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 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 EVENTHANDLER_INVOKE(process_exit, p); 354 355 /* 356 * XXX: imho, the eventhandler stuff is much cleaner than this. 357 * Maybe we should move everything to use eventhandler. 358 */ 359 TAILQ_FOREACH(ep, &exit_list, next) 360 (*ep->function)(td); 361 362 if (p->p_flags & P_PROFIL) 363 stopprofclock(p); 364 365 SIGEMPTYSET(p->p_siglist); 366 SIGEMPTYSET(lp->lwp_siglist); 367 if (timevalisset(&p->p_realtimer.it_value)) 368 callout_stop_sync(&p->p_ithandle); 369 370 /* 371 * Reset any sigio structures pointing to us as a result of 372 * F_SETOWN with our pid. 373 */ 374 funsetownlst(&p->p_sigiolst); 375 376 /* 377 * Close open files and release open-file table. 378 * This may block! 379 */ 380 fdfree(p, NULL); 381 382 if (p->p_leader->p_peers) { 383 q = p->p_leader; 384 while(q->p_peers != p) 385 q = q->p_peers; 386 q->p_peers = p->p_peers; 387 wakeup((caddr_t)p->p_leader); 388 } 389 390 /* 391 * XXX Shutdown SYSV semaphores 392 */ 393 semexit(p); 394 395 KKASSERT(p->p_numposixlocks == 0); 396 397 /* The next two chunks should probably be moved to vmspace_exit. */ 398 vm = p->p_vmspace; 399 400 /* 401 * Clean up data related to virtual kernel operation. Clean up 402 * any vkernel context related to the current lwp now so we can 403 * destroy p_vkernel. 404 */ 405 if (p->p_vkernel) { 406 vkernel_lwp_exit(lp); 407 vkernel_exit(p); 408 } 409 410 /* 411 * Release the user portion of address space. The exitbump prevents 412 * the vmspace from being completely eradicated (using holdcnt). 413 * This releases references to vnodes, which could cause I/O if the 414 * file has been unlinked. We need to do this early enough that 415 * we can still sleep. 416 * 417 * We can't free the entire vmspace as the kernel stack may be mapped 418 * within that space also. 419 * 420 * Processes sharing the same vmspace may exit in one order, and 421 * get cleaned up by vmspace_exit() in a different order. The 422 * last exiting process to reach this point releases as much of 423 * the environment as it can, and the last process cleaned up 424 * by vmspace_exit() (which decrements exitingcnt) cleans up the 425 * remainder. 426 * 427 * NOTE: Releasing p_token around this call is helpful if the 428 * vmspace had a huge RSS. Otherwise some other process 429 * trying to do an allproc or other scan (like 'ps') may 430 * stall for a long time. 431 */ 432 lwkt_reltoken(&p->p_token); 433 vmspace_relexit(vm); 434 lwkt_gettoken(&p->p_token); 435 436 if (SESS_LEADER(p)) { 437 struct session *sp = p->p_session; 438 439 if (sp->s_ttyvp) { 440 /* 441 * We are the controlling process. Signal the 442 * foreground process group, drain the controlling 443 * terminal, and revoke access to the controlling 444 * terminal. 445 * 446 * NOTE: while waiting for the process group to exit 447 * it is possible that one of the processes in the 448 * group will revoke the tty, so the ttyclosesession() 449 * function will re-check sp->s_ttyvp. 450 */ 451 if (sp->s_ttyp && (sp->s_ttyp->t_session == sp)) { 452 if (sp->s_ttyp->t_pgrp) 453 pgsignal(sp->s_ttyp->t_pgrp, SIGHUP, 1); 454 ttywait(sp->s_ttyp); 455 ttyclosesession(sp, 1); /* also revoke */ 456 } 457 /* 458 * Release the tty. If someone has it open via 459 * /dev/tty then close it (since they no longer can 460 * once we've NULL'd it out). 