xref: /dragonfly/sys/kern/kern_fork.c (revision 2b7dbe20)
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_fork.c	8.6 (Berkeley) 4/8/94
35  * $FreeBSD: src/sys/kern/kern_fork.c,v 1.72.2.14 2003/06/26 04:15:10 silby Exp $
36  */
37 
38 #include "opt_ktrace.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/sysmsg.h>
43 #include <sys/filedesc.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/malloc.h>
47 #include <sys/proc.h>
48 #include <sys/resourcevar.h>
49 #include <sys/vnode.h>
50 #include <sys/acct.h>
51 #include <sys/ktrace.h>
52 #include <sys/unistd.h>
53 #include <sys/jail.h>
54 #include <sys/lwp.h>
55 
56 #include <vm/vm.h>
57 #include <sys/lock.h>
58 #include <vm/pmap.h>
59 #include <vm/vm_map.h>
60 #include <vm/vm_extern.h>
61 
62 #include <sys/vmmeter.h>
63 #include <sys/refcount.h>
64 #include <sys/thread2.h>
65 #include <sys/signal2.h>
66 #include <sys/spinlock2.h>
67 
68 #include <sys/dsched.h>
69 
70 static MALLOC_DEFINE(M_ATFORK, "atfork", "atfork callback");
71 static MALLOC_DEFINE(M_REAPER, "reaper", "process reapers");
72 
73 /*
74  * These are the stuctures used to create a callout list for things to do
75  * when forking a process
76  */
77 struct forklist {
78 	forklist_fn function;
79 	TAILQ_ENTRY(forklist) next;
80 };
81 
82 TAILQ_HEAD(forklist_head, forklist);
83 static struct forklist_head fork_list = TAILQ_HEAD_INITIALIZER(fork_list);
84 
85 static struct lwp	*lwp_fork1(struct lwp *, struct proc *, int flags,
86 			    const cpumask_t *mask);
87 static void		lwp_fork2(struct lwp *lp1, struct proc *destproc,
88 			    struct lwp *lp2, int flags);
89 static int		lwp_create1(struct lwp_params *params,
90 			    const cpumask_t *mask);
91 static struct lock reaper_lock = LOCK_INITIALIZER("reapgl", 0, 0);
92 
93 int forksleep; /* Place for fork1() to sleep on. */
94 
95 /*
96  * Red-Black tree support for LWPs
97  */
98 
99 static int
100 rb_lwp_compare(struct lwp *lp1, struct lwp *lp2)
101 {
102 	if (lp1->lwp_tid < lp2->lwp_tid)
103 		return(-1);
104 	if (lp1->lwp_tid > lp2->lwp_tid)
105 		return(1);
106 	return(0);
107 }
108 
109 RB_GENERATE2(lwp_rb_tree, lwp, u.lwp_rbnode, rb_lwp_compare, lwpid_t, lwp_tid);
110 
111 /*
112  * When forking, memory underpinning umtx-supported mutexes may be set
113  * COW causing the physical address to change.  We must wakeup any threads
114  * blocked on the physical address to allow them to re-resolve their VM.
115  *
116  * (caller is holding p->p_token)
117  */
118 static void
119 wake_umtx_threads(struct proc *p1)
120 {
121 	struct lwp *lp;
122 	struct thread *td;
123 
124 	RB_FOREACH(lp, lwp_rb_tree, &p1->p_lwp_tree) {
125 		td = lp->lwp_thread;
126 		if (td && (td->td_flags & TDF_TSLEEPQ) &&
127 		    (td->td_wdomain & PDOMAIN_MASK) == PDOMAIN_UMTX) {
128 			wakeup_domain(td->td_wchan, PDOMAIN_UMTX);
129 		}
130 	}
131 }
132 
133 /*
134  * fork() system call
135  */
136 int
137 sys_fork(struct sysmsg *sysmsg, const struct fork_args *uap)
138 {
139 	struct lwp *lp = curthread->td_lwp;
140 	struct proc *p2;
141 	int error;
142 
143 	error = fork1(lp, RFFDG | RFPROC | RFPGLOCK, &p2);
144 	if (error == 0) {
145 		PHOLD(p2);
146 		start_forked_proc(lp, p2);
147 		sysmsg->sysmsg_fds[0] = p2->p_pid;
148 		sysmsg->sysmsg_fds[1] = 0;
149 		PRELE(p2);
150 	}
151 	return error;
152 }
153 
154 /*
155  * vfork() system call
156  */
157 int
158 sys_vfork(struct sysmsg *sysmsg, const struct vfork_args *uap)
159 {
160 	struct lwp *lp = curthread->td_lwp;
161 	struct proc *p2;
162 	int error;
163 
164 	error = fork1(lp, RFFDG | RFPROC | RFPPWAIT | RFMEM | RFPGLOCK, &p2);
165 	if (error == 0) {
166 		PHOLD(p2);
167 		start_forked_proc(lp, p2);
168 		sysmsg->sysmsg_fds[0] = p2->p_pid;
169 		sysmsg->sysmsg_fds[1] = 0;
170 		PRELE(p2);
171 	}
172 	return error;
173 }
174 
175 /*
176  * Handle rforks.  An rfork may (1) operate on the current process without
177  * creating a new, (2) create a new process that shared the current process's
178  * vmspace, signals, and/or descriptors, or (3) create a new process that does
179  * not share these things (normal fork).
180  *
181  * Note that we only call start_forked_proc() if a new process is actually
182  * created.
