 PRELEZOMB(p); 1008 goto done; 1009 } 1010 1011 /* 1012 * Unlink the proc from its process group so that 1013 * the following operations won't lead to an 1014 * inconsistent state for processes running down 1015 * the zombie list. 1016 */ 1017 proc_remove_zombie(p); 1018 proc_userunmap(p); 1019 lwkt_reltoken(&p->p_token); 1020 leavepgrp(p); 1021 1022 p->p_xstat = 0; 1023 ruadd(&q->p_cru, &p->p_ru); 1024 1025 /* 1026 * Decrement the count of procs running with this uid. 1027 */ 1028 chgproccnt(p->p_ucred->cr_ruidinfo, -1, 0); 1029 1030 /* 1031 * Free up credentials. 1032 */ 1033 crfree(p->p_ucred); 1034 p->p_ucred = NULL; 1035 1036 /* 1037 * Remove unused arguments 1038 */ 1039 pa = p->p_args; 1040 p->p_args = NULL; 1041 if (pa && refcount_release(&pa->ar_ref)) { 1042 kfree(pa, M_PARGS); 1043 pa = NULL; 1044 } 1045 1046 ps = p->p_sigacts; 1047 p->p_sigacts = NULL; 1048 if (ps && refcount_release(&ps->ps_refcnt)) { 1049 kfree(ps, M_SUBPROC); 1050 ps = NULL; 1051 } 1052 1053 /* 1054 * Our exitingcount was incremented when the process 1055 * became a zombie, now that the process has been 1056 * removed from (almost) all lists we should be able 1057 * to safely destroy its vmspace. Wait for any current 1058 * holders to go away (so the vmspace remains stable), 1059 * then scrap it. 1060 */ 1061 PSTALL(p, "reap4", 0); 1062 vmspace_exitfree(p); 1063 PSTALL(p, "reap5", 0); 1064 1065 /* 1066 * NOTE: We have to officially release ZOMB in order 1067 * to ensure that a racing thread in kern_wait() 1068 * which blocked on ZOMB is woken up. 1069 */ 1070 PHOLD(p); 1071 PRELEZOMB(p); 1072 kfree(p, M_PROC); 1073 atomic_add_int(&nprocs, -1); 1074 error = 0; 1075 goto done; 1076 } 1077 if (p->p_stat == SSTOP && (p->p_flags & P_WAITED) == 0 && 1078 ((p->p_flags & P_TRACED) || (options & WUNTRACED))) { 1079 PHOLD(p); 1080 lwkt_gettoken(&p->p_token); 1081 if (p->p_pptr != q) { 1082 lwkt_reltoken(&p->p_token); 1083 PRELE(p); 1084 goto loop; 1085 } 1086 if (p->p_stat != SSTOP || 1087 (p->p_flags & P_WAITED) != 0 || 1088 ((p->p_flags & P_TRACED) == 0 && 1089 (options & WUNTRACED) == 0)) { 1090 lwkt_reltoken(&p->p_token); 1091 PRELE(p); 1092 goto loop; 1093 } 1094 1095 p->p_flags |= P_WAITED; 1096 1097 *res = p->p_pid; 1098 if (status) 1099 *status = W_STOPCODE(p->p_xstat); 1100 /* Zero rusage so we get something consistent. */ 1101 if (rusage) 1102 bzero(rusage, sizeof(*rusage)); 1103 error = 0; 1104 lwkt_reltoken(&p->p_token); 1105 PRELE(p); 1106 goto done; 1107 } 1108 if ((options & WCONTINUED) && (p->p_flags & P_CONTINUED)) { 1109 PHOLD(p); 1110 lwkt_gettoken(&p->p_token); 1111 if (p->p_pptr != q) { 1112 lwkt_reltoken(&p->p_token); 1113 PRELE(p); 1114 goto loop; 1115 } 1116 if ((p->p_flags & P_CONTINUED) == 0) { 1117 lwkt_reltoken(&p->p_token); 1118 PRELE(p); 1119 goto loop; 1120 } 1121 1122 *res = p->p_pid; 1123 p->p_flags &= ~P_CONTINUED; 1124 1125 if (status) 1126 *status = SIGCONT; 1127 error = 0; 1128 lwkt_reltoken(&p->p_token); 1129 PRELE(p); 1130 goto done; 1131 } 1132 } 1133 if (nfound == 0) { 1134 error = ECHILD; 1135 goto done; 1136 } 1137 if (options & WNOHANG) { 1138 *res = 0; 1139 error = 0; 1140 goto done; 1141 } 1142 1143 /* 1144 * Wait for signal - interlocked using q->p_waitgen. 1145 */ 1146 error = 0; 1147 while ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1148 tsleep_interlock(q, PCATCH); 1149 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 1150 if ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1151 error = tsleep(q, PCATCH | PINTERLOCKED, "wait", 0); 1152 break; 1153 } 1154 } 1155 if (error) { 1156 done: 1157 lwkt_reltoken(&q->p_token); 1158 return (error); 1159 } 1160 goto loop; 1161 } 1162 1163 /* 1164 * Change child's parent process to parent. 