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/spinlock2.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 int fakestop; 196 197 /* 198 * Interlock against P_WEXIT. Only one of the process's thread 199 * is allowed to do the master exit. 200 */ 201 if (p->p_flags & P_WEXIT) 202 return (EALREADY); 203 p->p_flags |= P_WEXIT; 204 205 /* 206 * Set temporary stopped state in case we are racing a coredump. 207 * Otherwise the coredump may hang forever. 208 */ 209 if (lp->lwp_mpflags & LWP_MP_WSTOP) { 210 fakestop = 0; 211 } else { 212 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 213 ++p->p_nstopped; 214 fakestop = 1; 215 wakeup(&p->p_nstopped); 216 } 217 218 /* 219 * Interlock with LWP_MP_WEXIT and kill any remaining LWPs 220 */ 221 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 222 if (p->p_nthreads > 1) 223 killlwps(lp); 224 225 /* 226 * Undo temporary stopped state 227 */ 228 if (fakestop) { 229 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 230 --p->p_nstopped; 231 } 232 233 /* 234 * If doing this for an exec, clean up the remaining thread 235 * (us) for continuing operation after all the other threads 236 * have been killed. 237 */ 238 if (forexec) { 239 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 240 p->p_flags &= ~P_WEXIT; 241 } 242 return(0); 243 } 244 245 /* 246 * Kill all LWPs except the current one. Do not try to signal 247 * LWPs which have exited on their own or have already been 248 * signaled. 249 */ 250 static void 251 killlwps(struct lwp *lp) 252 { 253 struct proc *p = lp->lwp_proc; 254 struct lwp *tlp; 255 256 /* 257 * Kill the remaining LWPs. We must send the signal before setting 258 * LWP_MP_WEXIT. The setting of WEXIT is optional but helps reduce 259 * races. tlp must be held across the call as it might block and 260 * allow the target lwp to rip itself out from under our loop. 261 */ 262 FOREACH_LWP_IN_PROC(tlp, p) { 263 LWPHOLD(tlp); 264 lwkt_gettoken(&tlp->lwp_token); 265 if ((tlp->lwp_mpflags & LWP_MP_WEXIT) == 0) { 266 atomic_set_int(&tlp->lwp_mpflags, LWP_MP_WEXIT); 267 lwpsignal(p, tlp, SIGKILL); 268 } 269 lwkt_reltoken(&tlp->lwp_token); 270 LWPRELE(tlp); 271 } 272 273 /* 274 * Wait for everything to clear out. Also make sure any tstop()s 275 * are signalled (we are holding p_token for the interlock). 276 */ 277 wakeup(p); 278 while (p->p_nthreads > 1) 279 tsleep(&p->p_nthreads, 0, "killlwps", 0); 280 } 281 282 /* 283 * Exit: deallocate address space and other resources, change proc state 284 * to zombie, and unlink proc from allproc and parent's lists. Save exit 285 * status and rusage for wait(). Check for child processes and orphan them. 286 */ 287 void 288 exit1(int rv) 289 { 290 struct thread *td = curthread; 291 struct proc *p = td->td_proc; 292 struct lwp *lp = td->td_lwp; 293 struct proc *q; 294 struct proc *pp; 295 struct proc *reproc; 296 struct sysreaper *reap; 297 struct vmspace *vm; 298 struct vnode *vtmp; 299 struct exitlist *ep; 300 int error; 301 302 lwkt_gettoken(&p->p_token); 303 304 if (p->p_pid == 1) { 305 kprintf("init died (signal %d, exit %d)\n", 306 WTERMSIG(rv), WEXITSTATUS(rv)); 307 panic("Going nowhere without my init!"); 308 } 309 varsymset_clean(&p->p_varsymset); 310 lockuninit(&p->p_varsymset.vx_lock); 311 312 /* 313 * Kill all lwps associated with the current process, return an 314 * error if we race another thread trying to do the same thing 315 * and lose the race. 316 */ 317 error = killalllwps(0); 318 if (error) { 319 lwp_exit(0, NULL); 320 /* NOT REACHED */ 321 } 322 323 /* are we a task leader? */ 324 if (p == p->p_leader) { 