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