461 */ 462 ttyclosesession(sp, 0); 463 464 /* 465 * s_ttyp is not zero'd; we use this to indicate 466 * that the session once had a controlling terminal. 467 * (for logging and informational purposes) 468 */ 469 } 470 sp->s_leader = NULL; 471 } 472 fixjobc(p, p->p_pgrp, 0); 473 (void)acct_process(p); 474 #ifdef KTRACE 475 /* 476 * release trace file 477 */ 478 if (p->p_tracenode) 479 ktrdestroy(&p->p_tracenode); 480 p->p_traceflag = 0; 481 #endif 482 /* 483 * Release reference to text vnode 484 */ 485 if ((vtmp = p->p_textvp) != NULL) { 486 p->p_textvp = NULL; 487 vrele(vtmp); 488 } 489 490 /* Release namecache handle to text file */ 491 if (p->p_textnch.ncp) 492 cache_drop(&p->p_textnch); 493 494 /* 495 * We have to handle PPWAIT here or proc_move_allproc_zombie() 496 * will block on the PHOLD() the parent is doing. 497 * 498 * We are using the flag as an interlock so an atomic op is 499 * necessary to synchronize with the parent's cpu. 500 */ 501 if (p->p_flags & P_PPWAIT) { 502 if (p->p_pptr && p->p_pptr->p_upmap) 503 atomic_add_int(&p->p_pptr->p_upmap->invfork, -1); 504 atomic_clear_int(&p->p_flags, P_PPWAIT); 505 wakeup(p->p_pptr); 506 } 507 508 /* 509 * Move the process to the zombie list. This will block 510 * until the process p_lock count reaches 0. The process will 511 * not be reaped until TDF_EXITING is set by cpu_thread_exit(), 512 * which is called from cpu_proc_exit(). 513 * 514 * Interlock against waiters using p_waitgen. We increment 515 * p_waitgen after completing the move of our process to the 516 * zombie list. 517 * 518 * WARNING: pp becomes stale when we block, clear it now as a 519 * reminder. 520 */ 521 proc_move_allproc_zombie(p); 522 pp = p->p_pptr; 523 atomic_add_long(&pp->p_waitgen, 1); 524 pp = NULL; 525 526 /* 527 * release controlled reaper for exit if we own it and return the 528 * remaining reaper (the one for us), which we will drop after we 529 * are done. 530 */ 531 reap = reaper_exit(p); 532 533 /* 534 * Reparent all of this process's children to the init process or 535 * to the designated reaper. We must hold the reaper's p_token in 536 * order to safely mess with p_children. 537 * 538 * We already hold p->p_token (to remove the children from our list). 539 */ 540 reproc = NULL; 541 q = LIST_FIRST(&p->p_children); 542 if (q) { 543 reproc = reaper_get(reap); 544 lwkt_gettoken(&reproc->p_token); 545 while ((q = LIST_FIRST(&p->p_children)) != NULL) { 546 PHOLD(q); 547 lwkt_gettoken(&q->p_token); 548 if (q != LIST_FIRST(&p->p_children)) { 549 lwkt_reltoken(&q->p_token); 550 PRELE(q); 551 continue; 552 } 553 LIST_REMOVE(q, p_sibling); 554 LIST_INSERT_HEAD(&reproc->p_children, q, p_sibling); 555 q->p_pptr = reproc; 556 q->p_sigparent = SIGCHLD; 557 558 /* 559 * Traced processes are killed 560 * since their existence means someone is screwing up. 561 */ 562 if (q->p_flags & P_TRACED) { 563 q->p_flags &= ~P_TRACED; 564 ksignal(q, SIGKILL); 565 } 566 lwkt_reltoken(&q->p_token); 567 PRELE(q); 568 } 569 lwkt_reltoken(&reproc->p_token); 570 wakeup(reproc); 571 } 572 573 /* 574 * Save exit status and final rusage info, adding in child rusage 575 * info and self times. 576 */ 577 calcru_proc(p, &p->p_ru); 578 ruadd(&p->p_ru, &p->p_cru); 579 580 /* 581 * notify interested parties of our demise. 582 */ 583 KNOTE(&p->p_klist, NOTE_EXIT); 584 585 /* 586 * Notify parent that we're gone. If parent has the PS_NOCLDWAIT 587 * flag set, or if the handler is set to SIG_IGN, notify the reaper 588 * instead (it will handle this situation). 