1 /* 2 * Copyright (c) 1982, 1986, 1989, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)kern_exit.c 8.7 (Berkeley) 2/12/94 35 * $FreeBSD: src/sys/kern/kern_exit.c,v 1.92.2.11 2003/01/13 22:51:16 dillon Exp $ 36 */ 37 38 #include "opt_ktrace.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/sysproto.h> 43 #include <sys/kernel.h> 44 #include <sys/malloc.h> 45 #include <sys/proc.h> 46 #include <sys/ktrace.h> 47 #include <sys/pioctl.h> 48 #include <sys/tty.h> 49 #include <sys/wait.h> 50 #include <sys/vnode.h> 51 #include <sys/resourcevar.h> 52 #include <sys/signalvar.h> 53 #include <sys/taskqueue.h> 54 #include <sys/ptrace.h> 55 #include <sys/acct.h> /* for acct_process() function prototype */ 56 #include <sys/filedesc.h> 57 #include <sys/shm.h> 58 #include <sys/sem.h> 59 #include <sys/jail.h> 60 #include <sys/kern_syscall.h> 61 #include <sys/unistd.h> 62 #include <sys/eventhandler.h> 63 #include <sys/dsched.h> 64 65 #include <vm/vm.h> 66 #include <vm/vm_param.h> 67 #include <sys/lock.h> 68 #include <vm/pmap.h> 69 #include <vm/vm_map.h> 70 #include <vm/vm_extern.h> 71 #include <sys/user.h> 72 73 #include <sys/refcount.h> 74 #include <sys/thread2.h> 75 #include <sys/spinlock2.h> 76 #include <sys/mplock2.h> 77 78 #include <machine/vmm.h> 79 80 static void reaplwps(void *context, int dummy); 81 static void reaplwp(struct lwp *lp); 82 static void killlwps(struct lwp *lp); 83 84 static MALLOC_DEFINE(M_ATEXIT, "atexit", "atexit callback"); 85 86 /* 87 * callout list for things to do at exit time 88 */ 89 struct exitlist { 90 exitlist_fn function; 91 TAILQ_ENTRY(exitlist) next; 92 }; 93 94 TAILQ_HEAD(exit_list_head, exitlist); 95 static struct exit_list_head exit_list = TAILQ_HEAD_INITIALIZER(exit_list); 96 97 /* 98 * LWP reaper data 99 */ 100 static struct task *deadlwp_task[MAXCPU]; 101 static struct lwplist deadlwp_list[MAXCPU]; 102 static struct lwkt_token deadlwp_token[MAXCPU]; 103 104 /* 105 * exit -- 106 * Death of process. 107 * 108 * SYS_EXIT_ARGS(int rval) 109 */ 110 int 111 sys_exit(struct exit_args *uap) 112 { 113 exit1(W_EXITCODE(uap->rval, 0)); 114 /* NOTREACHED */ 115 } 116 117 /* 118 * Extended exit -- 119 * Death of a lwp or process with optional bells and whistles. 120 */ 121 int 122 sys_extexit(struct extexit_args *uap) 123 { 124 struct proc *p = curproc; 125 int action, who; 126 int error; 127 128 action = EXTEXIT_ACTION(uap->how); 129 who = EXTEXIT_WHO(uap->how); 130 131 /* Check parameters before we might perform some action */ 132 switch (who) { 133 case EXTEXIT_PROC: 134 case EXTEXIT_LWP: 135 break; 136 default: 137 return (EINVAL); 138 } 139 140 switch (action) { 141 case EXTEXIT_SIMPLE: 142 break; 143 case EXTEXIT_SETINT: 144 error = copyout(&uap->status, uap->addr, sizeof(uap->status)); 145 if (error) 146 return (error); 147 break; 148 default: 149 return (EINVAL); 150 } 151 152 lwkt_gettoken(&p->p_token); 153 154 switch (who) { 155 case EXTEXIT_LWP: 156 /* 157 * Be sure only to perform a simple lwp exit if there is at 158 * least one more lwp in the proc, which will call exit1() 159 * later, otherwise the proc will be an UNDEAD and not even a 160 * SZOMB! 161 */ 162 if (p->p_nthreads > 1) { 163 lwp_exit(0, NULL); /* called w/ p_token held */ 164 /* NOT REACHED */ 165 } 166 /* else last lwp in proc: do the real thing */ 167 /* FALLTHROUGH */ 168 default: /* to help gcc */ 169 case EXTEXIT_PROC: 170 lwkt_reltoken(&p->p_token); 171 exit1(W_EXITCODE(uap->status, 0)); 172 /* NOTREACHED */ 173 } 174 175 /* NOTREACHED */ 176 lwkt_reltoken(&p->p_token); /* safety */ 177 } 178 179 /* 180 * Kill all lwps associated with the current process except the 181 * current lwp. Return an error if we race another thread trying to 182 * do the same thing and lose the race. 183 * 184 * If forexec is non-zero the current thread and process flags are 185 * cleaned up so they can be reused. 186 * 187 * Caller must hold curproc->p_token 188 */ 189 int 190 killalllwps(int forexec) 191 { 192 struct lwp *lp = curthread->td_lwp; 193 struct proc *p = lp->lwp_proc; 194 int fakestop; 195 196 /* 197 * Interlock against P_WEXIT. Only one of the process's thread 198 * is allowed to do the master exit. 199 */ 200 if (p->p_flags & P_WEXIT) 201 return (EALREADY); 202 p->p_flags |= P_WEXIT; 203 204 /* 205 * Set temporary stopped state in case we are racing a coredump. 206 * Otherwise the coredump may hang forever. 