325 struct kill_args killArgs; 326 killArgs.signum = SIGKILL; 327 q = p->p_peers; 328 while(q) { 329 killArgs.pid = q->p_pid; 330 /* 331 * The interface for kill is better 332 * than the internal signal 333 */ 334 sys_kill(&killArgs); 335 q = q->p_peers; 336 } 337 while (p->p_peers) 338 tsleep((caddr_t)p, 0, "exit1", 0); 339 } 340 341 #ifdef PGINPROF 342 vmsizmon(); 343 #endif 344 STOPEVENT(p, S_EXIT, rv); 345 p->p_flags |= P_POSTEXIT; /* stop procfs stepping */ 346 347 /* 348 * Check if any loadable modules need anything done at process exit. 349 * e.g. SYSV IPC stuff 350 * XXX what if one of these generates an error? 351 */ 352 p->p_xstat = rv; 353 354 /* 355 * XXX: imho, the eventhandler stuff is much cleaner than this. 356 * Maybe we should move everything to use eventhandler. 357 */ 358 TAILQ_FOREACH(ep, &exit_list, next) 359 (*ep->function)(td); 360 361 if (p->p_flags & P_PROFIL) 362 stopprofclock(p); 363 364 SIGEMPTYSET(p->p_siglist); 365 SIGEMPTYSET(lp->lwp_siglist); 366 if (timevalisset(&p->p_realtimer.it_value)) 367 callout_stop_sync(&p->p_ithandle); 368 369 /* 370 * Reset any sigio structures pointing to us as a result of 371 * F_SETOWN with our pid. 372 */ 373 funsetownlst(&p->p_sigiolst); 374 375 /* 376 * Close open files and release open-file table. 377 * This may block! 378 */ 379 fdfree(p, NULL); 380 381 if (p->p_leader->p_peers) { 382 q = p->p_leader; 383 while(q->p_peers != p) 384 q = q->p_peers; 385 q->p_peers = p->p_peers; 386 wakeup((caddr_t)p->p_leader); 387 } 388 389 /* 390 * XXX Shutdown SYSV semaphores 391 */ 392 semexit(p); 393 394 KKASSERT(p->p_numposixlocks == 0); 395 396 /* The next two chunks should probably be moved to vmspace_exit. */ 397 vm = p->p_vmspace; 398 399 /* 400 * Clean up data related to virtual kernel operation. Clean up 401 * any vkernel context related to the current lwp now so we can 402 * destroy p_vkernel. 403 */ 404 if (p->p_vkernel) { 405 vkernel_lwp_exit(lp); 406 vkernel_exit(p); 407 } 408 409 /* 410 * Release the user portion of address space. The exitbump prevents 411 * the vmspace from being completely eradicated (using holdcnt). 412 * This releases references to vnodes, which could cause I/O if the 413 * file has been unlinked. We need to do this early enough that 414 * we can still sleep. 415 * 416 * We can't free the entire vmspace as the kernel stack may be mapped 417 * within that space also. 418 * 419 * Processes sharing the same vmspace may exit in one order, and 420 * get cleaned up by vmspace_exit() in a different order. The 421 * last exiting process to reach this point releases as much of 422 * the environment as it can, and the last process cleaned up 423 * by vmspace_exit() (which decrements exitingcnt) cleans up the 424 * remainder. 425 * 426 * NOTE: Releasing p_token around this call is helpful if the 427 * vmspace had a huge RSS. Otherwise some other process 428 * trying to do an allproc or other scan (like 'ps') may 429 * stall for a long time. 430 */ 431 lwkt_reltoken(&p->p_token); 432 vmspace_relexit(vm); 433 lwkt_gettoken(&p->p_token); 434 435 if (SESS_LEADER(p)) { 436 struct session *sp = p->p_session; 437 438 if (sp->s_ttyvp) { 439 /* 440 * We are the controlling process. Signal the 441 * foreground process group, drain the controlling 442 * terminal, and revoke access to the controlling 443 * terminal. 444 * 445 * NOTE: while waiting for the process group to exit 446 * it is possible that one of the processes in the 447 * group will revoke the tty, so the ttyclosesession() 448 * function will re-check sp->s_ttyvp. 