589 * 590 * NOTE: The reaper can still be the parent process. 591 * 592 * (must reload pp) 593 */ 594 if (p->p_pptr->p_sigacts->ps_flag & (PS_NOCLDWAIT | PS_CLDSIGIGN)) { 595 if (reproc == NULL) 596 reproc = reaper_get(reap); 597 proc_reparent(p, reproc); 598 } 599 if (reproc) 600 PRELE(reproc); 601 if (reap) 602 reaper_drop(reap); 603 604 /* 605 * Signal (possibly new) parent. 606 */ 607 pp = p->p_pptr; 608 PHOLD(pp); 609 if (p->p_sigparent && pp != initproc) { 610 int sig = p->p_sigparent; 611 612 if (sig != SIGUSR1 && sig != SIGCHLD) 613 sig = SIGCHLD; 614 ksignal(pp, sig); 615 } else { 616 ksignal(pp, SIGCHLD); 617 } 618 p->p_flags &= ~P_TRACED; 619 PRELE(pp); 620 621 /* 622 * cpu_exit is responsible for clearing curproc, since 623 * it is heavily integrated with the thread/switching sequence. 624 * 625 * Other substructures are freed from wait(). 626 */ 627 plimit_free(p); 628 629 /* 630 * Finally, call machine-dependent code to release as many of the 631 * lwp's resources as we can and halt execution of this thread. 632 * 633 * pp is a wild pointer now but still the correct wakeup() target. 634 * lwp_exit() only uses it to send the wakeup() signal to the likely 635 * parent. Any reparenting race that occurs will get a signal 636 * automatically and not be an issue. 637 */ 638 lwp_exit(1, pp); 639 } 640 641 /* 642 * Eventually called by every exiting LWP 643 * 644 * p->p_token must be held. mplock may be held and will be released. 645 */ 646 void 647 lwp_exit(int masterexit, void *waddr) 648 { 649 struct thread *td = curthread; 650 struct lwp *lp = td->td_lwp; 651 struct proc *p = lp->lwp_proc; 652 int dowake = 0; 653 654 /* 655 * Release the current user process designation on the process so 656 * the userland scheduler can work in someone else. 657 */ 658 p->p_usched->release_curproc(lp); 659 660 /* 661 * lwp_exit() may be called without setting LWP_MP_WEXIT, so 662 * make sure it is set here. 663 */ 664 ASSERT_LWKT_TOKEN_HELD(&p->p_token); 665 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 666 667 /* 668 * Clean up any virtualization 669 */ 670 if (lp->lwp_vkernel) 671 vkernel_lwp_exit(lp); 672 673 if (td->td_vmm) 674 vmm_vmdestroy(); 675 676 /* 677 * Clean up select/poll support 678 */ 679 kqueue_terminate(&lp->lwp_kqueue); 680 681 /* 682 * Clean up any syscall-cached ucred 683 */ 684 if (td->td_ucred) { 685 crfree(td->td_ucred); 686 td->td_ucred = NULL; 687 } 688 689 /* 690 * Nobody actually wakes us when the lock 691 * count reaches zero, so just wait one tick. 692 */ 693 while (lp->lwp_lock > 0) 694 tsleep(lp, 0, "lwpexit", 1); 695 696 /* Hand down resource usage to our proc */ 697 ruadd(&p->p_ru, &lp->lwp_ru); 698 699 /* 700 * If we don't hold the process until the LWP is reaped wait*() 701 * may try to dispose of its vmspace before all the LWPs have 702 * actually terminated. 703 */ 704 PHOLD(p); 705 706 /* 707 * Do any remaining work that might block on us. We should be 708 * coded such that further blocking is ok after decrementing 709 * p_nthreads but don't take the chance. 710 */ 711 dsched_exit_thread(td); 712 biosched_done(curthread); 713 714 /* 715 * We have to use the reaper for all the LWPs except the one doing 716 * the master exit. The LWP doing the master exit can just be 717 * left on p_lwps and the process reaper will deal with it 718 * synchronously, which is much faster. 719 * 720 * Wakeup anyone waiting on p_nthreads to drop to 1 or 0. 