207 */ 208 if (lp->lwp_mpflags & LWP_MP_WSTOP) { 209 fakestop = 0; 210 } else { 211 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 212 ++p->p_nstopped; 213 fakestop = 1; 214 wakeup(&p->p_nstopped); 215 } 216 217 /* 218 * Interlock with LWP_MP_WEXIT and kill any remaining LWPs 219 */ 220 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 221 if (p->p_nthreads > 1) 222 killlwps(lp); 223 224 /* 225 * Undo temporary stopped state 226 */ 227 if (fakestop) { 228 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WSTOP); 229 --p->p_nstopped; 230 } 231 232 /* 233 * If doing this for an exec, clean up the remaining thread 234 * (us) for continuing operation after all the other threads 235 * have been killed. 236 */ 237 if (forexec) { 238 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 239 p->p_flags &= ~P_WEXIT; 240 } 241 return(0); 242 } 243 244 /* 245 * Kill all LWPs except the current one. Do not try to signal 246 * LWPs which have exited on their own or have already been 247 * signaled. 248 */ 249 static void 250 killlwps(struct lwp *lp) 251 { 252 struct proc *p = lp->lwp_proc; 253 struct lwp *tlp; 254 255 /* 256 * Kill the remaining LWPs. We must send the signal before setting 257 * LWP_MP_WEXIT. The setting of WEXIT is optional but helps reduce 258 * races. tlp must be held across the call as it might block and 259 * allow the target lwp to rip itself out from under our loop. 260 */ 261 FOREACH_LWP_IN_PROC(tlp, p) { 262 LWPHOLD(tlp); 263 lwkt_gettoken(&tlp->lwp_token); 264 if ((tlp->lwp_mpflags & LWP_MP_WEXIT) == 0) { 265 atomic_set_int(&tlp->lwp_mpflags, LWP_MP_WEXIT); 266 lwpsignal(p, tlp, SIGKILL); 267 } 268 lwkt_reltoken(&tlp->lwp_token); 269 LWPRELE(tlp); 270 } 271 272 /* 273 * Wait for everything to clear out. Also make sure any tstop()s 274 * are signalled (we are holding p_token for the interlock). 275 */ 276 wakeup(p); 277 while (p->p_nthreads > 1) 278 tsleep(&p->p_nthreads, 0, "killlwps", 0); 279 } 280 281 /* 282 * Exit: deallocate address space and other resources, change proc state 283 * to zombie, and unlink proc from allproc and parent's lists. Save exit 284 * status and rusage for wait(). Check for child processes and orphan them. 285 */ 286 void 287 exit1(int rv) 288 { 289 struct thread *td = curthread; 290 struct proc *p = td->td_proc; 291 struct lwp *lp = td->td_lwp; 292 struct proc *q; 293 struct proc *pp; 294 struct proc *reproc; 295 struct sysreaper *reap; 296 struct vmspace *vm; 297 struct vnode *vtmp; 298 struct exitlist *ep; 299 int error; 300 301 lwkt_gettoken(&p->p_token); 302 303 if (p->p_pid == 1) { 304 kprintf("init died (signal %d, exit %d)\n", 305 WTERMSIG(rv), WEXITSTATUS(rv)); 306 panic("Going nowhere without my init!"); 307 } 308 varsymset_clean(&p->p_varsymset); 309 lockuninit(&p->p_varsymset.vx_lock); 310 311 /* 312 * Kill all lwps associated with the current process, return an 313 * error if we race another thread trying to do the same thing 314 * and lose the race. 315 */ 316 error = killalllwps(0); 317 if (error) { 318 lwp_exit(0, NULL); 319 /* NOT REACHED */ 320 } 321 322 /* are we a task leader? */ 323 if (p == p->p_leader) { 324 struct kill_args killArgs; 325 killArgs.signum = SIGKILL; 326 q = p->p_peers; 327 while(q) { 328 killArgs.pid = q->p_pid; 329 /* 330 * The interface for kill is better 331 * than the internal signal 332 */ 333 sys_kill(&killArgs); 334 q = q->p_peers; 335 } 336 while (p->p_peers) 337 tsleep((caddr_t)p, 0, "exit1", 0); 338 } 339 340 #ifdef PGINPROF 341 vmsizmon(); 342 #endif 343 STOPEVENT(p, S_EXIT, rv); 344 p->p_flags |= P_POSTEXIT; /* stop procfs stepping */ 345 346 /* 347 * Check if any loadable modules need anything done at process exit. 348 * e.g. SYSV IPC stuff 349 * XXX what if one of these generates an error? 350 */ 351 p->p_xstat = rv; 352 353 /* 354 * XXX: imho, the eventhandler stuff is much cleaner than this. 355 * Maybe we should move everything to use eventhandler. 356 */ 357 TAILQ_FOREACH(ep, &exit_list, next) 358 (*ep->function)(td); 359 360 if (p->p_flags & P_PROFIL) 361 stopprofclock(p); 362 363 SIGEMPTYSET(p->p_siglist); 364 SIGEMPTYSET(lp->lwp_siglist); 365 if (timevalisset(&p->p_realtimer.it_value)) 366 callout_stop_sync(&p->p_ithandle); 367 368 /* 369 * Reset any sigio structures pointing to us as a result of 370 * F_SETOWN with our pid. 371 */ 372 funsetownlst(&p->p_sigiolst); 373 374 /* 375 * Close open files and release open-file table. 376 * This may block! 