449 */ 450 if (sp->s_ttyp && (sp->s_ttyp->t_session == sp)) { 451 if (sp->s_ttyp->t_pgrp) 452 pgsignal(sp->s_ttyp->t_pgrp, SIGHUP, 1); 453 ttywait(sp->s_ttyp); 454 ttyclosesession(sp, 1); /* also revoke */ 455 } 456 /* 457 * Release the tty. If someone has it open via 458 * /dev/tty then close it (since they no longer can 459 * once we've NULL'd it out). 460 */ 461 ttyclosesession(sp, 0); 462 463 /* 464 * s_ttyp is not zero'd; we use this to indicate 465 * that the session once had a controlling terminal. 466 * (for logging and informational purposes) 467 */ 468 } 469 sp->s_leader = NULL; 470 } 471 fixjobc(p, p->p_pgrp, 0); 472 (void)acct_process(p); 473 #ifdef KTRACE 474 /* 475 * release trace file 476 */ 477 if (p->p_tracenode) 478 ktrdestroy(&p->p_tracenode); 479 p->p_traceflag = 0; 480 #endif 481 /* 482 * Release reference to text vnode 483 */ 484 if ((vtmp = p->p_textvp) != NULL) { 485 p->p_textvp = NULL; 486 vrele(vtmp); 487 } 488 489 /* Release namecache handle to text file */ 490 if (p->p_textnch.ncp) 491 cache_drop(&p->p_textnch); 492 493 /* 494 * We have to handle PPWAIT here or proc_move_allproc_zombie() 495 * will block on the PHOLD() the parent is doing. 496 * 497 * We are using the flag as an interlock so an atomic op is 498 * necessary to synchronize with the parent's cpu. 499 */ 500 if (p->p_flags & P_PPWAIT) { 501 if (p->p_pptr && p->p_pptr->p_upmap) 502 atomic_add_int(&p->p_pptr->p_upmap->invfork, -1); 503 atomic_clear_int(&p->p_flags, P_PPWAIT); 504 wakeup(p->p_pptr); 505 } 506 507 /* 508 * Move the process to the zombie list. This will block 509 * until the process p_lock count reaches 0. The process will 510 * not be reaped until TDF_EXITING is set by cpu_thread_exit(), 511 * which is called from cpu_proc_exit(). 512 * 513 * Interlock against waiters using p_waitgen. We increment 514 * p_waitgen after completing the move of our process to the 515 * zombie list. 516 * 517 * WARNING: pp becomes stale when we block, clear it now as a 518 * reminder. 519 */ 520 proc_move_allproc_zombie(p); 521 pp = p->p_pptr; 522 atomic_add_long(&pp->p_waitgen, 1); 523 pp = NULL; 524 525 /* 526 * release controlled reaper for exit if we own it and return the 527 * remaining reaper (the one for us), which we will drop after we 528 * are done. 529 */ 530 reap = reaper_exit(p); 531 532 /* 533 * Reparent all of this process's children to the init process or 534 * to the designated reaper. We must hold the reaper's p_token in 535 * order to safely mess with p_children. 536 * 537 * We already hold p->p_token (to remove the children from our list). 538 */ 539 reproc = NULL; 540 q = LIST_FIRST(&p->p_children); 541 if (q) { 542 reproc = reaper_get(reap); 543 lwkt_gettoken(&reproc->p_token); 544 while ((q = LIST_FIRST(&p->p_children)) != NULL) { 545 PHOLD(q); 546 lwkt_gettoken(&q->p_token); 547 if (q != LIST_FIRST(&p->p_children)) { 548 lwkt_reltoken(&q->p_token); 549 PRELE(q); 550 continue; 551 } 552 LIST_REMOVE(q, p_sibling); 553 LIST_INSERT_HEAD(&reproc->p_children, q, p_sibling); 554 q->p_pptr = reproc; 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 688 */ 689 if (td->td_ucred) { 690 crfree(td->td_ucred); 691 td->td_ucred = NULL; 692 } 693 694 /* 695 * Nobody actually wakes us when the lock 696 * count reaches zero, so just wait one tick. 697 */ 698 while (lp->lwp_lock > 0) 699 tsleep(lp, 0, "lwpexit", 1); 700 701 /* Hand down resource usage to our proc */ 702 ruadd(&p->p_ru, &lp->lwp_ru); 703 704 /* 705 * If we don't hold the process until the LWP is reaped wait*() 706 * may try to dispose of its vmspace before all the LWPs have 707 * actually terminated. 