721 * 722 * The process is left held until the reaper calls lwp_dispose() on 723 * the lp (after calling lwp_wait()). 724 */ 725 if (masterexit == 0) { 726 int cpu = mycpuid; 727 728 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 729 --p->p_nthreads; 730 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 731 dowake = 1; 732 lwkt_gettoken(&deadlwp_token[cpu]); 733 LIST_INSERT_HEAD(&deadlwp_list[cpu], lp, u.lwp_reap_entry); 734 taskqueue_enqueue(taskqueue_thread[cpu], deadlwp_task[cpu]); 735 lwkt_reltoken(&deadlwp_token[cpu]); 736 } else { 737 --p->p_nthreads; 738 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 739 dowake = 1; 740 } 741 742 /* 743 * We no longer need p_token. 744 * 745 * Tell the userland scheduler that we are going away 746 */ 747 lwkt_reltoken(&p->p_token); 748 p->p_usched->heuristic_exiting(lp, p); 749 750 /* 751 * Issue late wakeups after releasing our token to give us a chance 752 * to deschedule and switch away before another cpu in a wait*() 753 * reaps us. This is done as late as possible to reduce contention. 754 */ 755 if (dowake) 756 wakeup(&p->p_nthreads); 757 if (waddr) 758 wakeup(waddr); 759 760 cpu_lwp_exit(); 761 } 762 763 /* 764 * Wait until a lwp is completely dead. The final interlock in this drama 765 * is when TDF_EXITING is set in cpu_thread_exit() just before the final 766 * switchout. 767 * 768 * At the point TDF_EXITING is set a complete exit is accomplished when 769 * TDF_RUNNING and TDF_PREEMPT_LOCK are both clear. td_mpflags has two 770 * post-switch interlock flags that can be used to wait for the TDF_ 771 * flags to clear. 772 * 773 * Returns non-zero on success, and zero if the caller needs to retry 774 * the lwp_wait(). 775 */ 776 static int 777 lwp_wait(struct lwp *lp) 778 { 779 struct thread *td = lp->lwp_thread; 780 u_int mpflags; 781 782 KKASSERT(lwkt_preempted_proc() != lp); 783 784 /* 785 * This bit of code uses the thread destruction interlock 786 * managed by lwkt_switch_return() to wait for the lwp's 787 * thread to completely disengage. 788 * 789 * It is possible for us to race another cpu core so we 790 * have to do this correctly. 791 */ 792 for (;;) { 793 mpflags = td->td_mpflags; 794 cpu_ccfence(); 795 if (mpflags & TDF_MP_EXITSIG) 796 break; 797 tsleep_interlock(td, 0); 798 if (atomic_cmpset_int(&td->td_mpflags, mpflags, 799 mpflags | TDF_MP_EXITWAIT)) { 800 tsleep(td, PINTERLOCKED, "lwpxt", 0); 801 } 802 } 803 804 /* 805 * We've already waited for the core exit but there can still 806 * be other refs from e.g. process scans and such. 807 */ 808 if (lp->lwp_lock > 0) { 809 tsleep(lp, 0, "lwpwait1", 1); 810 return(0); 811 } 812 if (td->td_refs) { 813 tsleep(td, 0, "lwpwait2", 1); 814 return(0); 815 } 816 817 /* 818 * Now that we have the thread destruction interlock these flags 819 * really should already be cleaned up, keep a check for safety. 820 * 821 * We can't rip its stack out from under it until TDF_EXITING is 822 * set and both TDF_RUNNING and TDF_PREEMPT_LOCK are clear. 823 * TDF_PREEMPT_LOCK must be checked because TDF_RUNNING 824 * will be cleared temporarily if a thread gets preempted. 