1 /* 2 * Copyright (c) 1982, 1986, 1989, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)kern_exit.c 8.7 (Berkeley) 2/12/94 35 * $FreeBSD: src/sys/kern/kern_exit.c,v 1.92.2.11 2003/01/13 22:51:16 dillon Exp $ 36 */ 37 38 #include "opt_compat.h" 39 #include "opt_ktrace.h" 40 41 #include <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/sysproto.h> 44 #include <sys/kernel.h> 45 #include <sys/malloc.h> 46 #include <sys/proc.h> 47 #include <sys/ktrace.h> 48 #include <sys/pioctl.h> 49 #include <sys/tty.h> 50 #include <sys/wait.h> 51 #include <sys/vnode.h> 52 #include <sys/resourcevar.h> 53 #include <sys/signalvar.h> 54 #include <sys/taskqueue.h> 55 #include <sys/ptrace.h> 56 #include <sys/acct.h> /* for acct_process() function prototype */ 57 #include <sys/filedesc.h> 58 #include <sys/shm.h> 59 #include <sys/sem.h> 60 #include <sys/jail.h> 61 #include <sys/kern_syscall.h> 62 #include <sys/unistd.h> 63 #include <sys/eventhandler.h> 64 #include <sys/dsched.h> 65 66 #include <vm/vm.h> 67 #include <vm/vm_param.h> 68 #include <sys/lock.h> 69 #include <vm/pmap.h> 70 #include <vm/vm_map.h> 71 #include <vm/vm_extern.h> 72 #include <sys/user.h> 73 74 #include <sys/refcount.h> 75 #include <sys/thread2.h> 76 #include <sys/sysref2.h> 77 #include <sys/mplock2.h> 78 79 #include <machine/vmm.h> 80 81 static void reaplwps(void *context, int dummy); 82 static void reaplwp(struct lwp *lp); 83 static void killlwps(struct lwp *lp); 84 85 static MALLOC_DEFINE(M_ATEXIT, "atexit", "atexit callback"); 86 87 /* 88 * callout list for things to do at exit time 89 */ 90 struct exitlist { 91 exitlist_fn function; 92 TAILQ_ENTRY(exitlist) next; 93 }; 94 95 TAILQ_HEAD(exit_list_head, exitlist); 96 static struct exit_list_head exit_list = TAILQ_HEAD_INITIALIZER(exit_list); 97 98 /* 99 * LWP reaper data 100 */ 101 static struct task *deadlwp_task[MAXCPU]; 102 static struct lwplist deadlwp_list[MAXCPU]; 103 static struct lwkt_token deadlwp_token[MAXCPU]; 104 105 /* 106 * exit -- 107 * Death of process. 108 * 109 * SYS_EXIT_ARGS(int rval) 110 */ 111 int 112 sys_exit(struct exit_args *uap) 113 { 114 exit1(W_EXITCODE(uap->rval, 0)); 115 /* NOTREACHED */ 116 } 117 118 /* 119 * Extended exit -- 120 * Death of a lwp or process with optional bells and whistles. 121 */ 122 int 123 sys_extexit(struct extexit_args *uap) 124 { 125 struct proc *p = curproc; 126 int action, who; 127 int error; 128 129 action = EXTEXIT_ACTION(uap->how); 130 who = EXTEXIT_WHO(uap->how); 131 132 /* Check parameters before we might perform some action */ 133 switch (who) { 134 case EXTEXIT_PROC: 135 case EXTEXIT_LWP: 136 break; 137 default: 138 return (EINVAL); 139 } 140 141 switch (action) { 142 case EXTEXIT_SIMPLE: 143 break; 144 case EXTEXIT_SETINT: 145 error = copyout(&uap->status, uap->addr, sizeof(uap->status)); 146 if (error) 147 return (error); 148 break; 149 default: 150 return (EINVAL); 151 } 152 153 lwkt_gettoken(&p->p_token); 154 155 switch (who) { 156 case EXTEXIT_LWP: 157 /* 158 * Be sure only to perform a simple lwp exit if there is at 159 * least one more lwp in the proc, which will call exit1() 160 * later, otherwise the proc will be an UNDEAD and not even a 161 * SZOMB! 162 */ 163 if (p->p_nthreads > 1) { 164 lwp_exit(0, NULL); /* called w/ p_token held */ 165 /* NOT REACHED */ 166 } 167 /* else last lwp in proc: do the real thing */ 168 /* FALLTHROUGH */ 169 default: /* to help gcc */ 170 case EXTEXIT_PROC: 171 lwkt_reltoken(&p->p_token); 172 exit1(W_EXITCODE(uap->status, 0)); 173 /* NOTREACHED */ 174 } 175 176 /* NOTREACHED */ 177 lwkt_reltoken(&p->p_token); /* safety */ 178 } 179 180 /* 181 * Kill all lwps associated with the current process except the 182 * current lwp. Return an error if we race another thread trying to 183 * do the same thing and lose the race. 184 * 185 * If forexec is non-zero the current thread and process flags are 186 * cleaned up so they can be reused. 