708 */ 709 PHOLD(p); 710 711 /* 712 * Do any remaining work that might block on us. We should be 713 * coded such that further blocking is ok after decrementing 714 * p_nthreads but don't take the chance. 715 */ 716 dsched_exit_thread(td); 717 biosched_done(curthread); 718 719 /* 720 * We have to use the reaper for all the LWPs except the one doing 721 * the master exit. The LWP doing the master exit can just be 722 * left on p_lwps and the process reaper will deal with it 723 * synchronously, which is much faster. 724 * 725 * Wakeup anyone waiting on p_nthreads to drop to 1 or 0. 726 * 727 * The process is left held until the reaper calls lwp_dispose() on 728 * the lp (after calling lwp_wait()). 729 */ 730 if (masterexit == 0) { 731 int cpu = mycpuid; 732 733 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 734 --p->p_nthreads; 735 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 736 dowake = 1; 737 lwkt_gettoken(&deadlwp_token[cpu]); 738 LIST_INSERT_HEAD(&deadlwp_list[cpu], lp, u.lwp_reap_entry); 739 taskqueue_enqueue(taskqueue_thread[cpu], deadlwp_task[cpu]); 740 lwkt_reltoken(&deadlwp_token[cpu]); 741 } else { 742 --p->p_nthreads; 743 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 744 dowake = 1; 745 } 746 747 /* 748 * We no longer need p_token. 749 * 750 * Tell the userland scheduler that we are going away 751 */ 752 lwkt_reltoken(&p->p_token); 753 p->p_usched->heuristic_exiting(lp, p); 754 755 /* 756 * Issue late wakeups after releasing our token to give us a chance 757 * to deschedule and switch away before another cpu in a wait*() 758 * reaps us. This is done as late as possible to reduce contention. 759 */ 760 if (dowake) 761 wakeup(&p->p_nthreads); 762 if (waddr) 763 wakeup(waddr); 764 765 cpu_lwp_exit(); 766 } 767 768 /* 769 * Wait until a lwp is completely dead. The final interlock in this drama 770 * is when TDF_EXITING is set in cpu_thread_exit() just before the final 771 * switchout. 772 * 773 * At the point TDF_EXITING is set a complete exit is accomplished when 774 * TDF_RUNNING and TDF_PREEMPT_LOCK are both clear. td_mpflags has two 775 * post-switch interlock flags that can be used to wait for the TDF_ 776 * flags to clear. 777 * 778 * Returns non-zero on success, and zero if the caller needs to retry 779 * the lwp_wait(). 780 */ 781 static int 782 lwp_wait(struct lwp *lp) 783 { 784 struct thread *td = lp->lwp_thread; 785 u_int mpflags; 786 787 KKASSERT(lwkt_preempted_proc() != lp); 788 789 /* 790 * This bit of code uses the thread destruction interlock 791 * managed by lwkt_switch_return() to wait for the lwp's 792 * thread to completely disengage. 793 * 794 * It is possible for us to race another cpu core so we 795 * have to do this correctly. 796 */ 797 for (;;) { 798 mpflags = td->td_mpflags; 799 cpu_ccfence(); 800 if (mpflags & TDF_MP_EXITSIG) 801 break; 802 tsleep_interlock(td, 0); 803 if (atomic_cmpset_int(&td->td_mpflags, mpflags, 804 mpflags | TDF_MP_EXITWAIT)) { 805 tsleep(td, PINTERLOCKED, "lwpxt", 0); 806 } 807 } 808 809 /* 810 * We've already waited for the core exit but there can still 811 * be other refs from e.g. process scans and such. 812 */ 813 if (lp->lwp_lock > 0) { 814 tsleep(lp, 0, "lwpwait1", 1); 815 return(0); 816 } 817 if (td->td_refs) { 818 tsleep(td, 0, "lwpwait2", 1); 819 return(0); 820 } 821 822 /* 823 * Now that we have the thread destruction interlock these flags 824 * really should already be cleaned up, keep a check for safety. 825 * 826 * We can't rip its stack out from under it until TDF_EXITING is 827 * set and both TDF_RUNNING and TDF_PREEMPT_LOCK are clear. 