825 */ 826 while ((td->td_flags & (TDF_RUNNING | 827 TDF_RUNQ | 828 TDF_PREEMPT_LOCK | 829 TDF_EXITING)) != TDF_EXITING) { 830 tsleep(lp, 0, "lwpwait3", 1); 831 return (0); 832 } 833 834 KASSERT((td->td_flags & (TDF_RUNQ|TDF_TSLEEPQ)) == 0, 835 ("lwp_wait: td %p (%s) still on run or sleep queue", 836 td, td->td_comm)); 837 return (1); 838 } 839 840 /* 841 * Release the resources associated with a lwp. 842 * The lwp must be completely dead. 843 */ 844 void 845 lwp_dispose(struct lwp *lp) 846 { 847 struct thread *td = lp->lwp_thread; 848 849 KKASSERT(lwkt_preempted_proc() != lp); 850 KKASSERT(lp->lwp_lock == 0); 851 KKASSERT(td->td_refs == 0); 852 KKASSERT((td->td_flags & (TDF_RUNNING | 853 TDF_RUNQ | 854 TDF_PREEMPT_LOCK | 855 TDF_EXITING)) == TDF_EXITING); 856 857 PRELE(lp->lwp_proc); 858 lp->lwp_proc = NULL; 859 if (td != NULL) { 860 td->td_proc = NULL; 861 td->td_lwp = NULL; 862 lp->lwp_thread = NULL; 863 lwkt_free_thread(td); 864 } 865 kfree(lp, M_LWP); 866 } 867 868 int 869 sys_wait4(struct wait_args *uap) 870 { 871 struct rusage rusage; 872 int error, status; 873 874 error = kern_wait(uap->pid, (uap->status ? &status : NULL), 875 uap->options, (uap->rusage ? &rusage : NULL), 876 &uap->sysmsg_result); 877 878 if (error == 0 && uap->status) 879 error = copyout(&status, uap->status, sizeof(*uap->status)); 880 if (error == 0 && uap->rusage) 881 error = copyout(&rusage, uap->rusage, sizeof(*uap->rusage)); 882 return (error); 883 } 884 885 /* 886 * wait1() 887 * 888 * wait_args(int pid, int *status, int options, struct rusage *rusage) 889 */ 890 int 891 kern_wait(pid_t pid, int *status, int options, struct rusage *rusage, int *res) 892 { 893 struct thread *td = curthread; 894 struct lwp *lp; 895 struct proc *q = td->td_proc; 896 struct proc *p, *t; 897 struct pargs *pa; 898 struct sigacts *ps; 899 int nfound, error; 900 long waitgen; 901 902 if (pid == 0) 903 pid = -q->p_pgid; 904 if (options &~ (WUNTRACED|WNOHANG|WCONTINUED|WLINUXCLONE)) 905 return (EINVAL); 906 907 /* 908 * Protect the q->p_children list 909 */ 910 lwkt_gettoken(&q->p_token); 911 loop: 912 /* 913 * All sorts of things can change due to blocking so we have to loop 914 * all the way back up here. 915 * 916 * The problem is that if a process group is stopped and the parent 917 * is doing a wait*(..., WUNTRACED, ...), it will see the STOP 918 * of the child and then stop itself when it tries to return from the 919 * system call. When the process group is resumed the parent will 920 * then get the STOP status even though the child has now resumed 921 * (a followup wait*() will get the CONT status). 922 * 923 * Previously the CONT would overwrite the STOP because the tstop 924 * was handled within tsleep(), and the parent would only see 925 * the CONT when both are stopped and continued together. This little 926 * two-line hack restores this effect. 927 */ 928 if (STOPLWP(q, td->td_lwp)) 929 tstop(); 930 931 nfound = 0; 932 933 /* 934 * Loop on children. 935 * 936 * NOTE: We don't want to break q's p_token in the loop for the 937 * case where no children are found or we risk breaking the 938 * interlock between child and parent. 939 */ 940 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 941 LIST_FOREACH(p, &q->p_children, p_sibling) { 942 if (pid != WAIT_ANY && 943 p->p_pid != pid && p->p_pgid != -pid) { 944 continue; 945 } 946 947 /* 948 * This special case handles a kthread spawned by linux_clone 949 * (see linux_misc.c). The linux_wait4 and linux_waitpid 950 * functions need to be able to distinguish between waiting 951 * on a process and waiting on a thread. It is a thread if 952 * p_sigparent is not SIGCHLD, and the WLINUXCLONE option 953 * signifies we want to wait for threads and not processes. 