183  *
184  * rfork { int flags }
185  */
186 int
187 sys_rfork(struct sysmsg *sysmsg, const struct rfork_args *uap)
188 {
189 	struct lwp *lp = curthread->td_lwp;
190 	struct proc *p2;
191 	int error;
192 
193 	if ((uap->flags & RFKERNELONLY) != 0)
194 		return (EINVAL);
195 
196 	error = fork1(lp, uap->flags | RFPGLOCK, &p2);
197 	if (error == 0) {
198 		if (p2) {
199 			PHOLD(p2);
200 			start_forked_proc(lp, p2);
201 			sysmsg->sysmsg_fds[0] = p2->p_pid;
202 			sysmsg->sysmsg_fds[1] = 0;
203 			PRELE(p2);
204 		} else {
205 			sysmsg->sysmsg_fds[0] = 0;
206 			sysmsg->sysmsg_fds[1] = 0;
207 		}
208 	}
209 	return error;
210 }
211 
212 static int
213 lwp_create1(struct lwp_params *uprm, const cpumask_t *umask)
214 {
215 	struct proc *p = curproc;
216 	struct lwp *lp;
217 	struct lwp_params params;
218 	cpumask_t *mask = NULL, mask0;
219 	int error;
220 
221 	error = copyin(uprm, &params, sizeof(params));
222 	if (error)
223 		goto fail2;
224 
225 	if (umask != NULL) {
226 		error = copyin(umask, &mask0, sizeof(mask0));
227 		if (error)
228 			goto fail2;
229 		CPUMASK_ANDMASK(mask0, smp_active_mask);
230 		if (CPUMASK_TESTNZERO(mask0))
231 			mask = &mask0;
232 	}
233 
234 	lwkt_gettoken(&p->p_token);
235 	plimit_lwp_fork(p);	/* force exclusive access */
236 	lp = lwp_fork1(curthread->td_lwp, p, RFPROC | RFMEM, mask);
237 	lwp_fork2(curthread->td_lwp, p, lp, RFPROC | RFMEM);
238 	error = cpu_prepare_lwp(lp, &params);
239 	if (error)
240 		goto fail;
241 	if (params.lwp_tid1 != NULL &&
242 	    (error = copyout(&lp->lwp_tid, params.lwp_tid1, sizeof(lp->lwp_tid))))
243 		goto fail;
244 	if (params.lwp_tid2 != NULL &&
245 	    (error = copyout(&lp->lwp_tid, params.lwp_tid2, sizeof(lp->lwp_tid))))
246 		goto fail;
247 
248 	/*
249 	 * Now schedule the new lwp.
250 	 */
251 	p->p_usched->resetpriority(lp);
252 	crit_enter();
253 	lp->lwp_stat = LSRUN;
254 	p->p_usched->setrunqueue(lp);
255 	crit_exit();
256 	lwkt_reltoken(&p->p_token);
257 
258 	return (0);
259 
260 fail:
261 	/*
262 	 * Make sure no one is using this lwp, before it is removed from
263 	 * the tree.  If we didn't wait it here, lwp tree iteration with
264 	 * blocking operation would be broken.
265 	 */
266 	while (lp->lwp_lock > 0)
267 		tsleep(lp, 0, "lwpfail", 1);
268 	lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp);
269 	--p->p_nthreads;
270 	/* lwp_dispose expects an exited lwp, and a held proc */
271 	atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT);
272 	lp->lwp_thread->td_flags |= TDF_EXITING;
273 	lwkt_remove_tdallq(lp->lwp_thread);
274 	PHOLD(p);
275 	biosched_done(lp->lwp_thread);
276 	dsched_exit_thread(lp->lwp_thread);
277 	lwp_dispose(lp);
278 	lwkt_reltoken(&p->p_token);
279 fail2:
280 	return (error);
281 }
282 
283 /*
284  * Low level thread create used by pthreads.
285  */
286 int
287 sys_lwp_create(struct sysmsg *sysmsg, const struct lwp_create_args *uap)
288 {
289 
290 	return (lwp_create1(uap->params, NULL));
291 }
292 
293 int
294 sys_lwp_create2(struct sysmsg *sysmsg, const struct lwp_create2_args *uap)
295 {
296 
297 	return (lwp_create1(uap->params, uap->mask));
298 }
299 
300 int	nprocs = 1;		/* process 0 */
301 
302 int
303 fork1(struct lwp *lp1, int flags, struct proc **procp)
304 {
305 	struct proc *p1 = lp1->lwp_proc;
306 	struct proc *p2;
307 	struct proc *pptr;
308 	struct pgrp *p1grp;
309 	struct pgrp *plkgrp;
310 	struct lwp  *lp2;
311 	struct sysreaper *reap;
312 	uid_t uid;
313 	int ok, error;
314 	static int curfail = 0;
315 	static struct timeval lastfail;
316 	struct forklist *ep;
317 	struct filedesc_to_leader *fdtol;
318 
319 	if ((flags & (RFFDG|RFCFDG)) == (RFFDG|RFCFDG))
320 		return (EINVAL);
321 
322 	lwkt_gettoken(&p1->p_token);
323 	plkgrp = NULL;
324 	p2 = NULL;
325 
326 	/*
327 	 * Here we don't create a new process, but we divorce
328 	 * certain parts of a process from itself.
329 	 */
330 	if ((flags & RFPROC) == 0) {
331 		/*
332 		 * This kind of stunt does not work anymore if
333 		 * there are native threads (lwps) running
334 		 */
335 		if (p1->p_nthreads != 1) {
336 			error = EINVAL;
337 			goto done;
338 		}
339 
340 		vm_fork(p1, NULL, NULL, flags);
341 		if ((flags & RFMEM) == 0)
342 			wake_umtx_threads(p1);
343 
344 		/*
345 		 * Close all file descriptors.
346 		 */
347 		if (flags & RFCFDG) {
348 			struct filedesc *fdtmp;
349 			fdtmp = fdinit(p1);
350 			fdfree(p1, fdtmp);
351 		}
352 
353 		/*
354 		 * Unshare file descriptors (from parent.)