377 */ 378 fdfree(p, NULL); 379 380 if (p->p_leader->p_peers) { 381 q = p->p_leader; 382 while(q->p_peers != p) 383 q = q->p_peers; 384 q->p_peers = p->p_peers; 385 wakeup((caddr_t)p->p_leader); 386 } 387 388 /* 389 * XXX Shutdown SYSV semaphores 390 */ 391 semexit(p); 392 393 /* The next two chunks should probably be moved to vmspace_exit. */ 394 vm = p->p_vmspace; 395 396 /* 397 * Clean up data related to virtual kernel operation. Clean up 398 * any vkernel context related to the current lwp now so we can 399 * destroy p_vkernel. 400 */ 401 if (p->p_vkernel) { 402 vkernel_lwp_exit(lp); 403 vkernel_exit(p); 404 } 405 406 /* 407 * Release the user portion of address space. The exitbump prevents 408 * the vmspace from being completely eradicated (using holdcnt). 409 * This releases references to vnodes, which could cause I/O if the 410 * file has been unlinked. We need to do this early enough that 411 * we can still sleep. 412 * 413 * We can't free the entire vmspace as the kernel stack may be mapped 414 * within that space also. 415 * 416 * Processes sharing the same vmspace may exit in one order, and 417 * get cleaned up by vmspace_exit() in a different order. The 418 * last exiting process to reach this point releases as much of 419 * the environment as it can, and the last process cleaned up 420 * by vmspace_exit() (which decrements exitingcnt) cleans up the 421 * remainder. 422 * 423 * NOTE: Releasing p_token around this call is helpful if the 424 * vmspace had a huge RSS. Otherwise some other process 425 * trying to do an allproc or other scan (like 'ps') may 426 * stall for a long time. 427 */ 428 lwkt_reltoken(&p->p_token); 429 vmspace_relexit(vm); 430 lwkt_gettoken(&p->p_token); 431 432 if (SESS_LEADER(p)) { 433 struct session *sp = p->p_session; 434 435 if (sp->s_ttyvp) { 436 /* 437 * We are the controlling process. Signal the 438 * foreground process group, drain the controlling 439 * terminal, and revoke access to the controlling 440 * terminal. 441 * 442 * NOTE: while waiting for the process group to exit 443 * it is possible that one of the processes in the 444 * group will revoke the tty, so the ttyclosesession() 445 * function will re-check sp->s_ttyvp. 446 */ 447 if (sp->s_ttyp && (sp->s_ttyp->t_session == sp)) { 448 if (sp->s_ttyp->t_pgrp) 449 pgsignal(sp->s_ttyp->t_pgrp, SIGHUP, 1); 450 ttywait(sp->s_ttyp); 451 ttyclosesession(sp, 1); /* also revoke */ 452 } 453 /* 454 * Release the tty. If someone has it open via 455 * /dev/tty then close it (since they no longer can 456 * once we've NULL'd it out). 457 */ 458 ttyclosesession(sp, 0); 459 460 /* 461 * s_ttyp is not zero'd; we use this to indicate 462 * that the session once had a controlling terminal. 463 * (for logging and informational purposes) 464 */ 465 } 466 sp->s_leader = NULL; 467 } 468 fixjobc(p, p->p_pgrp, 0); 469 (void)acct_process(p); 470 #ifdef KTRACE 471 /* 472 * release trace file 473 */ 474 if (p->p_tracenode) 475 ktrdestroy(&p->p_tracenode); 476 p->p_traceflag = 0; 477 #endif 478 /* 479 * Release reference to text vnode 480 */ 481 if ((vtmp = p->p_textvp) != NULL) { 482 p->p_textvp = NULL; 483 vrele(vtmp); 484 } 485 486 /* Release namecache handle to text file */ 487 if (p->p_textnch.ncp) 488 cache_drop(&p->p_textnch); 489 490 /* 491 * We have to handle PPWAIT here or proc_move_allproc_zombie() 492 * will block on the PHOLD() the parent is doing. 493 * 494 * We are using the flag as an interlock so an atomic op is 495 * necessary to synchronize with the parent's cpu. 496 */ 497 if (p->p_flags & P_PPWAIT) { 498 if (p->p_pptr && p->p_pptr->p_upmap) 499 atomic_add_int(&p->p_pptr->p_upmap->invfork, -1); 500 atomic_clear_int(&p->p_flags, P_PPWAIT); 501 wakeup(p->p_pptr); 502 } 503 504 /* 505 * Move the process to the zombie list. This will block 506 * until the process p_lock count reaches 0. The process will 507 * not be reaped until TDF_EXITING is set by cpu_thread_exit(), 508 * which is called from cpu_proc_exit(). 509 * 510 * Interlock against waiters using p_waitgen. We increment 511 * p_waitgen after completing the move of our process to the 512 * zombie list. 513 * 514 * WARNING: pp becomes stale when we block, clear it now as a 515 * reminder. 516 */ 517 proc_move_allproc_zombie(p); 518 pp = p->p_pptr; 519 atomic_add_long(&pp->p_waitgen, 1); 520 pp = NULL; 521 522 /* 523 * release controlled reaper for exit if we own it and return the 524 * remaining reaper (the one for us), which we will drop after we 525 * are done. 526 */ 527 reap = reaper_exit(p); 528 529 /* 530 * Reparent all of this process's children to the init process or 531 * to the designated reaper. We must hold the reaper's p_token in 532 * order to safely mess with p_children. 