954 */ 955 if ((p->p_sigparent != SIGCHLD) ^ 956 ((options & WLINUXCLONE) != 0)) { 957 continue; 958 } 959 960 nfound++; 961 if (p->p_stat == SZOMB) { 962 /* 963 * We may go into SZOMB with threads still present. 964 * We must wait for them to exit before we can reap 965 * the master thread, otherwise we may race reaping 966 * non-master threads. 967 * 968 * Only this routine can remove a process from 969 * the zombie list and destroy it, use PACQUIREZOMB() 970 * to serialize us and loop if it blocks (interlocked 971 * by the parent's q->p_token). 972 * 973 * WARNING! (p) can be invalid when PHOLDZOMB(p) 974 * returns non-zero. Be sure not to 975 * mess with it. 976 */ 977 if (PHOLDZOMB(p)) 978 goto loop; 979 lwkt_gettoken(&p->p_token); 980 if (p->p_pptr != q) { 981 lwkt_reltoken(&p->p_token); 982 PRELEZOMB(p); 983 goto loop; 984 } 985 while (p->p_nthreads > 0) { 986 tsleep(&p->p_nthreads, 0, "lwpzomb", hz); 987 } 988 989 /* 990 * Reap any LWPs left in p->p_lwps. This is usually 991 * just the last LWP. This must be done before 992 * we loop on p_lock since the lwps hold a ref on 993 * it as a vmspace interlock. 994 * 995 * Once that is accomplished p_nthreads had better 996 * be zero. 997 */ 998 while ((lp = RB_ROOT(&p->p_lwp_tree)) != NULL) { 999 /* 1000 * Make sure no one is using this lwp, before 1001 * it is removed from the tree. If we didn't 1002 * wait it here, lwp tree iteration with 1003 * blocking operation would be broken. 1004 */ 1005 while (lp->lwp_lock > 0) 1006 tsleep(lp, 0, "zomblwp", 1); 1007 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 1008 reaplwp(lp); 1009 } 1010 KKASSERT(p->p_nthreads == 0); 1011 1012 /* 1013 * Don't do anything really bad until all references 1014 * to the process go away. This may include other 1015 * LWPs which are still in the process of being 1016 * reaped. We can't just pull the rug out from under 1017 * them because they may still be using the VM space. 1018 * 1019 * Certain kernel facilities such as /proc will also 1020 * put a hold on the process for short periods of 1021 * time. 1022 */ 1023 PRELE(p); 1024 PSTALL(p, "reap3", 0); 1025 1026 /* Take care of our return values. */ 1027 *res = p->p_pid; 1028 1029 if (status) 1030 *status = p->p_xstat; 1031 if (rusage) 1032 *rusage = p->p_ru; 1033 1034 /* 1035 * If we got the child via a ptrace 'attach', 1036 * we need to give it back to the old parent. 1037 */ 1038 if (p->p_oppid && (t = pfind(p->p_oppid)) != NULL) { 1039 PHOLD(p); 1040 p->p_oppid = 0; 1041 proc_reparent(p, t); 1042 ksignal(t, SIGCHLD); 1043 wakeup((caddr_t)t); 1044 error = 0; 1045 PRELE(t); 1046 lwkt_reltoken(&p->p_token); 1047 PRELEZOMB(p); 1048 goto done; 1049 } 1050 1051 /* 1052 * Unlink the proc from its process group so that 1053 * the following operations won't lead to an 1054 * inconsistent state for processes running down 1055 * the zombie list. 1056 */ 1057 proc_remove_zombie(p); 1058 proc_userunmap(p); 1059 lwkt_reltoken(&p->p_token); 1060 leavepgrp(p); 1061 1062 p->p_xstat = 0; 1063 ruadd(&q->p_cru, &p->p_ru); 1064 1065 /* 1066 * Decrement the count of procs running with this uid. 1067 */ 1068 chgproccnt(p->p_ucred->cr_ruidinfo, -1, 0); 1069 1070 /* 1071 * Free up credentials. 