187 * 188 * Caller must hold curproc->p_token 189 */ 190 int 191 killalllwps(int forexec) 192 { 193 struct lwp *lp = curthread->td_lwp; 194 struct proc *p = lp->lwp_proc; 195 196 /* 197 * Interlock against P_WEXIT. Only one of the process's thread 198 * is allowed to do the master exit. 199 */ 200 if (p->p_flags & P_WEXIT) 201 return (EALREADY); 202 p->p_flags |= P_WEXIT; 203 204 /* 205 * Interlock with LWP_MP_WEXIT and kill any remaining LWPs 206 */ 207 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 208 if (p->p_nthreads > 1) 209 killlwps(lp); 210 211 /* 212 * If doing this for an exec, clean up the remaining thread 213 * (us) for continuing operation after all the other threads 214 * have been killed. 215 */ 216 if (forexec) { 217 atomic_clear_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 218 p->p_flags &= ~P_WEXIT; 219 } 220 return(0); 221 } 222 223 /* 224 * Kill all LWPs except the current one. Do not try to signal 225 * LWPs which have exited on their own or have already been 226 * signaled. 227 */ 228 static void 229 killlwps(struct lwp *lp) 230 { 231 struct proc *p = lp->lwp_proc; 232 struct lwp *tlp; 233 234 /* 235 * Kill the remaining LWPs. We must send the signal before setting 236 * LWP_MP_WEXIT. The setting of WEXIT is optional but helps reduce 237 * races. tlp must be held across the call as it might block and 238 * allow the target lwp to rip itself out from under our loop. 239 */ 240 FOREACH_LWP_IN_PROC(tlp, p) { 241 LWPHOLD(tlp); 242 lwkt_gettoken(&tlp->lwp_token); 243 if ((tlp->lwp_mpflags & LWP_MP_WEXIT) == 0) { 244 lwpsignal(p, tlp, SIGKILL); 245 atomic_set_int(&tlp->lwp_mpflags, LWP_MP_WEXIT); 246 } 247 lwkt_reltoken(&tlp->lwp_token); 248 LWPRELE(tlp); 249 } 250 251 /* 252 * Wait for everything to clear out. 253 */ 254 while (p->p_nthreads > 1) 255 tsleep(&p->p_nthreads, 0, "killlwps", 0); 256 } 257 258 /* 259 * Exit: deallocate address space and other resources, change proc state 260 * to zombie, and unlink proc from allproc and parent's lists. Save exit 261 * status and rusage for wait(). Check for child processes and orphan them. 262 */ 263 void 264 exit1(int rv) 265 { 266 struct thread *td = curthread; 267 struct proc *p = td->td_proc; 268 struct lwp *lp = td->td_lwp; 269 struct proc *q; 270 struct proc *pp; 271 struct vmspace *vm; 272 struct vnode *vtmp; 273 struct exitlist *ep; 274 int error; 275 276 lwkt_gettoken(&p->p_token); 277 278 if (p->p_pid == 1) { 279 kprintf("init died (signal %d, exit %d)\n", 280 WTERMSIG(rv), WEXITSTATUS(rv)); 281 panic("Going nowhere without my init!"); 282 } 283 varsymset_clean(&p->p_varsymset); 284 lockuninit(&p->p_varsymset.vx_lock); 285 286 /* 287 * Kill all lwps associated with the current process, return an 288 * error if we race another thread trying to do the same thing 289 * and lose the race. 290 */ 291 error = killalllwps(0); 292 if (error) { 293 lwp_exit(0, NULL); 294 /* NOT REACHED */ 295 } 296 297 /* are we a task leader? */ 298 if (p == p->p_leader) { 299 struct kill_args killArgs; 300 killArgs.signum = SIGKILL; 301 q = p->p_peers; 302 while(q) { 303 killArgs.pid = q->p_pid; 304 /* 305 * The interface for kill is better 306 * than the internal signal 307 */ 308 sys_kill(&killArgs); 309 q = q->p_peers; 310 } 311 while (p->p_peers) 312 tsleep((caddr_t)p, 0, "exit1", 0); 313 } 314 315 #ifdef PGINPROF 316 vmsizmon(); 317 #endif 318 STOPEVENT(p, S_EXIT, rv); 319 p->p_flags |= P_POSTEXIT; /* stop procfs stepping */ 320 321 /* 322 * Check if any loadable modules need anything done at process exit. 323 * e.g. SYSV IPC stuff 324 * XXX what if one of these generates an error? 325 */ 326 p->p_xstat = rv; 327 EVENTHANDLER_INVOKE(process_exit, p); 328 329 /* 330 * XXX: imho, the eventhandler stuff is much cleaner than this. 331 * Maybe we should move everything to use eventhandler. 332 */ 333 TAILQ_FOREACH(ep, &exit_list, next) 334 (*ep->function)(td); 335 336 if (p->p_flags & P_PROFIL) 337 stopprofclock(p); 338 339 SIGEMPTYSET(p->p_siglist); 340 SIGEMPTYSET(lp->lwp_siglist); 341 if (timevalisset(&p->p_realtimer.it_value)) 342 callout_stop_sync(&p->p_ithandle); 343 344 /* 345 * Reset any sigio structures pointing to us as a result of 346 * F_SETOWN with our pid. 347 */ 348 funsetownlst(&p->p_sigiolst); 349 350 /* 351 * Close open files and release open-file table. 352 * This may block! 