828 * TDF_PREEMPT_LOCK must be checked because TDF_RUNNING 829 * will be cleared temporarily if a thread gets preempted. 830 */ 831 while ((td->td_flags & (TDF_RUNNING | 832 TDF_RUNQ | 833 TDF_PREEMPT_LOCK | 834 TDF_EXITING)) != TDF_EXITING) { 835 tsleep(lp, 0, "lwpwait3", 1); 836 return (0); 837 } 838 839 KASSERT((td->td_flags & (TDF_RUNQ|TDF_TSLEEPQ)) == 0, 840 ("lwp_wait: td %p (%s) still on run or sleep queue", 841 td, td->td_comm)); 842 return (1); 843 } 844 845 /* 846 * Release the resources associated with a lwp. 847 * The lwp must be completely dead. 848 */ 849 void 850 lwp_dispose(struct lwp *lp) 851 { 852 struct thread *td = lp->lwp_thread; 853 854 KKASSERT(lwkt_preempted_proc() != lp); 855 KKASSERT(lp->lwp_lock == 0); 856 KKASSERT(td->td_refs == 0); 857 KKASSERT((td->td_flags & (TDF_RUNNING | 858 TDF_RUNQ | 859 TDF_PREEMPT_LOCK | 860 TDF_EXITING)) == TDF_EXITING); 861 862 PRELE(lp->lwp_proc); 863 lp->lwp_proc = NULL; 864 if (td != NULL) { 865 td->td_proc = NULL; 866 td->td_lwp = NULL; 867 lp->lwp_thread = NULL; 868 lwkt_free_thread(td); 869 } 870 kfree(lp, M_LWP); 871 } 872 873 int 874 sys_wait4(struct wait_args *uap) 875 { 876 struct rusage rusage; 877 int error, status; 878 879 error = kern_wait(uap->pid, (uap->status ? &status : NULL), 880 uap->options, (uap->rusage ? &rusage : NULL), 881 &uap->sysmsg_result); 882 883 if (error == 0 && uap->status) 884 error = copyout(&status, uap->status, sizeof(*uap->status)); 885 if (error == 0 && uap->rusage) 886 error = copyout(&rusage, uap->rusage, sizeof(*uap->rusage)); 887 return (error); 888 } 889 890 /* 891 * wait1() 892 * 893 * wait_args(int pid, int *status, int options, struct rusage *rusage) 894 */ 895 int 896 kern_wait(pid_t pid, int *status, int options, struct rusage *rusage, int *res) 897 { 898 struct thread *td = curthread; 899 struct lwp *lp; 900 struct proc *q = td->td_proc; 901 struct proc *p, *t; 902 struct ucred *cr; 903 struct pargs *pa; 904 struct sigacts *ps; 905 int nfound, error; 906 long waitgen; 907 908 if (pid == 0) 909 pid = -q->p_pgid; 910 if (options &~ (WUNTRACED|WNOHANG|WCONTINUED|WLINUXCLONE)) 911 return (EINVAL); 912 913 /* 914 * Protect the q->p_children list 915 */ 916 lwkt_gettoken(&q->p_token); 917 loop: 918 /* 919 * All sorts of things can change due to blocking so we have to loop 920 * all the way back up here. 921 * 922 * The problem is that if a process group is stopped and the parent 923 * is doing a wait*(..., WUNTRACED, ...), it will see the STOP 924 * of the child and then stop itself when it tries to return from the 925 * system call. When the process group is resumed the parent will 926 * then get the STOP status even though the child has now resumed 927 * (a followup wait*() will get the CONT status). 928 * 929 * Previously the CONT would overwrite the STOP because the tstop 930 * was handled within tsleep(), and the parent would only see 931 * the CONT when both are stopped and continued together. This little 932 * two-line hack restores this effect. 933 */ 934 if (STOPLWP(q, td->td_lwp)) 935 tstop(); 936 937 nfound = 0; 938 939 /* 940 * Loop on children. 941 * 942 * NOTE: We don't want to break q's p_token in the loop for the 943 * case where no children are found or we risk breaking the 944 * interlock between child and parent. 