1072 */ 1073 crfree(p->p_ucred); 1074 p->p_ucred = NULL; 1075 1076 /* 1077 * Remove unused arguments 1078 */ 1079 pa = p->p_args; 1080 p->p_args = NULL; 1081 if (pa && refcount_release(&pa->ar_ref)) { 1082 kfree(pa, M_PARGS); 1083 pa = NULL; 1084 } 1085 1086 ps = p->p_sigacts; 1087 p->p_sigacts = NULL; 1088 if (ps && refcount_release(&ps->ps_refcnt)) { 1089 kfree(ps, M_SUBPROC); 1090 ps = NULL; 1091 } 1092 1093 /* 1094 * Our exitingcount was incremented when the process 1095 * became a zombie, now that the process has been 1096 * removed from (almost) all lists we should be able 1097 * to safely destroy its vmspace. Wait for any current 1098 * holders to go away (so the vmspace remains stable), 1099 * then scrap it. 1100 * 1101 * NOTE: Releasing the parent process (q) p_token 1102 * across the vmspace_exitfree() call is 1103 * important here to reduce stalls on 1104 * interactions with (q) (such as 1105 * fork/exec/wait or 'ps'). 1106 */ 1107 PSTALL(p, "reap4", 0); 1108 lwkt_reltoken(&q->p_token); 1109 vmspace_exitfree(p); 1110 lwkt_gettoken(&q->p_token); 1111 PSTALL(p, "reap5", 0); 1112 1113 /* 1114 * NOTE: We have to officially release ZOMB in order 1115 * to ensure that a racing thread in kern_wait() 1116 * which blocked on ZOMB is woken up. 1117 */ 1118 PHOLD(p); 1119 PRELEZOMB(p); 1120 kfree(p, M_PROC); 1121 atomic_add_int(&nprocs, -1); 1122 error = 0; 1123 goto done; 1124 } 1125 if ((p->p_stat == SSTOP || p->p_stat == SCORE) && 1126 (p->p_flags & P_WAITED) == 0 && 1127 ((p->p_flags & P_TRACED) || (options & WUNTRACED))) { 1128 PHOLD(p); 1129 lwkt_gettoken(&p->p_token); 1130 if (p->p_pptr != q) { 1131 lwkt_reltoken(&p->p_token); 1132 PRELE(p); 1133 goto loop; 1134 } 1135 if ((p->p_stat != SSTOP && p->p_stat != SCORE) || 1136 (p->p_flags & P_WAITED) != 0 || 1137 ((p->p_flags & P_TRACED) == 0 && 1138 (options & WUNTRACED) == 0)) { 1139 lwkt_reltoken(&p->p_token); 1140 PRELE(p); 1141 goto loop; 1142 } 1143 1144 p->p_flags |= P_WAITED; 1145 1146 *res = p->p_pid; 1147 if (status) 1148 *status = W_STOPCODE(p->p_xstat); 1149 /* Zero rusage so we get something consistent. */ 1150 if (rusage) 1151 bzero(rusage, sizeof(*rusage)); 1152 error = 0; 1153 lwkt_reltoken(&p->p_token); 1154 PRELE(p); 1155 goto done; 1156 } 1157 if ((options & WCONTINUED) && (p->p_flags & P_CONTINUED)) { 1158 PHOLD(p); 1159 lwkt_gettoken(&p->p_token); 1160 if (p->p_pptr != q) { 1161 lwkt_reltoken(&p->p_token); 1162 PRELE(p); 1163 goto loop; 1164 } 1165 if ((p->p_flags & P_CONTINUED) == 0) { 1166 lwkt_reltoken(&p->p_token); 1167 PRELE(p); 1168 goto loop; 1169 } 1170 1171 *res = p->p_pid; 1172 p->p_flags &= ~P_CONTINUED; 1173 1174 if (status) 1175 *status = SIGCONT; 1176 error = 0; 1177 lwkt_reltoken(&p->p_token); 1178 PRELE(p); 1179 goto done; 1180 } 1181 } 1182 if (nfound == 0) { 1183 error = ECHILD; 1184 goto done; 1185 } 1186 if (options & WNOHANG) { 1187 *res = 0; 1188 error = 0; 1189 goto done; 1190 } 1191 1192 /* 1193 * Wait for signal - interlocked using q->p_waitgen. 1194 */ 1195 error = 0; 1196 while ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1197 tsleep_interlock(q, PCATCH); 1198 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 1199 if ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1200 error = tsleep(q, PCATCH | PINTERLOCKED, "wait", 0); 1201 break; 1202 } 1203 } 1204 if (error) { 1205 done: 1206 lwkt_reltoken(&q->p_token); 1207 return (error); 1208 } 1209 goto loop; 1210 } 1211 1212 /* 1213 * Change child's parent process to parent. 1214 * 1215 * p_children/p_sibling requires the parent's token, and 1216 * changing pptr requires the child's token, so we have to 1217 * get three tokens to do this operation. We also need to 1218 * hold pointers that might get ripped out from under us to 1219 * preserve structural integrity. 