353 */ 354 fdfree(p, NULL); 355 356 if (p->p_leader->p_peers) { 357 q = p->p_leader; 358 while(q->p_peers != p) 359 q = q->p_peers; 360 q->p_peers = p->p_peers; 361 wakeup((caddr_t)p->p_leader); 362 } 363 364 /* 365 * XXX Shutdown SYSV semaphores 366 */ 367 semexit(p); 368 369 KKASSERT(p->p_numposixlocks == 0); 370 371 /* The next two chunks should probably be moved to vmspace_exit. */ 372 vm = p->p_vmspace; 373 374 /* 375 * Clean up data related to virtual kernel operation. Clean up 376 * any vkernel context related to the current lwp now so we can 377 * destroy p_vkernel. 378 */ 379 if (p->p_vkernel) { 380 vkernel_lwp_exit(lp); 381 vkernel_exit(p); 382 } 383 384 /* 385 * Release user portion of address space. 386 * This releases references to vnodes, 387 * which could cause I/O if the file has been unlinked. 388 * Need to do this early enough that we can still sleep. 389 * Can't free the entire vmspace as the kernel stack 390 * may be mapped within that space also. 391 * 392 * Processes sharing the same vmspace may exit in one order, and 393 * get cleaned up by vmspace_exit() in a different order. The 394 * last exiting process to reach this point releases as much of 395 * the environment as it can, and the last process cleaned up 396 * by vmspace_exit() (which decrements exitingcnt) cleans up the 397 * remainder. 398 */ 399 vmspace_exitbump(vm); 400 sysref_put(&vm->vm_sysref); 401 402 if (SESS_LEADER(p)) { 403 struct session *sp = p->p_session; 404 405 if (sp->s_ttyvp) { 406 /* 407 * We are the controlling process. Signal the 408 * foreground process group, drain the controlling 409 * terminal, and revoke access to the controlling 410 * terminal. 411 * 412 * NOTE: while waiting for the process group to exit 413 * it is possible that one of the processes in the 414 * group will revoke the tty, so the ttyclosesession() 415 * function will re-check sp->s_ttyvp. 416 */ 417 if (sp->s_ttyp && (sp->s_ttyp->t_session == sp)) { 418 if (sp->s_ttyp->t_pgrp) 419 pgsignal(sp->s_ttyp->t_pgrp, SIGHUP, 1); 420 ttywait(sp->s_ttyp); 421 ttyclosesession(sp, 1); /* also revoke */ 422 } 423 /* 424 * Release the tty. If someone has it open via 425 * /dev/tty then close it (since they no longer can 426 * once we've NULL'd it out). 427 */ 428 ttyclosesession(sp, 0); 429 430 /* 431 * s_ttyp is not zero'd; we use this to indicate 432 * that the session once had a controlling terminal. 433 * (for logging and informational purposes) 434 */ 435 } 436 sp->s_leader = NULL; 437 } 438 fixjobc(p, p->p_pgrp, 0); 439 (void)acct_process(p); 440 #ifdef KTRACE 441 /* 442 * release trace file 443 */ 444 if (p->p_tracenode) 445 ktrdestroy(&p->p_tracenode); 446 p->p_traceflag = 0; 447 #endif 448 /* 449 * Release reference to text vnode 450 */ 451 if ((vtmp = p->p_textvp) != NULL) { 452 p->p_textvp = NULL; 453 vrele(vtmp); 454 } 455 456 /* Release namecache handle to text file */ 457 if (p->p_textnch.ncp) 458 cache_drop(&p->p_textnch); 459 460 /* 461 * We have to handle PPWAIT here or proc_move_allproc_zombie() 462 * will block on the PHOLD() the parent is doing. 463 * 464 * We are using the flag as an interlock so an atomic op is 465 * necessary to synchronize with the parent's cpu. 466 */ 467 if (p->p_flags & P_PPWAIT) { 468 atomic_clear_int(&p->p_flags, P_PPWAIT); 469 wakeup(p->p_pptr); 470 } 471 472 /* 473 * Move the process to the zombie list. This will block 474 * until the process p_lock count reaches 0. The process will 475 * not be reaped until TDF_EXITING is set by cpu_thread_exit(), 476 * which is called from cpu_proc_exit(). 477 * 478 * Interlock against waiters using p_waitgen. We increment 479 * p_waitgen after completing the move of our process to the 480 * zombie list. 481 * 482 * WARNING: pp becomes stale when we block, clear it now as a 483 * reminder. 484 */ 485 proc_move_allproc_zombie(p); 486 pp = p->p_pptr; 487 atomic_add_long(&pp->p_waitgen, 1); 488 pp = NULL; 489 490 /* 491 * Reparent all of this process's children to the init process. 492 * We must hold initproc->p_token in order to mess with 493 * initproc->p_children. We already hold p->p_token (to remove 494 * the children from our list). 