533 * 534 * We already hold p->p_token (to remove the children from our list). 535 */ 536 reproc = NULL; 537 q = LIST_FIRST(&p->p_children); 538 if (q) { 539 reproc = reaper_get(reap); 540 lwkt_gettoken(&reproc->p_token); 541 while ((q = LIST_FIRST(&p->p_children)) != NULL) { 542 PHOLD(q); 543 lwkt_gettoken(&q->p_token); 544 if (q != LIST_FIRST(&p->p_children)) { 545 lwkt_reltoken(&q->p_token); 546 PRELE(q); 547 continue; 548 } 549 LIST_REMOVE(q, p_sibling); 550 LIST_INSERT_HEAD(&reproc->p_children, q, p_sibling); 551 q->p_pptr = reproc; 552 q->p_ppid = reproc->p_pid; 553 q->p_sigparent = SIGCHLD; 554 555 /* 556 * Traced processes are killed 557 * since their existence means someone is screwing up. 558 */ 559 if (q->p_flags & P_TRACED) { 560 q->p_flags &= ~P_TRACED; 561 ksignal(q, SIGKILL); 562 } 563 lwkt_reltoken(&q->p_token); 564 PRELE(q); 565 } 566 lwkt_reltoken(&reproc->p_token); 567 wakeup(reproc); 568 } 569 570 /* 571 * Save exit status and final rusage info, adding in child rusage 572 * info and self times. 573 */ 574 calcru_proc(p, &p->p_ru); 575 ruadd(&p->p_ru, &p->p_cru); 576 577 /* 578 * notify interested parties of our demise. 579 */ 580 KNOTE(&p->p_klist, NOTE_EXIT); 581 582 /* 583 * Notify parent that we're gone. If parent has the PS_NOCLDWAIT 584 * flag set, or if the handler is set to SIG_IGN, notify the reaper 585 * instead (it will handle this situation). 586 * 587 * NOTE: The reaper can still be the parent process. 588 * 589 * (must reload pp) 590 */ 591 if (p->p_pptr->p_sigacts->ps_flag & (PS_NOCLDWAIT | PS_CLDSIGIGN)) { 592 if (reproc == NULL) 593 reproc = reaper_get(reap); 594 proc_reparent(p, reproc); 595 } 596 if (reproc) 597 PRELE(reproc); 598 if (reap) 599 reaper_drop(reap); 600 601 /* 602 * Signal (possibly new) parent. 603 */ 604 pp = p->p_pptr; 605 PHOLD(pp); 606 if (p->p_sigparent && pp != initproc) { 607 int sig = p->p_sigparent; 608 609 if (sig != SIGUSR1 && sig != SIGCHLD) 610 sig = SIGCHLD; 611 ksignal(pp, sig); 612 } else { 613 ksignal(pp, SIGCHLD); 614 } 615 p->p_flags &= ~P_TRACED; 616 PRELE(pp); 617 618 /* 619 * cpu_exit is responsible for clearing curproc, since 620 * it is heavily integrated with the thread/switching sequence. 621 * 622 * Other substructures are freed from wait(). 623 */ 624 if (p->p_limit) { 625 struct plimit *rlimit; 626 627 rlimit = p->p_limit; 628 p->p_limit = NULL; 629 plimit_free(rlimit); 630 } 631 632 /* 633 * Finally, call machine-dependent code to release as many of the 634 * lwp's resources as we can and halt execution of this thread. 635 * 636 * pp is a wild pointer now but still the correct wakeup() target. 637 * lwp_exit() only uses it to send the wakeup() signal to the likely 638 * parent. Any reparenting race that occurs will get a signal 639 * automatically and not be an issue. 640 */ 641 lwp_exit(1, pp); 642 } 643 644 /* 645 * Eventually called by every exiting LWP 646 * 647 * p->p_token must be held. mplock may be held and will be released. 648 */ 649 void 650 lwp_exit(int masterexit, void *waddr) 651 { 652 struct thread *td = curthread; 653 struct lwp *lp = td->td_lwp; 654 struct proc *p = lp->lwp_proc; 655 int dowake = 0; 656 657 /* 658 * Release the current user process designation on the process so 659 * the userland scheduler can work in someone else. 660 */ 661 p->p_usched->release_curproc(lp); 662 663 /* 664 * lwp_exit() may be called without setting LWP_MP_WEXIT, so 665 * make sure it is set here. 666 */ 667 ASSERT_LWKT_TOKEN_HELD(&p->p_token); 668 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 669 670 /* 671 * Clean up any virtualization 672 */ 673 if (lp->lwp_vkernel) 674 vkernel_lwp_exit(lp); 675 676 if (td->td_vmm) 677 vmm_vmdestroy(); 678 679 /* 680 * Clean up select/poll support 681 */ 682 kqueue_terminate(&lp->lwp_kqueue); 683 684 /* 685 * Clean up any syscall-cached ucred or rlimit. 686 */ 687 if (td->td_ucred) { 688 crfree(td->td_ucred); 689 td->td_ucred = NULL; 690 } 691 if (td->td_limit) { 692 struct plimit *rlimit; 693 694 rlimit = td->td_limit; 695 td->td_limit = NULL; 696 plimit_free(rlimit); 697 } 698 699 /* 700 * Cleanup any cached descriptors for this thread 701 */ 702 if (p->p_fd) 703 fexitcache(td); 704 705 /* 706 * Nobody actually wakes us when the lock 707 * count reaches zero, so just wait one tick. 708 */ 709 while (lp->lwp_lock > 0) 710 tsleep(lp, 0, "lwpexit", 1); 711 712 /* Hand down resource usage to our proc */ 713 ruadd(&p->p_ru, &lp->lwp_ru); 714 715 /* 716 * If we don't hold the process until the LWP is reaped wait*() 717 * may try to dispose of its vmspace before all the LWPs have 718 * actually terminated. 