1165 * 1166 * p_children/p_sibling requires the parent's token, and 1167 * changing pptr requires the child's token, so we have to 1168 * get three tokens to do this operation. We also need to 1169 * hold pointers that might get ripped out from under us to 1170 * preserve structural integrity. 1171 * 1172 * It is possible to race another reparent or disconnect or other 1173 * similar operation. We must retry when this situation occurs. 1174 * Once we successfully reparent the process we no longer care 1175 * about any races. 1176 */ 1177 void 1178 proc_reparent(struct proc *child, struct proc *parent) 1179 { 1180 struct proc *opp; 1181 1182 PHOLD(parent); 1183 while ((opp = child->p_pptr) != parent) { 1184 PHOLD(opp); 1185 lwkt_gettoken(&opp->p_token); 1186 lwkt_gettoken(&child->p_token); 1187 lwkt_gettoken(&parent->p_token); 1188 if (child->p_pptr != opp) { 1189 lwkt_reltoken(&parent->p_token); 1190 lwkt_reltoken(&child->p_token); 1191 lwkt_reltoken(&opp->p_token); 1192 PRELE(opp); 1193 continue; 1194 } 1195 LIST_REMOVE(child, p_sibling); 1196 LIST_INSERT_HEAD(&parent->p_children, child, p_sibling); 1197 child->p_pptr = parent; 1198 lwkt_reltoken(&parent->p_token); 1199 lwkt_reltoken(&child->p_token); 1200 lwkt_reltoken(&opp->p_token); 1201 if (LIST_EMPTY(&opp->p_children)) 1202 wakeup(opp); 1203 PRELE(opp); 1204 break; 1205 } 1206 PRELE(parent); 1207 } 1208 1209 /* 1210 * The next two functions are to handle adding/deleting items on the 1211 * exit callout list 1212 * 1213 * at_exit(): 1214 * Take the arguments given and put them onto the exit callout list, 1215 * However first make sure that it's not already there. 1216 * returns 0 on success. 1217 */ 1218 1219 int 1220 at_exit(exitlist_fn function) 1221 { 1222 struct exitlist *ep; 1223 1224 #ifdef INVARIANTS 1225 /* Be noisy if the programmer has lost track of things */ 1226 if (rm_at_exit(function)) 1227 kprintf("WARNING: exit callout entry (%p) already present\n", 1228 function); 1229 #endif 1230 ep = kmalloc(sizeof(*ep), M_ATEXIT, M_NOWAIT); 1231 if (ep == NULL) 1232 return (ENOMEM); 1233 ep->function = function; 1234 TAILQ_INSERT_TAIL(&exit_list, ep, next); 1235 return (0); 1236 } 1237 1238 /* 1239 * Scan the exit callout list for the given item and remove it. 1240 * Returns the number of items removed (0 or 1) 1241 */ 1242 int 1243 rm_at_exit(exitlist_fn function) 1244 { 1245 struct exitlist *ep; 1246 1247 TAILQ_FOREACH(ep, &exit_list, next) { 1248 if (ep->function == function) { 1249 TAILQ_REMOVE(&exit_list, ep, next); 1250 kfree(ep, M_ATEXIT); 1251 return(1); 1252 } 1253 } 1254 return (0); 1255 } 1256 1257 /* 1258 * LWP reaper related code. 1259 */ 1260 static void 1261 reaplwps(void *context, int dummy) 1262 { 1263 struct lwplist *lwplist = context; 1264 struct lwp *lp; 1265 int cpu = mycpuid; 1266 1267 lwkt_gettoken(&deadlwp_token[cpu]); 1268 while ((lp = LIST_FIRST(lwplist))) { 1269 LIST_REMOVE(lp, u.lwp_reap_entry); 1270 reaplwp(lp); 1271 } 1272 lwkt_reltoken(&deadlwp_token[cpu]); 1273 } 1274 1275 static void 1276 reaplwp(struct lwp *lp) 1277 { 1278 while (lwp_wait(lp) == 0) 1279 ; 1280 lwp_dispose(lp); 1281 } 1282 1283 static void 1284 deadlwp_init(void) 1285 { 1286 int cpu; 1287 1288 for (cpu = 0; cpu < ncpus; cpu++) { 1289 lwkt_token_init(&deadlwp_token[cpu], "deadlwpl"); 1290 LIST_INIT(&deadlwp_list[cpu]); 1291 deadlwp_task[cpu] = kmalloc(sizeof(*deadlwp_task[cpu]), 1292 M_DEVBUF, M_WAITOK); 1293 TASK_INIT(deadlwp_task[cpu], 0, reaplwps, &deadlwp_list[cpu]); 1294 } 1295 } 1296 1297 SYSINIT(deadlwpinit, SI_SUB_CONFIGURE, SI_ORDER_ANY, deadlwp_init, NULL); 1298