495 */ 496 q = LIST_FIRST(&p->p_children); 497 if (q) { 498 lwkt_gettoken(&initproc->p_token); 499 while ((q = LIST_FIRST(&p->p_children)) != NULL) { 500 PHOLD(q); 501 lwkt_gettoken(&q->p_token); 502 if (q != LIST_FIRST(&p->p_children)) { 503 lwkt_reltoken(&q->p_token); 504 PRELE(q); 505 continue; 506 } 507 LIST_REMOVE(q, p_sibling); 508 LIST_INSERT_HEAD(&initproc->p_children, q, p_sibling); 509 q->p_pptr = initproc; 510 q->p_sigparent = SIGCHLD; 511 512 /* 513 * Traced processes are killed 514 * since their existence means someone is screwing up. 515 */ 516 if (q->p_flags & P_TRACED) { 517 q->p_flags &= ~P_TRACED; 518 ksignal(q, SIGKILL); 519 } 520 lwkt_reltoken(&q->p_token); 521 PRELE(q); 522 } 523 lwkt_reltoken(&initproc->p_token); 524 wakeup(initproc); 525 } 526 527 /* 528 * Save exit status and final rusage info, adding in child rusage 529 * info and self times. 530 */ 531 calcru_proc(p, &p->p_ru); 532 ruadd(&p->p_ru, &p->p_cru); 533 534 /* 535 * notify interested parties of our demise. 536 */ 537 KNOTE(&p->p_klist, NOTE_EXIT); 538 539 /* 540 * Notify parent that we're gone. If parent has the PS_NOCLDWAIT 541 * flag set, or if the handler is set to SIG_IGN, notify process 1 542 * instead (and hope it will handle this situation). 543 * 544 * (must reload pp) 545 */ 546 if (p->p_pptr->p_sigacts->ps_flag & (PS_NOCLDWAIT | PS_CLDSIGIGN)) { 547 proc_reparent(p, initproc); 548 } 549 550 pp = p->p_pptr; 551 PHOLD(pp); 552 if (p->p_sigparent && pp != initproc) { 553 ksignal(pp, p->p_sigparent); 554 } else { 555 ksignal(pp, SIGCHLD); 556 } 557 p->p_flags &= ~P_TRACED; 558 PRELE(pp); 559 560 /* 561 * cpu_exit is responsible for clearing curproc, since 562 * it is heavily integrated with the thread/switching sequence. 563 * 564 * Other substructures are freed from wait(). 565 */ 566 plimit_free(p); 567 568 /* 569 * Finally, call machine-dependent code to release as many of the 570 * lwp's resources as we can and halt execution of this thread. 571 * 572 * pp is a wild pointer now but still the correct wakeup() target. 573 * lwp_exit() only uses it to send the wakeup() signal to the likely 574 * parent. Any reparenting race that occurs will get a signal 575 * automatically and not be an issue. 576 */ 577 lwp_exit(1, pp); 578 } 579 580 /* 581 * Eventually called by every exiting LWP 582 * 583 * p->p_token must be held. mplock may be held and will be released. 584 */ 585 void 586 lwp_exit(int masterexit, void *waddr) 587 { 588 struct thread *td = curthread; 589 struct lwp *lp = td->td_lwp; 590 struct proc *p = lp->lwp_proc; 591 int dowake = 0; 592 593 /* 594 * Release the current user process designation on the process so 595 * the userland scheduler can work in someone else. 596 */ 597 p->p_usched->release_curproc(lp); 598 599 /* 600 * lwp_exit() may be called without setting LWP_MP_WEXIT, so 601 * make sure it is set here. 602 */ 603 ASSERT_LWKT_TOKEN_HELD(&p->p_token); 604 atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT); 605 606 /* 607 * Clean up any virtualization 608 */ 609 if (lp->lwp_vkernel) 610 vkernel_lwp_exit(lp); 611 612 if (td->td_vmm) 613 vmm_vmdestroy(); 614 615 /* 616 * Clean up select/poll support 617 */ 618 kqueue_terminate(&lp->lwp_kqueue); 619 620 /* 621 * Clean up any syscall-cached ucred 622 */ 623 if (td->td_ucred) { 624 crfree(td->td_ucred); 625 td->td_ucred = NULL; 626 } 627 628 /* 629 * Nobody actually wakes us when the lock 630 * count reaches zero, so just wait one tick. 631 */ 632 while (lp->lwp_lock > 0) 633 tsleep(lp, 0, "lwpexit", 1); 634 635 /* Hand down resource usage to our proc */ 636 ruadd(&p->p_ru, &lp->lwp_ru); 637 638 /* 639 * If we don't hold the process until the LWP is reaped wait*() 640 * may try to dispose of its vmspace before all the LWPs have 641 * actually terminated. 642 */ 643 PHOLD(p); 644 645 /* 646 * Do any remaining work that might block on us. We should be 647 * coded such that further blocking is ok after decrementing 648 * p_nthreads but don't take the chance. 649 */ 650 dsched_exit_thread(td); 651 biosched_done(curthread); 652 653 /* 654 * We have to use the reaper for all the LWPs except the one doing 655 * the master exit. The LWP doing the master exit can just be 656 * left on p_lwps and the process reaper will deal with it 657 * synchronously, which is much faster. 