719 */ 720 PHOLD(p); 721 722 /* 723 * Do any remaining work that might block on us. We should be 724 * coded such that further blocking is ok after decrementing 725 * p_nthreads but don't take the chance. 726 */ 727 dsched_exit_thread(td); 728 biosched_done(curthread); 729 730 /* 731 * We have to use the reaper for all the LWPs except the one doing 732 * the master exit. The LWP doing the master exit can just be 733 * left on p_lwps and the process reaper will deal with it 734 * synchronously, which is much faster. 735 * 736 * Wakeup anyone waiting on p_nthreads to drop to 1 or 0. 737 * 738 * The process is left held until the reaper calls lwp_dispose() on 739 * the lp (after calling lwp_wait()). 740 */ 741 if (masterexit == 0) { 742 int cpu = mycpuid; 743 744 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 745 --p->p_nthreads; 746 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 747 dowake = 1; 748 lwkt_gettoken(&deadlwp_token[cpu]); 749 LIST_INSERT_HEAD(&deadlwp_list[cpu], lp, u.lwp_reap_entry); 750 taskqueue_enqueue(taskqueue_thread[cpu], deadlwp_task[cpu]); 751 lwkt_reltoken(&deadlwp_token[cpu]); 752 } else { 753 --p->p_nthreads; 754 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 755 dowake = 1; 756 } 757 758 /* 759 * We no longer need p_token. 760 * 761 * Tell the userland scheduler that we are going away 762 */ 763 lwkt_reltoken(&p->p_token); 764 p->p_usched->heuristic_exiting(lp, p); 765 766 /* 767 * Issue late wakeups after releasing our token to give us a chance 768 * to deschedule and switch away before another cpu in a wait*() 769 * reaps us. This is done as late as possible to reduce contention. 770 */ 771 if (dowake) 772 wakeup(&p->p_nthreads); 773 if (waddr) 774 wakeup(waddr); 775 776 cpu_lwp_exit(); 777 } 778 779 /* 780 * Wait until a lwp is completely dead. The final interlock in this drama 781 * is when TDF_EXITING is set in cpu_thread_exit() just before the final 782 * switchout. 783 * 784 * At the point TDF_EXITING is set a complete exit is accomplished when 785 * TDF_RUNNING and TDF_PREEMPT_LOCK are both clear. td_mpflags has two 786 * post-switch interlock flags that can be used to wait for the TDF_ 787 * flags to clear. 788 * 789 * Returns non-zero on success, and zero if the caller needs to retry 790 * the lwp_wait(). 791 */ 792 static int 793 lwp_wait(struct lwp *lp) 794 { 795 struct thread *td = lp->lwp_thread; 796 u_int mpflags; 797 798 KKASSERT(lwkt_preempted_proc() != lp); 799 800 /* 801 * This bit of code uses the thread destruction interlock 802 * managed by lwkt_switch_return() to wait for the lwp's 803 * thread to completely disengage. 804 * 805 * It is possible for us to race another cpu core so we 806 * have to do this correctly. 807 */ 808 for (;;) { 809 mpflags = td->td_mpflags; 810 cpu_ccfence(); 811 if (mpflags & TDF_MP_EXITSIG) 812 break; 813 tsleep_interlock(td, 0); 814 if (atomic_cmpset_int(&td->td_mpflags, mpflags, 815 mpflags | TDF_MP_EXITWAIT)) { 816 tsleep(td, PINTERLOCKED, "lwpxt", 0); 817 } 818 } 819 820 /* 821 * We've already waited for the core exit but there can still 822 * be other refs from e.g. process scans and such. 823 */ 824 if (lp->lwp_lock > 0) { 825 tsleep(lp, 0, "lwpwait1", 1); 826 return(0); 827 } 828 if (td->td_refs) { 829 tsleep(td, 0, "lwpwait2", 1); 830 return(0); 831 } 832 833 /* 834 * Now that we have the thread destruction interlock these flags 835 * really should already be cleaned up, keep a check for safety. 836 * 837 * We can't rip its stack out from under it until TDF_EXITING is 838 * set and both TDF_RUNNING and TDF_PREEMPT_LOCK are clear. 839 * TDF_PREEMPT_LOCK must be checked because TDF_RUNNING 840 * will be cleared temporarily if a thread gets preempted. 841 */ 842 while ((td->td_flags & (TDF_RUNNING | 843 TDF_RUNQ | 844 TDF_PREEMPT_LOCK | 845 TDF_EXITING)) != TDF_EXITING) { 846 tsleep(lp, 0, "lwpwait3", 1); 847 return (0); 848 } 849 850 KASSERT((td->td_flags & (TDF_RUNQ|TDF_TSLEEPQ)) == 0, 851 ("lwp_wait: td %p (%s) still on run or sleep queue", 852 td, td->td_comm)); 853 return (1); 854 } 855 856 /* 857 * Release the resources associated with a lwp. 858 * The lwp must be completely dead. 859 */ 860 void 861 lwp_dispose(struct lwp *lp) 862 { 863 struct thread *td = lp->lwp_thread; 864 865 KKASSERT(lwkt_preempted_proc() != lp); 866 KKASSERT(lp->lwp_lock == 0); 867 KKASSERT(td->td_refs == 0); 868 KKASSERT((td->td_flags & (TDF_RUNNING | 869 TDF_RUNQ | 870 TDF_PREEMPT_LOCK | 871 TDF_EXITING)) == TDF_EXITING); 872 873 PRELE(lp->lwp_proc); 874 lp->lwp_proc = NULL; 875 if (td != NULL) { 876 td->td_proc = NULL; 877 td->td_lwp = NULL; 878 lp->lwp_thread = NULL; 879 lwkt_free_thread(td); 880 } 881 kfree(lp, M_LWP); 882 } 883 884 int 885 sys_wait4(struct wait_args *uap) 886 { 887 struct rusage rusage; 888 int error, status; 889 890 error = kern_wait(uap->pid, (uap->status ? &status : NULL), 891 uap->options, (uap->rusage ? &rusage : NULL), 892 &uap->sysmsg_result); 893 894 if (error == 0 && uap->status) 895 error = copyout(&status, uap->status, sizeof(*uap->status)); 896 if (error == 0 && uap->rusage) 897 error = copyout(&rusage, uap->rusage, sizeof(*uap->rusage)); 898 return (error); 899 } 900 901 /* 902 * wait1() 903 * 904 * wait_args(int pid, int *status, int options, struct rusage *rusage) 905 */ 906 int 907 kern_wait(pid_t pid, int *status, int options, struct rusage *rusage, int *res) 908 { 909 struct thread *td = curthread; 910 struct lwp *lp; 911 struct proc *q = td->td_proc; 912 struct proc *p, *t; 913 struct ucred *cr; 914 struct pargs *pa; 915 struct sigacts *ps; 916 int nfound, error; 917 long waitgen; 918 919 if (pid == 0) 920 pid = -q->p_pgid; 921 if (options &~ (WUNTRACED|WNOHANG|WCONTINUED|WLINUXCLONE)) 922 return (EINVAL); 923 924 /* 925 * Protect the q->p_children list 926 */ 927 lwkt_gettoken(&q->p_token); 928 loop: 929 /* 930 * All sorts of things can change due to blocking so we have to loop 931 * all the way back up here. 932 * 933 * The problem is that if a process group is stopped and the parent 934 * is doing a wait*(..., WUNTRACED, ...), it will see the STOP 935 * of the child and then stop itself when it tries to return from the 936 * system call. When the process group is resumed the parent will 937 * then get the STOP status even though the child has now resumed 938 * (a followup wait*() will get the CONT status). 939 * 940 * Previously the CONT would overwrite the STOP because the tstop 941 * was handled within tsleep(), and the parent would only see 942 * the CONT when both are stopped and continued together. This little 943 * two-line hack restores this effect. 944 */ 945 if (STOPLWP(q, td->td_lwp)) 946 tstop(); 947 948 nfound = 0; 949 950 /* 951 * Loop on children. 952 * 953 * NOTE: We don't want to break q's p_token in the loop for the 954 * case where no children are found or we risk breaking the 955 * interlock between child and parent. 956 */ 957 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 958 LIST_FOREACH(p, &q->p_children, p_sibling) { 959 if (pid != WAIT_ANY && 960 p->p_pid != pid && p->p_pgid != -pid) { 961 continue; 962 } 963 964 /* 965 * This special case handles a kthread spawned by linux_clone 966 * (see linux_misc.c). The linux_wait4 and linux_waitpid 967 * functions need to be able to distinguish between waiting 968 * on a process and waiting on a thread. It is a thread if 969 * p_sigparent is not SIGCHLD, and the WLINUXCLONE option 970 * signifies we want to wait for threads and not processes. 971 */ 972 if ((p->p_sigparent != SIGCHLD) ^ 973 ((options & WLINUXCLONE) != 0)) { 974 continue; 975 } 976 977 nfound++; 978 if (p->p_stat == SZOMB) { 979 /* 980 * We may go into SZOMB with threads still present. 981 * We must wait for them to exit before we can reap 982 * the master thread, otherwise we may race reaping 983 * non-master threads. 984 * 985 * Only this routine can remove a process from 986 * the zombie list and destroy it, use PACQUIREZOMB() 987 * to serialize us and loop if it blocks (interlocked 988 * by the parent's q->p_token). 989 * 990 * WARNING! (p) can be invalid when PHOLDZOMB(p) 991 * returns non-zero. Be sure not to 992 * mess with it. 993 */ 994 if (PHOLDZOMB(p)) 995 goto loop; 996 lwkt_gettoken(&p->p_token); 997 if (p->p_pptr != q) { 998 lwkt_reltoken(&p->p_token); 999 PRELEZOMB(p); 1000 goto loop; 1001 } 1002 while (p->p_nthreads > 0) { 1003 tsleep(&p->p_nthreads, 0, "lwpzomb", hz); 1004 } 1005 1006 /* 1007 * Reap any LWPs left in p->p_lwps. This is usually 1008 * just the last LWP. This must be done before 1009 * we loop on p_lock since the lwps hold a ref on 1010 * it as a vmspace interlock. 1011 * 1012 * Once that is accomplished p_nthreads had better 1013 * be zero. 1014 */ 1015 while ((lp = RB_ROOT(&p->p_lwp_tree)) != NULL) { 1016 /* 1017 * Make sure no one is using this lwp, before 1018 * it is removed from the tree. If we didn't 1019 * wait it here, lwp tree iteration with 1020 * blocking operation would be broken. 