945 */ 946 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 947 LIST_FOREACH(p, &q->p_children, p_sibling) { 948 if (pid != WAIT_ANY && 949 p->p_pid != pid && p->p_pgid != -pid) { 950 continue; 951 } 952 953 /* 954 * This special case handles a kthread spawned by linux_clone 955 * (see linux_misc.c). The linux_wait4 and linux_waitpid 956 * functions need to be able to distinguish between waiting 957 * on a process and waiting on a thread. It is a thread if 958 * p_sigparent is not SIGCHLD, and the WLINUXCLONE option 959 * signifies we want to wait for threads and not processes. 960 */ 961 if ((p->p_sigparent != SIGCHLD) ^ 962 ((options & WLINUXCLONE) != 0)) { 963 continue; 964 } 965 966 nfound++; 967 if (p->p_stat == SZOMB) { 968 /* 969 * We may go into SZOMB with threads still present. 970 * We must wait for them to exit before we can reap 971 * the master thread, otherwise we may race reaping 972 * non-master threads. 973 * 974 * Only this routine can remove a process from 975 * the zombie list and destroy it, use PACQUIREZOMB() 976 * to serialize us and loop if it blocks (interlocked 977 * by the parent's q->p_token). 978 * 979 * WARNING! (p) can be invalid when PHOLDZOMB(p) 980 * returns non-zero. Be sure not to 981 * mess with it. 982 */ 983 if (PHOLDZOMB(p)) 984 goto loop; 985 lwkt_gettoken(&p->p_token); 986 if (p->p_pptr != q) { 987 lwkt_reltoken(&p->p_token); 988 PRELEZOMB(p); 989 goto loop; 990 } 991 while (p->p_nthreads > 0) { 992 tsleep(&p->p_nthreads, 0, "lwpzomb", hz); 993 } 994 995 /* 996 * Reap any LWPs left in p->p_lwps. This is usually 997 * just the last LWP. This must be done before 998 * we loop on p_lock since the lwps hold a ref on 999 * it as a vmspace interlock. 1000 * 1001 * Once that is accomplished p_nthreads had better 1002 * be zero. 1003 */ 1004 while ((lp = RB_ROOT(&p->p_lwp_tree)) != NULL) { 1005 /* 1006 * Make sure no one is using this lwp, before 1007 * it is removed from the tree. If we didn't 1008 * wait it here, lwp tree iteration with 1009 * blocking operation would be broken. 1010 */ 1011 while (lp->lwp_lock > 0) 1012 tsleep(lp, 0, "zomblwp", 1); 1013 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 1014 reaplwp(lp); 1015 } 1016 KKASSERT(p->p_nthreads == 0); 1017 1018 /* 1019 * Don't do anything really bad until all references 1020 * to the process go away. This may include other 1021 * LWPs which are still in the process of being 1022 * reaped. We can't just pull the rug out from under 1023 * them because they may still be using the VM space. 1024 * 1025 * Certain kernel facilities such as /proc will also 1026 * put a hold on the process for short periods of 1027 * time. 1028 */ 1029 PRELE(p); 1030 PSTALL(p, "reap3", 0); 1031 1032 /* Take care of our return values. */ 1033 *res = p->p_pid; 1034 1035 if (status) 1036 *status = p->p_xstat; 1037 if (rusage) 1038 *rusage = p->p_ru; 1039 1040 /* 1041 * If we got the child via a ptrace 'attach', 1042 * we need to give it back to the old parent. 1043 */ 1044 if (p->p_oppid && (t = pfind(p->p_oppid)) != NULL) { 1045 PHOLD(p); 1046 p->p_oppid = 0; 1047 proc_reparent(p, t); 1048 ksignal(t, SIGCHLD); 1049 wakeup((caddr_t)t); 1050 error = 0; 1051 PRELE(t); 1052 lwkt_reltoken(&p->p_token); 1053 PRELEZOMB(p); 1054 goto done; 1055 } 1056 1057 /* 1058 * Unlink the proc from its process group so that 1059 * the following operations won't lead to an 1060 * inconsistent state for processes running down 1061 * the zombie list. 1062 */ 1063 proc_remove_zombie(p); 1064 proc_userunmap(p); 1065 lwkt_reltoken(&p->p_token); 1066 leavepgrp(p); 1067 1068 p->p_xstat = 0; 1069 ruadd(&q->p_cru, &p->p_ru); 1070 1071 /* 1072 * Decrement the count of procs running with this uid. 1073 */ 1074 chgproccnt(p->p_ucred->cr_ruidinfo, -1, 0); 1075 1076 /* 1077 * Free up credentials. p_spin is required to 1078 * avoid races against allproc scans. 