355 		 */
356 		if (flags & RFFDG) {
357 			if (p1->p_fd->fd_refcnt > 1) {
358 				struct filedesc *newfd;
359 				error = fdcopy(p1, &newfd);
360 				if (error != 0) {
361 					error = ENOMEM;
362 					goto done;
363 				}
364 				fdfree(p1, newfd);
365 			}
366 		}
367 		*procp = NULL;
368 		error = 0;
369 		goto done;
370 	}
371 
372 	/*
373 	 * Interlock against process group signal delivery.  If signals
374 	 * are pending after the interlock is obtained we have to restart
375 	 * the system call to process the signals.  If we don't the child
376 	 * can miss a pgsignal (such as ^C) sent during the fork.
377 	 *
378 	 * We can't use CURSIG() here because it will process any STOPs
379 	 * and cause the process group lock to be held indefinitely.  If
380 	 * a STOP occurs, the fork will be restarted after the CONT.
381 	 */
382 	p1grp = p1->p_pgrp;
383 	if ((flags & RFPGLOCK) && (plkgrp = p1->p_pgrp) != NULL) {
384 		pgref(plkgrp);
385 		lockmgr(&plkgrp->pg_lock, LK_SHARED);
386 		if (CURSIG_NOBLOCK(lp1)) {
387 			error = ERESTART;
388 			goto done;
389 		}
390 	}
391 
392 	/*
393 	 * Although process entries are dynamically created, we still keep
394 	 * a global limit on the maximum number we will create.  Don't allow
395 	 * a nonprivileged user to use the last ten processes; don't let root
396 	 * exceed the limit. The variable nprocs is the current number of
397 	 * processes, maxproc is the limit.
398 	 */
399 	uid = lp1->lwp_thread->td_ucred->cr_ruid;
400 	if ((nprocs >= maxproc - 10 && uid != 0) || nprocs >= maxproc) {
401 		if (ppsratecheck(&lastfail, &curfail, 1))
402 			kprintf("maxproc limit exceeded by uid %d, please "
403 			       "see tuning(7) and login.conf(5).\n", uid);
404 		tsleep(&forksleep, 0, "fork", hz / 2);
405 		error = EAGAIN;
406 		goto done;
407 	}
408 
409 	/*
410 	 * Increment the nprocs resource before blocking can occur.  There
411 	 * are hard-limits as to the number of processes that can run.
412 	 */
413 	atomic_add_int(&nprocs, 1);
414 
415 	/*
416 	 * Increment the count of procs running with this uid.  This also
417 	 * applies to root.
418 	 */
419 	ok = chgproccnt(lp1->lwp_thread->td_ucred->cr_ruidinfo, 1,
420 			plimit_getadjvalue(RLIMIT_NPROC));
421 	if (!ok) {
422 		/*
423 		 * Back out the process count
424 		 */
425 		atomic_add_int(&nprocs, -1);
426 		if (ppsratecheck(&lastfail, &curfail, 1)) {
427 			kprintf("maxproc limit of %jd "
428 				"exceeded by \"%s\" uid %d, "
429 				"please see tuning(7) and login.conf(5).\n",
430 				plimit_getadjvalue(RLIMIT_NPROC),
431 				p1->p_comm,
432 				uid);
433 		}
434 		tsleep(&forksleep, 0, "fork", hz / 2);
435 		error = EAGAIN;
436 		goto done;
437 	}
438 
439 	/*
440 	 * Allocate a new process, don't get fancy: zero the structure.
441 	 */
442 	p2 = kmalloc(sizeof(struct proc), M_PROC, M_WAITOK|M_ZERO);
443 
444 	/*
445 	 * Core initialization.  SIDL is a safety state that protects the
446 	 * partially initialized process once it starts getting hooked
447 	 * into system structures and becomes addressable.
448 	 *
449 	 * We must be sure to acquire p2->p_token as well, we must hold it
450 	 * once the process is on the allproc list to avoid things such
451 	 * as competing modifications to p_flags.
452 	 */
453 	mycpu->gd_forkid += ncpus;
454 	p2->p_forkid = mycpu->gd_forkid + mycpu->gd_cpuid;
455 	p2->p_lasttid = 0;	/* first tid will be 1 */
456 	p2->p_stat = SIDL;
457 
458 	/*
459 	 * NOTE: Process 0 will not have a reaper, but process 1 (init) and
460 	 *	 all other processes always will.
461 	 */
462 	if ((reap = p1->p_reaper) != NULL) {
463 		reaper_hold(reap);
464 		p2->p_reaper = reap;
465 	} else {
466 		p2->p_reaper = NULL;
467 	}
468 
469 	RB_INIT(&p2->p_lwp_tree);
470 	spin_init(&p2->p_spin, "procfork1");
471 	lwkt_token_init(&p2->p_token, "proc");
472 	lwkt_gettoken(&p2->p_token);
473 	p2->p_uidpcpu = kmalloc(sizeof(*p2->p_uidpcpu) * ncpus,
474 				M_SUBPROC, M_WAITOK | M_ZERO);
475 
476 	/*
477 	 * Setup linkage for kernel based threading XXX lwp.  Also add the
478 	 * process to the allproclist.
479 	 *
480 	 * The process structure is addressable after this point.
481 	 */
482 	if (flags & RFTHREAD) {
483 		p2->p_peers = p1->p_peers;
484 		p1->p_peers = p2;
485 		p2->p_leader = p1->p_leader;
486 	} else {
487 		p2->p_leader = p2;
488 	}
489 	proc_add_allproc(p2);
490 
491 	/*
492 	 * Initialize the section which is copied verbatim from the parent.
493 	 */
494 	bcopy(&p1->p_startcopy, &p2->p_startcopy,
495 	      ((caddr_t)&p2->p_endcopy - (caddr_t)&p2->p_startcopy));
496 
497 	/*
498 	 * Duplicate sub-structures as needed.  Increase reference counts
499 	 * on shared objects.
500 	 *
501 	 * NOTE: because we are now on the allproc list it is possible for
502 	 *	 other consumers to gain temporary references to p2
503 	 *	 (p2->p_lock can change).