1220 * 1221 * It is possible to race another reparent or disconnect or other 1222 * similar operation. We must retry when this situation occurs. 1223 * Once we successfully reparent the process we no longer care 1224 * about any races. 1225 */ 1226 void 1227 proc_reparent(struct proc *child, struct proc *parent) 1228 { 1229 struct proc *opp; 1230 1231 PHOLD(parent); 1232 while ((opp = child->p_pptr) != parent) { 1233 PHOLD(opp); 1234 lwkt_gettoken(&opp->p_token); 1235 lwkt_gettoken(&child->p_token); 1236 lwkt_gettoken(&parent->p_token); 1237 if (child->p_pptr != opp) { 1238 lwkt_reltoken(&parent->p_token); 1239 lwkt_reltoken(&child->p_token); 1240 lwkt_reltoken(&opp->p_token); 1241 PRELE(opp); 1242 continue; 1243 } 1244 LIST_REMOVE(child, p_sibling); 1245 LIST_INSERT_HEAD(&parent->p_children, child, p_sibling); 1246 child->p_pptr = parent; 1247 lwkt_reltoken(&parent->p_token); 1248 lwkt_reltoken(&child->p_token); 1249 lwkt_reltoken(&opp->p_token); 1250 if (LIST_EMPTY(&opp->p_children)) 1251 wakeup(opp); 1252 PRELE(opp); 1253 break; 1254 } 1255 PRELE(parent); 1256 } 1257 1258 /* 1259 * The next two functions are to handle adding/deleting items on the 1260 * exit callout list 1261 * 1262 * at_exit(): 1263 * Take the arguments given and put them onto the exit callout list, 1264 * However first make sure that it's not already there. 1265 * returns 0 on success. 1266 */ 1267 1268 int 1269 at_exit(exitlist_fn function) 1270 { 1271 struct exitlist *ep; 1272 1273 #ifdef INVARIANTS 1274 /* Be noisy if the programmer has lost track of things */ 1275 if (rm_at_exit(function)) 1276 kprintf("WARNING: exit callout entry (%p) already present\n", 1277 function); 1278 #endif 1279 ep = kmalloc(sizeof(*ep), M_ATEXIT, M_NOWAIT); 1280 if (ep == NULL) 1281 return (ENOMEM); 1282 ep->function = function; 1283 TAILQ_INSERT_TAIL(&exit_list, ep, next); 1284 return (0); 1285 } 1286 1287 /* 1288 * Scan the exit callout list for the given item and remove it. 1289 * Returns the number of items removed (0 or 1) 1290 */ 1291 int 1292 rm_at_exit(exitlist_fn function) 1293 { 1294 struct exitlist *ep; 1295 1296 TAILQ_FOREACH(ep, &exit_list, next) { 1297 if (ep->function == function) { 1298 TAILQ_REMOVE(&exit_list, ep, next); 1299 kfree(ep, M_ATEXIT); 1300 return(1); 1301 } 1302 } 1303 return (0); 1304 } 1305 1306 /* 1307 * LWP reaper related code. 1308 */ 1309 static void 1310 reaplwps(void *context, int dummy) 1311 { 1312 struct lwplist *lwplist = context; 1313 struct lwp *lp; 1314 int cpu = mycpuid; 1315 1316 lwkt_gettoken(&deadlwp_token[cpu]); 1317 while ((lp = LIST_FIRST(lwplist))) { 1318 LIST_REMOVE(lp, u.lwp_reap_entry); 1319 reaplwp(lp); 1320 } 1321 lwkt_reltoken(&deadlwp_token[cpu]); 1322 } 1323 1324 static void 1325 reaplwp(struct lwp *lp) 1326 { 1327 while (lwp_wait(lp) == 0) 1328 ; 1329 lwp_dispose(lp); 1330 } 1331 1332 static void 1333 deadlwp_init(void) 1334 { 1335 int cpu; 1336 1337 for (cpu = 0; cpu < ncpus; cpu++) { 1338 lwkt_token_init(&deadlwp_token[cpu], "deadlwpl"); 1339 LIST_INIT(&deadlwp_list[cpu]); 1340 deadlwp_task[cpu] = kmalloc(sizeof(*deadlwp_task[cpu]), 1341 M_DEVBUF, M_WAITOK); 1342 TASK_INIT(deadlwp_task[cpu], 0, reaplwps, &deadlwp_list[cpu]); 1343 } 1344 } 1345 1346 SYSINIT(deadlwpinit, SI_SUB_CONFIGURE, SI_ORDER_ANY, deadlwp_init, NULL); 1347