1021 */ 1022 while (lp->lwp_lock > 0) 1023 tsleep(lp, 0, "zomblwp", 1); 1024 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 1025 reaplwp(lp); 1026 } 1027 KKASSERT(p->p_nthreads == 0); 1028 1029 /* 1030 * Don't do anything really bad until all references 1031 * to the process go away. This may include other 1032 * LWPs which are still in the process of being 1033 * reaped. We can't just pull the rug out from under 1034 * them because they may still be using the VM space. 1035 * 1036 * Certain kernel facilities such as /proc will also 1037 * put a hold on the process for short periods of 1038 * time. 1039 */ 1040 PRELE(p); 1041 PSTALL(p, "reap3", 0); 1042 1043 /* Take care of our return values. */ 1044 *res = p->p_pid; 1045 1046 if (status) 1047 *status = p->p_xstat; 1048 if (rusage) 1049 *rusage = p->p_ru; 1050 1051 /* 1052 * If we got the child via a ptrace 'attach', 1053 * we need to give it back to the old parent. 1054 */ 1055 if (p->p_oppid && (t = pfind(p->p_oppid)) != NULL) { 1056 PHOLD(p); 1057 p->p_oppid = 0; 1058 proc_reparent(p, t); 1059 ksignal(t, SIGCHLD); 1060 wakeup((caddr_t)t); 1061 error = 0; 1062 PRELE(t); 1063 lwkt_reltoken(&p->p_token); 1064 PRELEZOMB(p); 1065 goto done; 1066 } 1067 1068 /* 1069 * Unlink the proc from its process group so that 1070 * the following operations won't lead to an 1071 * inconsistent state for processes running down 1072 * the zombie list. 1073 */ 1074 proc_remove_zombie(p); 1075 proc_userunmap(p); 1076 lwkt_reltoken(&p->p_token); 1077 leavepgrp(p); 1078 1079 p->p_xstat = 0; 1080 ruadd(&q->p_cru, &p->p_ru); 1081 1082 /* 1083 * Decrement the count of procs running with this uid. 1084 */ 1085 chgproccnt(p->p_ucred->cr_ruidinfo, -1, 0); 1086 1087 /* 1088 * Free up credentials. p_spin is required to 1089 * avoid races against allproc scans. 1090 */ 1091 spin_lock(&p->p_spin); 1092 cr = p->p_ucred; 1093 p->p_ucred = NULL; 1094 spin_unlock(&p->p_spin); 1095 crfree(cr); 1096 1097 /* 1098 * Remove unused arguments 1099 */ 1100 pa = p->p_args; 1101 p->p_args = NULL; 1102 if (pa && refcount_release(&pa->ar_ref)) { 1103 kfree(pa, M_PARGS); 1104 pa = NULL; 1105 } 1106 1107 ps = p->p_sigacts; 1108 p->p_sigacts = NULL; 1109 if (ps && refcount_release(&ps->ps_refcnt)) { 1110 kfree(ps, M_SUBPROC); 1111 ps = NULL; 1112 } 1113 1114 /* 1115 * Our exitingcount was incremented when the process 1116 * became a zombie, now that the process has been 1117 * removed from (almost) all lists we should be able 1118 * to safely destroy its vmspace. Wait for any current 1119 * holders to go away (so the vmspace remains stable), 1120 * then scrap it. 1121 * 1122 * NOTE: Releasing the parent process (q) p_token 1123 * across the vmspace_exitfree() call is 1124 * important here to reduce stalls on 1125 * interactions with (q) (such as 1126 * fork/exec/wait or 'ps'). 1127 */ 1128 PSTALL(p, "reap4", 0); 1129 lwkt_reltoken(&q->p_token); 1130 vmspace_exitfree(p); 1131 lwkt_gettoken(&q->p_token); 1132 PSTALL(p, "reap5", 0); 1133 1134 /* 1135 * NOTE: We have to officially release ZOMB in order 1136 * to ensure that a racing thread in kern_wait() 1137 * which blocked on ZOMB is woken up. 1138 */ 1139 PHOLD(p); 1140 PRELEZOMB(p); 1141 kfree(p->p_uidpcpu, M_SUBPROC); 1142 kfree(p, M_PROC); 1143 atomic_add_int(&nprocs, -1); 1144 error = 0; 1145 goto done; 1146 } 1147 if ((p->p_stat == SSTOP || p->p_stat == SCORE) && 1148 (p->p_flags & P_WAITED) == 0 && 1149 ((p->p_flags & P_TRACED) || (options & WUNTRACED))) { 1150 PHOLD(p); 1151 lwkt_gettoken(&p->p_token); 1152 if (p->p_pptr != q) { 1153 lwkt_reltoken(&p->p_token); 1154 PRELE(p); 1155 goto loop; 1156 } 1157 if ((p->p_stat != SSTOP && p->p_stat != SCORE) || 1158 (p->p_flags & P_WAITED) != 0 || 1159 ((p->p_flags & P_TRACED) == 0 && 1160 (options & WUNTRACED) == 0)) { 1161 lwkt_reltoken(&p->p_token); 1162 PRELE(p); 1163 goto loop; 1164 } 1165 1166 p->p_flags |= P_WAITED; 1167 1168 *res = p->p_pid; 1169 if (status) 1170 *status = W_STOPCODE(p->p_xstat); 1171 /* Zero rusage so we get something consistent. */ 1172 if (rusage) 1173 bzero(rusage, sizeof(*rusage)); 1174 error = 0; 1175 lwkt_reltoken(&p->p_token); 1176 PRELE(p); 1177 goto done; 1178 } 1179 if ((options & WCONTINUED) && (p->p_flags & P_CONTINUED)) { 1180 PHOLD(p); 1181 lwkt_gettoken(&p->p_token); 1182 if (p->p_pptr != q) { 1183 lwkt_reltoken(&p->p_token); 1184 PRELE(p); 1185 goto loop; 1186 } 1187 if ((p->p_flags & P_CONTINUED) == 0) { 1188 lwkt_reltoken(&p->p_token); 1189 PRELE(p); 1190 goto loop; 1191 } 1192 1193 *res = p->p_pid; 1194 p->p_flags &= ~P_CONTINUED; 1195 1196 if (status) 1197 *status = SIGCONT; 1198 error = 0; 1199 lwkt_reltoken(&p->p_token); 1200 PRELE(p); 1201 goto done; 1202 } 1203 } 1204 if (nfound == 0) { 1205 error = ECHILD; 1206 goto done; 1207 } 1208 if (options & WNOHANG) { 1209 *res = 0; 1210 error = 0; 1211 goto done; 1212 } 1213 1214 /* 1215 * Wait for signal - interlocked using q->p_waitgen. 