504 	 */
505 	if (p1->p_flags & P_PROFIL)
506 		startprofclock(p2);
507 	p2->p_ucred = crhold(lp1->lwp_thread->td_ucred);
508 
509 	if (jailed(p2->p_ucred))
510 		p2->p_flags |= P_JAILED;
511 
512 	if (p2->p_args)
513 		refcount_acquire(&p2->p_args->ar_ref);
514 
515 	p2->p_usched = p1->p_usched;
516 	/* XXX: verify copy of the secondary iosched stuff */
517 	dsched_enter_proc(p2);
518 
519 	if (flags & RFSIGSHARE) {
520 		p2->p_sigacts = p1->p_sigacts;
521 		refcount_acquire(&p2->p_sigacts->ps_refcnt);
522 	} else {
523 		p2->p_sigacts = kmalloc(sizeof(*p2->p_sigacts),
524 					M_SUBPROC, M_WAITOK);
525 		bcopy(p1->p_sigacts, p2->p_sigacts, sizeof(*p2->p_sigacts));
526 		refcount_init(&p2->p_sigacts->ps_refcnt, 1);
527 	}
528 	if (flags & RFLINUXTHPN)
529 	        p2->p_sigparent = SIGUSR1;
530 	else
531 	        p2->p_sigparent = SIGCHLD;
532 
533 	/* bump references to the text vnode (for procfs) */
534 	p2->p_textvp = p1->p_textvp;
535 	if (p2->p_textvp)
536 		vref(p2->p_textvp);
537 
538 	/* copy namecache handle to the text file */
539 	if (p1->p_textnch.mount)
540 		cache_copy(&p1->p_textnch, &p2->p_textnch);
541 
542 	/*
543 	 * Handle file descriptors
544 	 */
545 	if (flags & RFCFDG) {
546 		p2->p_fd = fdinit(p1);
547 		fdtol = NULL;
548 	} else if (flags & RFFDG) {
549 		error = fdcopy(p1, &p2->p_fd);
550 		if (error != 0) {
551 			error = ENOMEM;
552 			goto done;
553 		}
554 		fdtol = NULL;
555 	} else {
556 		p2->p_fd = fdshare(p1);
557 		if (p1->p_fdtol == NULL) {
558 			p1->p_fdtol = filedesc_to_leader_alloc(NULL,
559 							       p1->p_leader);
560 		}
561 		if ((flags & RFTHREAD) != 0) {
562 			/*
563 			 * Shared file descriptor table and
564 			 * shared process leaders.
565 			 */
566 			fdtol = p1->p_fdtol;
567 			fdtol->fdl_refcount++;
568 		} else {
569 			/*
570 			 * Shared file descriptor table, and
571 			 * different process leaders
572 			 */
573 			fdtol = filedesc_to_leader_alloc(p1->p_fdtol, p2);
574 		}
575 	}
576 	p2->p_fdtol = fdtol;
577 	p2->p_limit = plimit_fork(p1);
578 
579 	/*
580 	 * Adjust depth for resource downscaling
581 	 */
582 	if ((p2->p_depth & 31) != 31)
583 		++p2->p_depth;
584 
585 	/*
586 	 * Preserve some more flags in subprocess.  P_PROFIL has already
587 	 * been preserved.
588 	 */
589 	p2->p_flags |= p1->p_flags & P_SUGID;
590 	if (p1->p_session->s_ttyvp != NULL && (p1->p_flags & P_CONTROLT))
591 		p2->p_flags |= P_CONTROLT;
592 	if (flags & RFPPWAIT) {
593 		p2->p_flags |= P_PPWAIT;
594 		if (p1->p_upmap)
595 			atomic_add_int(&p1->p_upmap->invfork, 1);
596 	}
597 
598 	/*
599 	 * Inherit the virtual kernel structure (allows a virtual kernel
600 	 * to fork to simulate multiple cpus).
601 	 */
602 	if (p1->p_vkernel)
603 		vkernel_inherit(p1, p2);
604 
605 	/*
606 	 * Once we are on a pglist we may receive signals.  XXX we might
607 	 * race a ^C being sent to the process group by not receiving it
608 	 * at all prior to this line.
609 	 */
610 	pgref(p1grp);
611 	lwkt_gettoken(&p1grp->pg_token);
612 	LIST_INSERT_AFTER(p1, p2, p_pglist);
613 	lwkt_reltoken(&p1grp->pg_token);
614 
615 	/*
616 	 * Attach the new process to its parent.
617 	 *
618 	 * If RFNOWAIT is set, the newly created process becomes a child
619 	 * of the reaper (typically init).  This effectively disassociates
620 	 * the child from the parent.
621 	 *
622 	 * Temporarily hold pptr for the RFNOWAIT case to avoid ripouts.
623 	 */
624 	if (flags & RFNOWAIT) {
625 		pptr = reaper_get(reap);
626 		if (pptr == NULL) {
627 			pptr = initproc;
628 			PHOLD(pptr);
629 		}
630 	} else {
631 		pptr = p1;
632 	}
633 	p2->p_pptr = pptr;
634 	p2->p_ppid = pptr->p_pid;
635 	LIST_INIT(&p2->p_children);
636 
637 	lwkt_gettoken(&pptr->p_token);
638 	LIST_INSERT_HEAD(&pptr->p_children, p2, p_sibling);
639 	lwkt_reltoken(&pptr->p_token);
640 
641 	if (flags & RFNOWAIT)
642 		PRELE(pptr);
643 
644 	varsymset_init(&p2->p_varsymset, &p1->p_varsymset);
645 	callout_init_mp(&p2->p_ithandle);
646 
647 #ifdef KTRACE
648 	/*
649 	 * Copy traceflag and tracefile if enabled.  If not inherited,
650 	 * these were zeroed above but we still could have a trace race
651 	 * so make sure p2's p_tracenode is NULL.