1079 */ 1080 spin_lock(&p->p_spin); 1081 cr = p->p_ucred; 1082 p->p_ucred = NULL; 1083 spin_unlock(&p->p_spin); 1084 crfree(cr); 1085 1086 /* 1087 * Remove unused arguments 1088 */ 1089 pa = p->p_args; 1090 p->p_args = NULL; 1091 if (pa && refcount_release(&pa->ar_ref)) { 1092 kfree(pa, M_PARGS); 1093 pa = NULL; 1094 } 1095 1096 ps = p->p_sigacts; 1097 p->p_sigacts = NULL; 1098 if (ps && refcount_release(&ps->ps_refcnt)) { 1099 kfree(ps, M_SUBPROC); 1100 ps = NULL; 1101 } 1102 1103 /* 1104 * Our exitingcount was incremented when the process 1105 * became a zombie, now that the process has been 1106 * removed from (almost) all lists we should be able 1107 * to safely destroy its vmspace. Wait for any current 1108 * holders to go away (so the vmspace remains stable), 1109 * then scrap it. 1110 * 1111 * NOTE: Releasing the parent process (q) p_token 1112 * across the vmspace_exitfree() call is 1113 * important here to reduce stalls on 1114 * interactions with (q) (such as 1115 * fork/exec/wait or 'ps'). 1116 */ 1117 PSTALL(p, "reap4", 0); 1118 lwkt_reltoken(&q->p_token); 1119 vmspace_exitfree(p); 1120 lwkt_gettoken(&q->p_token); 1121 PSTALL(p, "reap5", 0); 1122 1123 /* 1124 * NOTE: We have to officially release ZOMB in order 1125 * to ensure that a racing thread in kern_wait() 1126 * which blocked on ZOMB is woken up. 1127 */ 1128 PHOLD(p); 1129 PRELEZOMB(p); 1130 kfree(p, M_PROC); 1131 atomic_add_int(&nprocs, -1); 1132 error = 0; 1133 goto done; 1134 } 1135 if ((p->p_stat == SSTOP || p->p_stat == SCORE) && 1136 (p->p_flags & P_WAITED) == 0 && 1137 ((p->p_flags & P_TRACED) || (options & WUNTRACED))) { 1138 PHOLD(p); 1139 lwkt_gettoken(&p->p_token); 1140 if (p->p_pptr != q) { 1141 lwkt_reltoken(&p->p_token); 1142 PRELE(p); 1143 goto loop; 1144 } 1145 if ((p->p_stat != SSTOP && p->p_stat != SCORE) || 1146 (p->p_flags & P_WAITED) != 0 || 1147 ((p->p_flags & P_TRACED) == 0 && 1148 (options & WUNTRACED) == 0)) { 1149 lwkt_reltoken(&p->p_token); 1150 PRELE(p); 1151 goto loop; 1152 } 1153 1154 p->p_flags |= P_WAITED; 1155 1156 *res = p->p_pid; 1157 if (status) 1158 *status = W_STOPCODE(p->p_xstat); 1159 /* Zero rusage so we get something consistent. */ 1160 if (rusage) 1161 bzero(rusage, sizeof(*rusage)); 1162 error = 0; 1163 lwkt_reltoken(&p->p_token); 1164 PRELE(p); 1165 goto done; 1166 } 1167 if ((options & WCONTINUED) && (p->p_flags & P_CONTINUED)) { 1168 PHOLD(p); 1169 lwkt_gettoken(&p->p_token); 1170 if (p->p_pptr != q) { 1171 lwkt_reltoken(&p->p_token); 1172 PRELE(p); 1173 goto loop; 1174 } 1175 if ((p->p_flags & P_CONTINUED) == 0) { 1176 lwkt_reltoken(&p->p_token); 1177 PRELE(p); 1178 goto loop; 1179 } 1180 1181 *res = p->p_pid; 1182 p->p_flags &= ~P_CONTINUED; 1183 1184 if (status) 1185 *status = SIGCONT; 1186 error = 0; 1187 lwkt_reltoken(&p->p_token); 1188 PRELE(p); 1189 goto done; 1190 } 1191 } 1192 if (nfound == 0) { 1193 error = ECHILD; 1194 goto done; 1195 } 1196 if (options & WNOHANG) { 1197 *res = 0; 1198 error = 0; 1199 goto done; 1200 } 1201 1202 /* 1203 * Wait for signal - interlocked using q->p_waitgen. 1204 */ 1205 error = 0; 1206 while ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1207 tsleep_interlock(q, PCATCH); 1208 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 1209 if ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1210 error = tsleep(q, PCATCH | PINTERLOCKED, "wait", 0); 1211 break; 1212 } 1213 } 1214 if (error) { 1215 done: 1216 lwkt_reltoken(&q->p_token); 1217 return (error); 1218 } 1219 goto loop; 1220 } 1221 1222 /* 1223 * Change child's parent process to parent. 