658 * 659 * Wakeup anyone waiting on p_nthreads to drop to 1 or 0. 660 * 661 * The process is left held until the reaper calls lwp_dispose() on 662 * the lp (after calling lwp_wait()). 663 */ 664 if (masterexit == 0) { 665 int cpu = mycpuid; 666 667 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 668 --p->p_nthreads; 669 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 670 dowake = 1; 671 lwkt_gettoken(&deadlwp_token[cpu]); 672 LIST_INSERT_HEAD(&deadlwp_list[cpu], lp, u.lwp_reap_entry); 673 taskqueue_enqueue(taskqueue_thread[cpu], deadlwp_task[cpu]); 674 lwkt_reltoken(&deadlwp_token[cpu]); 675 } else { 676 --p->p_nthreads; 677 if ((p->p_flags & P_MAYBETHREADED) && p->p_nthreads <= 1) 678 dowake = 1; 679 } 680 681 /* 682 * We no longer need p_token. 683 * 684 * Tell the userland scheduler that we are going away 685 */ 686 lwkt_reltoken(&p->p_token); 687 p->p_usched->heuristic_exiting(lp, p); 688 689 /* 690 * Issue late wakeups after releasing our token to give us a chance 691 * to deschedule and switch away before another cpu in a wait*() 692 * reaps us. This is done as late as possible to reduce contention. 693 */ 694 if (dowake) 695 wakeup(&p->p_nthreads); 696 if (waddr) 697 wakeup(waddr); 698 699 cpu_lwp_exit(); 700 } 701 702 /* 703 * Wait until a lwp is completely dead. The final interlock in this drama 704 * is when TDF_EXITING is set in cpu_thread_exit() just before the final 705 * switchout. 706 * 707 * At the point TDF_EXITING is set a complete exit is accomplished when 708 * TDF_RUNNING and TDF_PREEMPT_LOCK are both clear. td_mpflags has two 709 * post-switch interlock flags that can be used to wait for the TDF_ 710 * flags to clear. 711 * 712 * Returns non-zero on success, and zero if the caller needs to retry 713 * the lwp_wait(). 714 */ 715 static int 716 lwp_wait(struct lwp *lp) 717 { 718 struct thread *td = lp->lwp_thread; 719 u_int mpflags; 720 721 KKASSERT(lwkt_preempted_proc() != lp); 722 723 /* 724 * This bit of code uses the thread destruction interlock 725 * managed by lwkt_switch_return() to wait for the lwp's 726 * thread to completely disengage. 727 * 728 * It is possible for us to race another cpu core so we 729 * have to do this correctly. 730 */ 731 for (;;) { 732 mpflags = td->td_mpflags; 733 cpu_ccfence(); 734 if (mpflags & TDF_MP_EXITSIG) 735 break; 736 tsleep_interlock(td, 0); 737 if (atomic_cmpset_int(&td->td_mpflags, mpflags, 738 mpflags | TDF_MP_EXITWAIT)) { 739 tsleep(td, PINTERLOCKED, "lwpxt", 0); 740 } 741 } 742 743 /* 744 * We've already waited for the core exit but there can still 745 * be other refs from e.g. process scans and such. 746 */ 747 if (lp->lwp_lock > 0) { 748 tsleep(lp, 0, "lwpwait1", 1); 749 return(0); 750 } 751 if (td->td_refs) { 752 tsleep(td, 0, "lwpwait2", 1); 753 return(0); 754 } 755 756 /* 757 * Now that we have the thread destruction interlock these flags 758 * really should already be cleaned up, keep a check for safety. 759 * 760 * We can't rip its stack out from under it until TDF_EXITING is 761 * set and both TDF_RUNNING and TDF_PREEMPT_LOCK are clear. 762 * TDF_PREEMPT_LOCK must be checked because TDF_RUNNING 763 * will be cleared temporarily if a thread gets preempted. 764 */ 765 while ((td->td_flags & (TDF_RUNNING | 766 TDF_PREEMPT_LOCK | 767 TDF_EXITING)) != TDF_EXITING) { 768 tsleep(lp, 0, "lwpwait3", 1); 769 return (0); 770 } 771 772 KASSERT((td->td_flags & (TDF_RUNQ|TDF_TSLEEPQ)) == 0, 773 ("lwp_wait: td %p (%s) still on run or sleep queue", 774 td, td->td_comm)); 775 return (1); 776 } 777 778 /* 779 * Release the resources associated with a lwp. 780 * The lwp must be completely dead. 781 */ 782 void 783 lwp_dispose(struct lwp *lp) 784 { 785 struct thread *td = lp->lwp_thread; 786 787 KKASSERT(lwkt_preempted_proc() != lp); 788 KKASSERT(td->td_refs == 0); 789 KKASSERT((td->td_flags & (TDF_RUNNING | 790 TDF_PREEMPT_LOCK | 791 TDF_EXITING)) == TDF_EXITING); 792 793 PRELE(lp->lwp_proc); 794 lp->lwp_proc = NULL; 795 if (td != NULL) { 796 td->td_proc = NULL; 797 td->td_lwp = NULL; 798 lp->lwp_thread = NULL; 799 lwkt_free_thread(td); 800 } 801 kfree(lp, M_LWP); 802 } 803 804 int 805 sys_wait4(struct wait_args *uap) 806 { 807 struct rusage rusage; 808 int error, status; 809 810 error = kern_wait(uap->pid, (uap->status ? &status : NULL), 811 uap->options, (uap->rusage ? &rusage : NULL), 812 &uap->sysmsg_result); 813 814 if (error == 0 && uap->status) 815 error = copyout(&status, uap->status, sizeof(*uap->status)); 816 if (error == 0 && uap->rusage) 817 error = copyout(&rusage, uap->rusage, sizeof(*uap->rusage)); 818 return (error); 819 } 820 821 /* 822 * wait1() 823 * 824 * wait_args(int pid, int *status, int options, struct rusage *rusage) 825 */ 826 int 827 kern_wait(pid_t pid, int *status, int options, struct rusage *rusage, int *res) 828 { 829 struct thread *td = curthread; 830 struct lwp *lp; 831 struct proc *q = td->td_proc; 832 struct proc *p, *t; 833 struct pargs *pa; 834 struct sigacts *ps; 835 int nfound, error; 836 long waitgen; 837 838 if (pid == 0) 839 pid = -q->p_pgid; 840 if (options &~ (WUNTRACED|WNOHANG|WCONTINUED|WLINUXCLONE)) 841 return (EINVAL); 842 843 /* 844 * Protect the q->p_children list 845 */ 846 lwkt_gettoken(&q->p_token); 847 loop: 848 /* 849 * All sorts of things can change due to blocking so we have to loop 850 * all the way back up here. 851 * 852 * The problem is that if a process group is stopped and the parent 853 * is doing a wait*(..., WUNTRACED, ...), it will see the STOP 854 * of the child and then stop itself when it tries to return from the 855 * system call. When the process group is resumed the parent will 856 * then get the STOP status even though the child has now resumed 857 * (a followup wait*() will get the CONT status). 858 * 859 * Previously the CONT would overwrite the STOP because the tstop 860 * was handled within tsleep(), and the parent would only see 861 * the CONT when both are stopped and continued together. This little 862 * two-line hack restores this effect. 863 */ 864 while (q->p_stat == SSTOP) 865 tstop(); 866 867 nfound = 0; 868 869 /* 870 * Loop on children. 871 * 872 * NOTE: We don't want to break q's p_token in the loop for the 873 * case where no children are found or we risk breaking the 874 * interlock between child and parent. 875 */ 876 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 877 LIST_FOREACH(p, &q->p_children, p_sibling) { 878 if (pid != WAIT_ANY && 879 p->p_pid != pid && p->p_pgid != -pid) { 880 continue; 881 } 882 883 /* 884 * This special case handles a kthread spawned by linux_clone 885 * (see linux_misc.c). The linux_wait4 and linux_waitpid 886 * functions need to be able to distinguish between waiting 887 * on a process and waiting on a thread. It is a thread if 888 * p_sigparent is not SIGCHLD, and the WLINUXCLONE option 889 * signifies we want to wait for threads and not processes. 890 */ 891 if ((p->p_sigparent != SIGCHLD) ^ 892 ((options & WLINUXCLONE) != 0)) { 893 continue; 894 } 895 896 nfound++; 897 if (p->p_stat == SZOMB) { 898 /* 899 * We may go into SZOMB with threads still present. 900 * We must wait for them to exit before we can reap 901 * the master thread, otherwise we may race reaping 902 * non-master threads. 903 * 904 * Only this routine can remove a process from 905 * the zombie list and destroy it, use PACQUIREZOMB() 906 * to serialize us and loop if it blocks (interlocked 907 * by the parent's q->p_token). 908 * 909 * WARNING! (p) can be invalid when PHOLDZOMB(p) 910 * returns non-zero. Be sure not to 911 * mess with it. 912 */ 913 if (PHOLDZOMB(p)) 914 goto loop; 915 lwkt_gettoken(&p->p_token); 916 if (p->p_pptr != q) { 917 lwkt_reltoken(&p->p_token); 918 PRELEZOMB(p); 919 goto loop; 920 } 921 while (p->p_nthreads > 0) { 922 tsleep(&p->p_nthreads, 0, "lwpzomb", hz); 923 } 924 925 /* 926 * Reap any LWPs left in p->p_lwps. This is usually 927 * just the last LWP. This must be done before 928 * we loop on p_lock since the lwps hold a ref on 929 * it as a vmspace interlock. 