652 	 */
653 	if ((p1->p_traceflag & KTRFAC_INHERIT) && p2->p_tracenode == NULL) {
654 		p2->p_traceflag = p1->p_traceflag;
655 		p2->p_tracenode = ktrinherit(p1->p_tracenode);
656 	}
657 #endif
658 
659 	/*
660 	 * This begins the section where we must prevent the parent
661 	 * from being messed with too heavily while we run through the
662 	 * fork operation.
663 	 *
664 	 * Gets PRELE'd in the caller in start_forked_proc().
665 	 *
666 	 * Create the first lwp associated with the new proc.  It will
667 	 * return via a different execution path later, directly into
668 	 * userland, after it was put on the runq by start_forked_proc().
669 	 */
670 	PHOLD(p1);
671 
672 	lp2 = lwp_fork1(lp1, p2, flags, NULL);
673 	vm_fork(p1, p2, lp2, flags);
674 	if ((flags & RFMEM) == 0)
675 		wake_umtx_threads(p1);
676 	lwp_fork2(lp1, p2, lp2, flags);
677 
678 	if (flags == (RFFDG | RFPROC | RFPGLOCK)) {
679 		mycpu->gd_cnt.v_forks++;
680 		mycpu->gd_cnt.v_forkpages += btoc(p2->p_vmspace->vm_dsize) +
681 					     btoc(p2->p_vmspace->vm_ssize);
682 	} else if (flags == (RFFDG | RFPROC | RFPPWAIT | RFMEM | RFPGLOCK)) {
683 		mycpu->gd_cnt.v_vforks++;
684 		mycpu->gd_cnt.v_vforkpages += btoc(p2->p_vmspace->vm_dsize) +
685 					      btoc(p2->p_vmspace->vm_ssize);
686 	} else if (p1 == &proc0) {
687 		mycpu->gd_cnt.v_kthreads++;
688 		mycpu->gd_cnt.v_kthreadpages += btoc(p2->p_vmspace->vm_dsize) +
689 						btoc(p2->p_vmspace->vm_ssize);
690 	} else {
691 		mycpu->gd_cnt.v_rforks++;
692 		mycpu->gd_cnt.v_rforkpages += btoc(p2->p_vmspace->vm_dsize) +
693 					      btoc(p2->p_vmspace->vm_ssize);
694 	}
695 
696 	/*
697 	 * Both processes are set up, now check if any loadable modules want
698 	 * to adjust anything.
699 	 *   What if they have an error? XXX
700 	 */
701 	TAILQ_FOREACH(ep, &fork_list, next) {
702 		(*ep->function)(p1, p2, flags);
703 	}
704 
705 	/*
706 	 * Set the start time.  Note that the process is not runnable.  The
707 	 * caller is responsible for making it runnable.
708 	 */
709 	microtime(&p2->p_start);
710 	p2->p_acflag = AFORK;
711 
712 	/*
713 	 * tell any interested parties about the new process
714 	 */
715 	KNOTE(&p1->p_klist, NOTE_FORK | p2->p_pid);
716 
717 	/*
718 	 * Return child proc pointer to parent.
719 	 */
720 	*procp = p2;
721 	error = 0;
722 done:
723 	if (p2)
724 		lwkt_reltoken(&p2->p_token);
725 	lwkt_reltoken(&p1->p_token);
726 	if (plkgrp) {
727 		lockmgr(&plkgrp->pg_lock, LK_RELEASE);
728 		pgrel(plkgrp);
729 	}
730 	return (error);
731 }
732 
733 /*
734  * The first part of lwp_fork*() allocates enough of the new lwp that
735  * vm_fork() can use it to deal with /dev/lpmap mappings.
736  */
737 static struct lwp *
738 lwp_fork1(struct lwp *lp1, struct proc *destproc, int flags,
739 	 const cpumask_t *mask)
740 {
741 	struct lwp *lp2;
742 
743 	lp2 = kmalloc(sizeof(struct lwp), M_LWP, M_WAITOK|M_ZERO);
744 	lp2->lwp_proc = destproc;
745 	lp2->lwp_stat = LSRUN;
746 	bcopy(&lp1->lwp_startcopy, &lp2->lwp_startcopy,
747 	    (unsigned) ((caddr_t)&lp2->lwp_endcopy -
748 			(caddr_t)&lp2->lwp_startcopy));
749 	if (mask != NULL)
750 		lp2->lwp_cpumask = *mask;
751 
752 	lwkt_token_init(&lp2->lwp_token, "lwp_token");
753 	TAILQ_INIT(&lp2->lwp_lpmap_backing_list);
754 	spin_init(&lp2->lwp_spin, "lwptoken");
755 
756 	/*
757 	 * Use the same TID for the first thread in the new process after
758 	 * a fork or vfork.  This is needed to keep pthreads and /dev/lpmap
759 	 * sane.  In particular a consequence of implementing the per-thread
760 	 * /dev/lpmap map code makes this mandatory.
761 	 *
762 	 * NOTE: exec*() will reset the TID to 1 to keep things sane in that
763 	 *	 department too.
764 	 *
765 	 * NOTE: In the case of lwp_create(), this TID represents a conflict
766 	 *	 which will be resolved in lwp_fork2(), but in the case of
767 	 *	 a fork(), the TID has to be correct or vm_fork() will not
768 	 *	 keep the correct lpmap.
769 	 */
770 	lp2->lwp_tid = lp1->lwp_tid;
771 
772 	return lp2;
773 }
774 
775 /*
776  * The second part of lwp_fork*()
777  */
778 static void
779 lwp_fork2(struct lwp *lp1, struct proc *destproc, struct lwp *lp2, int flags)
780 {
781 	globaldata_t gd = mycpu;
782 	struct thread *td2;
783 
784 	lp2->lwp_vmspace = destproc->p_vmspace;
785 
786 	/*
787 	 * Reset the sigaltstack if memory is shared, otherwise inherit
788 	 * it.