1224 * 1225 * p_children/p_sibling requires the parent's token, and 1226 * changing pptr requires the child's token, so we have to 1227 * get three tokens to do this operation. We also need to 1228 * hold pointers that might get ripped out from under us to 1229 * preserve structural integrity. 1230 * 1231 * It is possible to race another reparent or disconnect or other 1232 * similar operation. We must retry when this situation occurs. 1233 * Once we successfully reparent the process we no longer care 1234 * about any races. 1235 */ 1236 void 1237 proc_reparent(struct proc *child, struct proc *parent) 1238 { 1239 struct proc *opp; 1240 1241 PHOLD(parent); 1242 while ((opp = child->p_pptr) != parent) { 1243 PHOLD(opp); 1244 lwkt_gettoken(&opp->p_token); 1245 lwkt_gettoken(&child->p_token); 1246 lwkt_gettoken(&parent->p_token); 1247 if (child->p_pptr != opp) { 1248 lwkt_reltoken(&parent->p_token); 1249 lwkt_reltoken(&child->p_token); 1250 lwkt_reltoken(&opp->p_token); 1251 PRELE(opp); 1252 continue; 1253 } 1254 LIST_REMOVE(child, p_sibling); 1255 LIST_INSERT_HEAD(&parent->p_children, child, p_sibling); 1256 child->p_pptr = parent; 1257 lwkt_reltoken(&parent->p_token); 1258 lwkt_reltoken(&child->p_token); 1259 lwkt_reltoken(&opp->p_token); 1260 if (LIST_EMPTY(&opp->p_children)) 1261 wakeup(opp); 1262 PRELE(opp); 1263 break; 1264 } 1265 PRELE(parent); 1266 } 1267 1268 /* 1269 * The next two functions are to handle adding/deleting items on the 1270 * exit callout list 1271 * 1272 * at_exit(): 1273 * Take the arguments given and put them onto the exit callout list, 1274 * However first make sure that it's not already there. 1275 * returns 0 on success. 1276 */ 1277 1278 int 1279 at_exit(exitlist_fn function) 1280 { 1281 struct exitlist *ep; 1282 1283 #ifdef INVARIANTS 1284 /* Be noisy if the programmer has lost track of things */ 1285 if (rm_at_exit(function)) 1286 kprintf("WARNING: exit callout entry (%p) already present\n", 1287 function); 1288 #endif 1289 ep = kmalloc(sizeof(*ep), M_ATEXIT, M_NOWAIT); 1290 if (ep == NULL) 1291 return (ENOMEM); 1292 ep->function = function; 1293 TAILQ_INSERT_TAIL(&exit_list, ep, next); 1294 return (0); 1295 } 1296 1297 /* 1298 * Scan the exit callout list for the given item and remove it. 1299 * Returns the number of items removed (0 or 1) 1300 */ 1301 int 1302 rm_at_exit(exitlist_fn function) 1303 { 1304 struct exitlist *ep; 1305 1306 TAILQ_FOREACH(ep, &exit_list, next) { 1307 if (ep->function == function) { 1308 TAILQ_REMOVE(&exit_list, ep, next); 1309 kfree(ep, M_ATEXIT); 1310 return(1); 1311 } 1312 } 1313 return (0); 1314 } 1315 1316 /* 1317 * LWP reaper related code. 1318 */ 1319 static void 1320 reaplwps(void *context, int dummy) 1321 { 1322 struct lwplist *lwplist = context; 1323 struct lwp *lp; 1324 int cpu = mycpuid; 1325 1326 lwkt_gettoken(&deadlwp_token[cpu]); 1327 while ((lp = LIST_FIRST(lwplist))) { 1328 LIST_REMOVE(lp, u.lwp_reap_entry); 1329 reaplwp(lp); 1330 } 1331 lwkt_reltoken(&deadlwp_token[cpu]); 1332 } 1333 1334 static void 1335 reaplwp(struct lwp *lp) 1336 { 1337 while (lwp_wait(lp) == 0) 1338 ; 1339 lwp_dispose(lp); 1340 } 1341 1342 static void 1343 deadlwp_init(void) 1344 { 1345 int cpu; 1346 1347 for (cpu = 0; cpu < ncpus; cpu++) { 1348 lwkt_token_init(&deadlwp_token[cpu], "deadlwpl"); 1349 LIST_INIT(&deadlwp_list[cpu]); 1350 deadlwp_task[cpu] = kmalloc(sizeof(*deadlwp_task[cpu]), 1351 M_DEVBUF, M_WAITOK); 1352 TASK_INIT(deadlwp_task[cpu], 0, reaplwps, &deadlwp_list[cpu]); 1353 } 1354 } 1355 1356 SYSINIT(deadlwpinit, SI_SUB_CONFIGURE, SI_ORDER_ANY, deadlwp_init, NULL); 1357