1216 */ 1217 error = 0; 1218 while ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1219 tsleep_interlock(q, PCATCH); 1220 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 1221 if ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1222 error = tsleep(q, PCATCH | PINTERLOCKED, "wait", 0); 1223 break; 1224 } 1225 } 1226 if (error) { 1227 done: 1228 lwkt_reltoken(&q->p_token); 1229 return (error); 1230 } 1231 goto loop; 1232 } 1233 1234 /* 1235 * Change child's parent process to parent. 1236 * 1237 * p_children/p_sibling requires the parent's token, and 1238 * changing pptr requires the child's token, so we have to 1239 * get three tokens to do this operation. We also need to 1240 * hold pointers that might get ripped out from under us to 1241 * preserve structural integrity. 1242 * 1243 * It is possible to race another reparent or disconnect or other 1244 * similar operation. We must retry when this situation occurs. 1245 * Once we successfully reparent the process we no longer care 1246 * about any races. 1247 */ 1248 void 1249 proc_reparent(struct proc *child, struct proc *parent) 1250 { 1251 struct proc *opp; 1252 1253 PHOLD(parent); 1254 while ((opp = child->p_pptr) != parent) { 1255 PHOLD(opp); 1256 lwkt_gettoken(&opp->p_token); 1257 lwkt_gettoken(&child->p_token); 1258 lwkt_gettoken(&parent->p_token); 1259 if (child->p_pptr != opp) { 1260 lwkt_reltoken(&parent->p_token); 1261 lwkt_reltoken(&child->p_token); 1262 lwkt_reltoken(&opp->p_token); 1263 PRELE(opp); 1264 continue; 1265 } 1266 LIST_REMOVE(child, p_sibling); 1267 LIST_INSERT_HEAD(&parent->p_children, child, p_sibling); 1268 child->p_pptr = parent; 1269 child->p_ppid = parent->p_pid; 1270 lwkt_reltoken(&parent->p_token); 1271 lwkt_reltoken(&child->p_token); 1272 lwkt_reltoken(&opp->p_token); 1273 if (LIST_EMPTY(&opp->p_children)) 1274 wakeup(opp); 1275 PRELE(opp); 1276 break; 1277 } 1278 PRELE(parent); 1279 } 1280 1281 /* 1282 * The next two functions are to handle adding/deleting items on the 1283 * exit callout list 1284 * 1285 * at_exit(): 1286 * Take the arguments given and put them onto the exit callout list, 1287 * However first make sure that it's not already there. 1288 * returns 0 on success. 1289 */ 1290 1291 int 1292 at_exit(exitlist_fn function) 1293 { 1294 struct exitlist *ep; 1295 1296 #ifdef INVARIANTS 1297 /* Be noisy if the programmer has lost track of things */ 1298 if (rm_at_exit(function)) 1299 kprintf("WARNING: exit callout entry (%p) already present\n", 1300 function); 1301 #endif 1302 ep = kmalloc(sizeof(*ep), M_ATEXIT, M_NOWAIT); 1303 if (ep == NULL) 1304 return (ENOMEM); 1305 ep->function = function; 1306 TAILQ_INSERT_TAIL(&exit_list, ep, next); 1307 return (0); 1308 } 1309 1310 /* 1311 * Scan the exit callout list for the given item and remove it. 1312 * Returns the number of items removed (0 or 1) 1313 */ 1314 int 1315 rm_at_exit(exitlist_fn function) 1316 { 1317 struct exitlist *ep; 1318 1319 TAILQ_FOREACH(ep, &exit_list, next) { 1320 if (ep->function == function) { 1321 TAILQ_REMOVE(&exit_list, ep, next); 1322 kfree(ep, M_ATEXIT); 1323 return(1); 1324 } 1325 } 1326 return (0); 1327 } 1328 1329 /* 1330 * LWP reaper related code. 1331 */ 1332 static void 1333 reaplwps(void *context, int dummy) 1334 { 1335 struct lwplist *lwplist = context; 1336 struct lwp *lp; 1337 int cpu = mycpuid; 1338 1339 lwkt_gettoken(&deadlwp_token[cpu]); 1340 while ((lp = LIST_FIRST(lwplist))) { 1341 LIST_REMOVE(lp, u.lwp_reap_entry); 1342 reaplwp(lp); 1343 } 1344 lwkt_reltoken(&deadlwp_token[cpu]); 1345 } 1346 1347 static void 1348 reaplwp(struct lwp *lp) 1349 { 1350 while (lwp_wait(lp) == 0) 1351 ; 1352 lwp_dispose(lp); 1353 } 1354 1355 static void 1356 deadlwp_init(void) 1357 { 1358 int cpu; 1359 1360 for (cpu = 0; cpu < ncpus; cpu++) { 1361 lwkt_token_init(&deadlwp_token[cpu], "deadlwpl"); 1362 LIST_INIT(&deadlwp_list[cpu]); 1363 deadlwp_task[cpu] = kmalloc(sizeof(*deadlwp_task[cpu]), 1364 M_DEVBUF, M_WAITOK); 1365 TASK_INIT(deadlwp_task[cpu], 0, reaplwps, &deadlwp_list[cpu]); 1366 } 1367 } 1368 1369 SYSINIT(deadlwpinit, SI_SUB_CONFIGURE, SI_ORDER_ANY, deadlwp_init, NULL); 1370