789 	 */
790 	if (flags & RFMEM) {
791 		lp2->lwp_sigstk.ss_flags = SS_DISABLE;
792 		lp2->lwp_sigstk.ss_size = 0;
793 		lp2->lwp_sigstk.ss_sp = NULL;
794 		lp2->lwp_flags &= ~LWP_ALTSTACK;
795 	} else {
796 		lp2->lwp_flags |= lp1->lwp_flags & LWP_ALTSTACK;
797 	}
798 
799 	/*
800 	 * Set cpbase to the last timeout that occured (not the upcoming
801 	 * timeout).
802 	 *
803 	 * A critical section is required since a timer IPI can update
804 	 * scheduler specific data.
805 	 */
806 	crit_enter();
807 	lp2->lwp_cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
808 	destproc->p_usched->heuristic_forking(lp1, lp2);
809 	crit_exit();
810 	CPUMASK_ANDMASK(lp2->lwp_cpumask, usched_mastermask);
811 
812 	/*
813 	 * Assign the thread to the current cpu to begin with so we
814 	 * can manipulate it.
815 	 */
816 	td2 = lwkt_alloc_thread(NULL, LWKT_THREAD_STACK, gd->gd_cpuid, 0);
817 	lp2->lwp_thread = td2;
818 	td2->td_wakefromcpu = gd->gd_cpuid;
819 	td2->td_ucred = crhold(destproc->p_ucred);
820 	td2->td_proc = destproc;
821 	td2->td_lwp = lp2;
822 	td2->td_switch = cpu_heavy_switch;
823 #ifdef NO_LWKT_SPLIT_USERPRI
824 	lwkt_setpri(td2, TDPRI_USER_NORM);
825 #else
826 	lwkt_setpri(td2, TDPRI_KERN_USER);
827 #endif
828 	lwkt_set_comm(td2, "%s", destproc->p_comm);
829 
830 	/*
831 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
832 	 * and make the child ready to run.
833 	 */
834 	cpu_fork(lp1, lp2, flags);
835 	kqueue_init(&lp2->lwp_kqueue, destproc->p_fd);
836 
837 	/*
838 	 * Associate the new thread with destproc, after we've set most of
839 	 * it up and gotten its related td2 installed.  Otherwise we can
840 	 * race other random kernel code that iterates LWPs and expects the
841 	 * thread to be assigned.
842 	 *
843 	 * Leave 2 bits open so the pthreads library can optimize locks
844 	 * by combining the TID with a few Lock-related flags.
845 	 */
846 	while (lwp_rb_tree_RB_INSERT(&destproc->p_lwp_tree, lp2) != NULL) {
847 		++lp2->lwp_tid;
848 		if (lp2->lwp_tid == 0 || lp2->lwp_tid == 0x3FFFFFFF)
849 			lp2->lwp_tid = 1;
850 	}
851 
852 	destproc->p_lasttid = lp2->lwp_tid;
853 	destproc->p_nthreads++;
854 
855 	/*
856 	 * This flag is set and never cleared.  It means that the process
857 	 * was threaded at some point.  Used to improve exit performance.
858 	 */
859 	pmap_maybethreaded(&destproc->p_vmspace->vm_pmap);
860 	destproc->p_flags |= P_MAYBETHREADED;
861 
862 	/*
863 	 * If the original lp had a lpmap and a non-zero blockallsigs
864 	 * count, give the lp for the forked process the same count.
865 	 *
866 	 * This makes the user code and expectations less confusing
867 	 * in terms of unwinding locks and also allows userland to start
868 	 * the forked process with signals blocked via the blockallsigs()
869 	 * mechanism if desired.
870 	 */
871 	if (lp1->lwp_lpmap &&
872 	    (lp1->lwp_lpmap->blockallsigs & 0x7FFFFFFF)) {
873 		lwp_usermap(lp2, 0);
874 		if (lp2->lwp_lpmap) {
875 			lp2->lwp_lpmap->blockallsigs =
876 				lp1->lwp_lpmap->blockallsigs;
877 		}
878 	}
879 }
880 
881 /*
882  * The next two functionms are general routines to handle adding/deleting
883  * items on the fork callout list.
884  *
885  * at_fork():
886  * Take the arguments given and put them onto the fork callout list,
887  * However first make sure that it's not already there.
888  * Returns 0 on success or a standard error number.
889  */
890 int
891 at_fork(forklist_fn function)
892 {
893 	struct forklist *ep;
894 
895 #ifdef INVARIANTS
896 	/* let the programmer know if he's been stupid */
897 	if (rm_at_fork(function)) {
898 		kprintf("WARNING: fork callout entry (%p) already present\n",
899 		    function);
900 	}
901 #endif
902 	ep = kmalloc(sizeof(*ep), M_ATFORK, M_WAITOK|M_ZERO);
903 	ep->function = function;
904 	TAILQ_INSERT_TAIL(&fork_list, ep, next);
905 	return (0);
906 }
907 
908 /*
909  * Scan the exit callout list for the given item and remove it..
910  * Returns the number of items removed (0 or 1)
911  */
912 int
913 rm_at_fork(forklist_fn function)
914 {
915 	struct forklist *ep;
916 
917 	TAILQ_FOREACH(ep, &fork_list, next) {
918 		if (ep->function == function) {
919 			TAILQ_REMOVE(&fork_list, ep, next);
920 			kfree(ep, M_ATFORK);
921 			return(1);
922 		}
923 	}
924 	return (0);
925 }
926 
927 /*
928  * Add a forked process to the run queue after any remaining setup, such
929  * as setting the fork handler, has been completed.
930  *
931  * p2 is held by the caller.
932  */
933 void
934 start_forked_proc(struct lwp *lp1, struct proc *p2)
935 {
936 	struct lwp *lp2 = ONLY_LWP_IN_PROC(p2);
937 	int pflags;
938 
939 	/*
940 	 * Move from SIDL to RUN queue, and activate the process's thread.