930 * 931 * Once that is accomplished p_nthreads had better 932 * be zero. 933 */ 934 while ((lp = RB_ROOT(&p->p_lwp_tree)) != NULL) { 935 lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp); 936 reaplwp(lp); 937 } 938 KKASSERT(p->p_nthreads == 0); 939 940 /* 941 * Don't do anything really bad until all references 942 * to the process go away. This may include other 943 * LWPs which are still in the process of being 944 * reaped. We can't just pull the rug out from under 945 * them because they may still be using the VM space. 946 * 947 * Certain kernel facilities such as /proc will also 948 * put a hold on the process for short periods of 949 * time. 950 */ 951 PRELE(p); 952 PSTALL(p, "reap3", 0); 953 954 /* Take care of our return values. */ 955 *res = p->p_pid; 956 957 if (status) 958 *status = p->p_xstat; 959 if (rusage) 960 *rusage = p->p_ru; 961 962 /* 963 * If we got the child via a ptrace 'attach', 964 * we need to give it back to the old parent. 965 */ 966 if (p->p_oppid && (t = pfind(p->p_oppid)) != NULL) { 967 PHOLD(p); 968 p->p_oppid = 0; 969 proc_reparent(p, t); 970 ksignal(t, SIGCHLD); 971 wakeup((caddr_t)t); 972 error = 0; 973 PRELE(t); 974 lwkt_reltoken(&p->p_token); 975 PRELEZOMB(p); 976 goto done; 977 } 978 979 /* 980 * Unlink the proc from its process group so that 981 * the following operations won't lead to an 982 * inconsistent state for processes running down 983 * the zombie list. 984 */ 985 proc_remove_zombie(p); 986 lwkt_reltoken(&p->p_token); 987 leavepgrp(p); 988 989 p->p_xstat = 0; 990 ruadd(&q->p_cru, &p->p_ru); 991 992 /* 993 * Decrement the count of procs running with this uid. 994 */ 995 chgproccnt(p->p_ucred->cr_ruidinfo, -1, 0); 996 997 /* 998 * Free up credentials. 999 */ 1000 crfree(p->p_ucred); 1001 p->p_ucred = NULL; 1002 1003 /* 1004 * Remove unused arguments 1005 */ 1006 pa = p->p_args; 1007 p->p_args = NULL; 1008 if (pa && refcount_release(&pa->ar_ref)) { 1009 kfree(pa, M_PARGS); 1010 pa = NULL; 1011 } 1012 1013 ps = p->p_sigacts; 1014 p->p_sigacts = NULL; 1015 if (ps && refcount_release(&ps->ps_refcnt)) { 1016 kfree(ps, M_SUBPROC); 1017 ps = NULL; 1018 } 1019 1020 /* 1021 * Our exitingcount was incremented when the process 1022 * became a zombie, now that the process has been 1023 * removed from (almost) all lists we should be able 1024 * to safely destroy its vmspace. Wait for any current 1025 * holders to go away (so the vmspace remains stable), 1026 * then scrap it. 1027 */ 1028 PSTALL(p, "reap4", 0); 1029 vmspace_exitfree(p); 1030 PSTALL(p, "reap5", 0); 1031 1032 /* 1033 * NOTE: We have to officially release ZOMB in order 1034 * to ensure that a racing thread in kern_wait() 1035 * which blocked on ZOMB is woken up. 1036 */ 1037 PHOLD(p); 1038 PRELEZOMB(p); 1039 kfree(p, M_PROC); 1040 atomic_add_int(&nprocs, -1); 1041 error = 0; 1042 goto done; 1043 } 1044 if (p->p_stat == SSTOP && (p->p_flags & P_WAITED) == 0 && 1045 ((p->p_flags & P_TRACED) || (options & WUNTRACED))) { 1046 PHOLD(p); 1047 lwkt_gettoken(&p->p_token); 1048 if (p->p_pptr != q) { 1049 lwkt_reltoken(&p->p_token); 1050 PRELE(p); 1051 goto loop; 1052 } 1053 if (p->p_stat != SSTOP || 1054 (p->p_flags & P_WAITED) != 0 || 1055 ((p->p_flags & P_TRACED) == 0 && 1056 (options & WUNTRACED) == 0)) { 1057 lwkt_reltoken(&p->p_token); 1058 PRELE(p); 1059 goto loop; 1060 } 1061 1062 p->p_flags |= P_WAITED; 1063 1064 *res = p->p_pid; 1065 if (status) 1066 *status = W_STOPCODE(p->p_xstat); 1067 /* Zero rusage so we get something consistent. */ 1068 if (rusage) 1069 bzero(rusage, sizeof(*rusage)); 1070 error = 0; 1071 lwkt_reltoken(&p->p_token); 1072 PRELE(p); 1073 goto done; 1074 } 1075 if ((options & WCONTINUED) && (p->p_flags & P_CONTINUED)) { 1076 PHOLD(p); 1077 lwkt_gettoken(&p->p_token); 1078 if (p->p_pptr != q) { 1079 lwkt_reltoken(&p->p_token); 1080 PRELE(p); 1081 goto loop; 1082 } 1083 if ((p->p_flags & P_CONTINUED) == 0) { 1084 lwkt_reltoken(&p->p_token); 1085 PRELE(p); 1086 goto loop; 1087 } 1088 1089 *res = p->p_pid; 1090 p->p_flags &= ~P_CONTINUED; 1091 1092 if (status) 1093 *status = SIGCONT; 1094 error = 0; 1095 lwkt_reltoken(&p->p_token); 1096 PRELE(p); 1097 goto done; 1098 } 1099 } 1100 if (nfound == 0) { 1101 error = ECHILD; 1102 goto done; 1103 } 1104 if (options & WNOHANG) { 1105 *res = 0; 1106 error = 0; 1107 goto done; 1108 } 1109 1110 /* 1111 * Wait for signal - interlocked using q->p_waitgen. 