941 	 * Activation of the thread effectively makes the process "a"
942 	 * current process, so we do not setrunqueue().
943 	 *
944 	 * YYY setrunqueue works here but we should clean up the trampoline
945 	 * code so we just schedule the LWKT thread and let the trampoline
946 	 * deal with the userland scheduler on return to userland.
947 	 */
948 	KASSERT(p2->p_stat == SIDL,
949 	    ("cannot start forked process, bad status: %p", p2));
950 	p2->p_usched->resetpriority(lp2);
951 	crit_enter();
952 	p2->p_stat = SACTIVE;
953 	lp2->lwp_stat = LSRUN;
954 	p2->p_usched->setrunqueue(lp2);
955 	crit_exit();
956 
957 	/*
958 	 * Now can be swapped.
959 	 */
960 	PRELE(lp1->lwp_proc);
961 
962 	/*
963 	 * Preserve synchronization semantics of vfork.  P_PPWAIT is set in
964 	 * the child until it has retired the parent's resources.  The parent
965 	 * must wait for the flag to be cleared by the child.
966 	 *
967 	 * Interlock the flag/tsleep with atomic ops to avoid unnecessary
968 	 * p_token conflicts.
969 	 *
970 	 * XXX Is this use of an atomic op on a field that is not normally
971 	 *     manipulated with atomic ops ok?
972 	 */
973 	while ((pflags = p2->p_flags) & P_PPWAIT) {
974 		cpu_ccfence();
975 		tsleep_interlock(lp1->lwp_proc, 0);
976 		if (atomic_cmpset_int(&p2->p_flags, pflags, pflags))
977 			tsleep(lp1->lwp_proc, PINTERLOCKED, "ppwait", 0);
978 	}
979 }
980 
981 /*
982  * procctl (idtype_t idtype, id_t id, int cmd, void *arg)
983  */
984 int
985 sys_procctl(struct sysmsg *sysmsg, const struct procctl_args *uap)
986 {
987 	struct proc *p = curproc;
988 	struct proc *p2;
989 	struct sysreaper *reap;
990 	union reaper_info udata;
991 	int error;
992 
993 	if (uap->idtype != P_PID || uap->id != (id_t)p->p_pid)
994 		return EINVAL;
995 
996 	switch(uap->cmd) {
997 	case PROC_REAP_ACQUIRE:
998 		lwkt_gettoken(&p->p_token);
999 		reap = kmalloc(sizeof(*reap), M_REAPER, M_WAITOK|M_ZERO);
1000 		if (p->p_reaper == NULL || p->p_reaper->p != p) {
1001 			reaper_init(p, reap);
1002 			error = 0;
1003 		} else {
1004 			kfree(reap, M_REAPER);
1005 			error = EALREADY;
1006 		}
1007 		lwkt_reltoken(&p->p_token);
1008 		break;
1009 	case PROC_REAP_RELEASE:
1010 		lwkt_gettoken(&p->p_token);
1011 release_again:
1012 		reap = p->p_reaper;
1013 		KKASSERT(reap != NULL);
1014 		if (reap->p == p) {
1015 			reaper_hold(reap);	/* in case of thread race */
1016 			lockmgr(&reap->lock, LK_EXCLUSIVE);
1017 			if (reap->p != p) {
1018 				lockmgr(&reap->lock, LK_RELEASE);
1019 				reaper_drop(reap);
1020 				goto release_again;
1021 			}
1022 			reap->p = NULL;
1023 			p->p_reaper = reap->parent;
1024 			if (p->p_reaper)
1025 				reaper_hold(p->p_reaper);
1026 			lockmgr(&reap->lock, LK_RELEASE);
1027 			reaper_drop(reap);	/* our ref */
1028 			reaper_drop(reap);	/* old p_reaper ref */
1029 			error = 0;
1030 		} else {
1031 			error = ENOTCONN;
1032 		}
1033 		lwkt_reltoken(&p->p_token);
1034 		break;
1035 	case PROC_REAP_STATUS:
1036 		bzero(&udata, sizeof(udata));
1037 		lwkt_gettoken_shared(&p->p_token);
1038 		if ((reap = p->p_reaper) != NULL && reap->p == p) {
1039 			udata.status.flags = reap->flags;
1040 			udata.status.refs = reap->refs - 1; /* minus ours */
1041 		}
1042 		p2 = LIST_FIRST(&p->p_children);
1043 		udata.status.pid_head = p2 ? p2->p_pid : -1;
1044 		lwkt_reltoken(&p->p_token);
1045 
1046 		if (uap->data) {
1047 			error = copyout(&udata, uap->data,
1048 					sizeof(udata.status));
1049 		} else {
1050 			error = 0;
1051 		}
1052 		break;
1053 	default:
1054 		error = EINVAL;
1055 		break;
1056 	}
1057 	return error;
1058 }
1059 
1060 /*
1061  * Bump ref on reaper, preventing destruction
1062  */
1063 void
1064 reaper_hold(struct sysreaper *reap)
1065 {
1066 	KKASSERT(reap->refs > 0);
1067 	refcount_acquire(&reap->refs);
1068 }
1069 
1070 /*
1071  * Drop ref on reaper, destroy the structure on the 1->0
1072  * transition and loop on the parent.
1073  */
1074 void
1075 reaper_drop(struct sysreaper *next)
1076 {
1077 	struct sysreaper *reap;
1078 
1079 	while ((reap = next) != NULL) {
1080 		if (refcount_release(&reap->refs)) {
1081 			next = reap->parent;
1082 			KKASSERT(reap->p == NULL);
1083 			lockmgr(&reaper_lock, LK_EXCLUSIVE);
1084 			reap->parent = NULL;
1085 			kfree(reap, M_REAPER);
1086 			lockmgr(&reaper_lock, LK_RELEASE);
1087 		} else {
1088 			next = NULL;
1089 		}
1090 	}
1091 }
1092 
1093 /*
1094  * Initialize a static or newly allocated reaper structure
1095  */
1096 void
1097 reaper_init(struct proc *p, struct sysreaper *reap)
1098 {
1099 	reap->parent = p->p_reaper;
1100 	reap->p = p;
1101 	if (p == initproc) {
1102 		reap->flags = REAPER_STAT_OWNED | REAPER_STAT_REALINIT;
1103 		reap->refs = 2;
1104 	} else {
1105 		reap->flags = REAPER_STAT_OWNED;
1106 		reap->refs = 1;
1107 	}
1108 	lockinit(&reap->lock, "subrp", 0, 0);
1109 	cpu_sfence();
1110 	p->p_reaper = reap;
1111 }
1112 
1113 /*
1114  * Called with p->p_token held during exit.
1115  *
1116  * This is a bit simpler than RELEASE because there are no threads remaining
1117  * to race.  We only release if we own the reaper, the exit code will handle
1118  * the final p_reaper release.
1119  */
1120 struct sysreaper *
1121 reaper_exit(struct proc *p)
1122 {
1123 	struct sysreaper *reap;
1124 
1125 	/*
1126 	 * Release acquired reaper
1127 	 */
1128 	if ((reap = p->p_reaper) != NULL && reap->p == p) {
1129 		lockmgr(&reap->lock, LK_EXCLUSIVE);
1130 		p->p_reaper = reap->parent;
1131 		if (p->p_reaper)
1132 			reaper_hold(p->p_reaper);
1133 		reap->p = NULL;
1134 		lockmgr(&reap->lock, LK_RELEASE);
1135 		reaper_drop(reap);
1136 	}
1137 
1138 	/*
1139 	 * Return and clear reaper (caller is holding p_token for us)
1140 	 * (reap->p does not equal p).  Caller must drop it.
1141 	 */
1142 	if ((reap = p->p_reaper) != NULL) {
1143 		p->p_reaper = NULL;
1144 	}
1145 	return reap;
1146 }
1147 
1148 /*
1149  * Return a held (PHOLD) process representing the reaper for process (p).
1150  * NULL should not normally be returned.  Caller should PRELE() the returned
1151  * reaper process when finished.
1152  *
1153  * Remove dead internal nodes while we are at it.
1154  *
1155  * Process (p)'s token must be held on call.
1156  * The returned process's token is NOT acquired by this routine.
1157  */
1158 struct proc *
1159 reaper_get(struct sysreaper *reap)
1160 {
1161 	struct sysreaper *next;
1162 	struct proc *reproc;
1163 
1164 	if (reap == NULL)
1165 		return NULL;
1166 
1167 	/*
1168 	 * Extra hold for loop
1169 	 */
1170 	reaper_hold(reap);
1171 
1172 	while (reap) {
1173 		lockmgr(&reap->lock, LK_SHARED);
1174 		if (reap->p) {
1175 			/*
1176 			 * Probable reaper
1177 			 */
1178 			if (reap->p) {
1179 				reproc = reap->p;
1180 				PHOLD(reproc);
1181 				lockmgr(&reap->lock, LK_RELEASE);
1182 				reaper_drop(reap);
1183 				return reproc;
1184 			}
1185 
1186 			/*
1187 			 * Raced, try again
1188 			 */
1189 			lockmgr(&reap->lock, LK_RELEASE);
1190 			continue;
1191 		}
1192 
1193 		/*
1194 		 * Traverse upwards in the reaper topology, destroy
1195 		 * dead internal nodes when possible.
1196 		 *
1197 		 * NOTE: Our ref on next means that a dead node should
1198 		 *	 have 2 (ours and reap->parent's).
1199 		 */
1200 		next = reap->parent;
1201 		while (next) {
1202 			reaper_hold(next);
1203 			if (next->refs == 2 && next->p == NULL) {
1204 				lockmgr(&reap->lock, LK_RELEASE);
1205 				lockmgr(&reap->lock, LK_EXCLUSIVE);
1206 				if (next->refs == 2 &&
1207 				    reap->parent == next &&
1208 				    next->p == NULL) {
1209 					/*
1210 					 * reap->parent inherits ref from next.
1211 					 */
1212 					reap->parent = next->parent;
1213 					next->parent = NULL;
1214 					reaper_drop(next);	/* ours */
1215 					reaper_drop(next);	/* old parent */
1216 					next = reap->parent;
1217 					continue;	/* possible chain */
1218 				}
1219 			}
1220 			break;
1221 		}
1222 		lockmgr(&reap->lock, LK_RELEASE);
1223 		reaper_drop(reap);
1224 		reap = next;
1225 	}
1226 	return NULL;
1227 }
1228 
1229 /*
1230  * Test that the sender is allowed to send a signal to the target.
1231  * The sender process is assumed to have a stable reaper.  The
1232  * target can be e.g. from a scan callback.
1233  *
1234  * Target cannot be the reaper process itself unless reaper_ok is specified,
1235  * or sender == target.
1236  */
1237 int
1238 reaper_sigtest(struct proc *sender, struct proc *target, int reaper_ok)
1239 {
1240 	struct sysreaper *sreap;
1241 	struct sysreaper *reap;
1242 	int r;
1243 
1244 	sreap = sender->p_reaper;
1245 	if (sreap == NULL)
1246 		return 1;
1247 
1248 	if (sreap == target->p_reaper) {
1249 		if (sreap->p == target && sreap->p != sender && reaper_ok == 0)
1250 			return 0;
1251 		return 1;
1252 	}
1253 	lockmgr(&reaper_lock, LK_SHARED);
1254 	r = 0;
1255 	for (reap = target->p_reaper; reap; reap = reap->parent) {
1256 		if (sreap == reap) {
1257 			if (sreap->p != target || reaper_ok)
1258 				r = 1;
1259 			break;
1260 		}
1261 	}
1262 	lockmgr(&reaper_lock, LK_RELEASE);
1263 
1264 	return r;
1265 }
1266