1112 */ 1113 error = 0; 1114 while ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1115 tsleep_interlock(q, PCATCH); 1116 waitgen = atomic_fetchadd_long(&q->p_waitgen, 0x80000000); 1117 if ((waitgen & 0x7FFFFFFF) == (q->p_waitgen & 0x7FFFFFFF)) { 1118 error = tsleep(q, PCATCH | PINTERLOCKED, "wait", 0); 1119 break; 1120 } 1121 } 1122 if (error) { 1123 done: 1124 lwkt_reltoken(&q->p_token); 1125 return (error); 1126 } 1127 goto loop; 1128 } 1129 1130 /* 1131 * Change child's parent process to parent. 1132 * 1133 * p_children/p_sibling requires the parent's token, and 1134 * changing pptr requires the child's token, so we have to 1135 * get three tokens to do this operation. We also need to 1136 * hold pointers that might get ripped out from under us to 1137 * preserve structural integrity. 1138 * 1139 * It is possible to race another reparent or disconnect or other 1140 * similar operation. We must retry when this situation occurs. 1141 * Once we successfully reparent the process we no longer care 1142 * about any races. 1143 */ 1144 void 1145 proc_reparent(struct proc *child, struct proc *parent) 1146 { 1147 struct proc *opp; 1148 1149 PHOLD(parent); 1150 while ((opp = child->p_pptr) != parent) { 1151 PHOLD(opp); 1152 lwkt_gettoken(&opp->p_token); 1153 lwkt_gettoken(&child->p_token); 1154 lwkt_gettoken(&parent->p_token); 1155 if (child->p_pptr != opp) { 1156 lwkt_reltoken(&parent->p_token); 1157 lwkt_reltoken(&child->p_token); 1158 lwkt_reltoken(&opp->p_token); 1159 PRELE(opp); 1160 continue; 1161 } 1162 LIST_REMOVE(child, p_sibling); 1163 LIST_INSERT_HEAD(&parent->p_children, child, p_sibling); 1164 child->p_pptr = parent; 1165 lwkt_reltoken(&parent->p_token); 1166 lwkt_reltoken(&child->p_token); 1167 lwkt_reltoken(&opp->p_token); 1168 if (LIST_EMPTY(&opp->p_children)) 1169 wakeup(opp); 1170 PRELE(opp); 1171 break; 1172 } 1173 PRELE(parent); 1174 } 1175 1176 /* 1177 * The next two functions are to handle adding/deleting items on the 1178 * exit callout list 1179 * 1180 * at_exit(): 1181 * Take the arguments given and put them onto the exit callout list, 1182 * However first make sure that it's not already there. 1183 * returns 0 on success. 1184 */ 1185 1186 int 1187 at_exit(exitlist_fn function) 1188 { 1189 struct exitlist *ep; 1190 1191 #ifdef INVARIANTS 1192 /* Be noisy if the programmer has lost track of things */ 1193 if (rm_at_exit(function)) 1194 kprintf("WARNING: exit callout entry (%p) already present\n", 1195 function); 1196 #endif 1197 ep = kmalloc(sizeof(*ep), M_ATEXIT, M_NOWAIT); 1198 if (ep == NULL) 1199 return (ENOMEM); 1200 ep->function = function; 1201 TAILQ_INSERT_TAIL(&exit_list, ep, next); 1202 return (0); 1203 } 1204 1205 /* 1206 * Scan the exit callout list for the given item and remove it. 1207 * Returns the number of items removed (0 or 1) 1208 */ 1209 int 1210 rm_at_exit(exitlist_fn function) 1211 { 1212 struct exitlist *ep; 1213 1214 TAILQ_FOREACH(ep, &exit_list, next) { 1215 if (ep->function == function) { 1216 TAILQ_REMOVE(&exit_list, ep, next); 1217 kfree(ep, M_ATEXIT); 1218 return(1); 1219 } 1220 } 1221 return (0); 1222 } 1223 1224 /* 1225 * LWP reaper related code. 1226 */ 1227 static void 1228 reaplwps(void *context, int dummy) 1229 { 1230 struct lwplist *lwplist = context; 1231 struct lwp *lp; 1232 int cpu = mycpuid; 1233 1234 lwkt_gettoken(&deadlwp_token[cpu]); 1235 while ((lp = LIST_FIRST(lwplist))) { 1236 LIST_REMOVE(lp, u.lwp_reap_entry); 1237 reaplwp(lp); 1238 } 1239 lwkt_reltoken(&deadlwp_token[cpu]); 1240 } 1241 1242 static void 1243 reaplwp(struct lwp *lp) 1244 { 1245 while (lwp_wait(lp) == 0) 1246 ; 1247 lwp_dispose(lp); 1248 } 1249 1250 static void 1251 deadlwp_init(void) 1252 { 1253 int cpu; 1254 1255 for (cpu = 0; cpu < ncpus; cpu++) { 1256 lwkt_token_init(&deadlwp_token[cpu], "deadlwpl"); 1257 LIST_INIT(&deadlwp_list[cpu]); 1258 deadlwp_task[cpu] = kmalloc(sizeof(*deadlwp_task[cpu]), 1259 M_DEVBUF, M_WAITOK); 1260 TASK_INIT(deadlwp_task[cpu], 0, reaplwps, &deadlwp_list[cpu]); 1261 } 1262 } 1263 1264 SYSINIT(deadlwpinit, SI_SUB_CONFIGURE, SI_ORDER_ANY, deadlwp_init, NULL); 1265