xref: /dflybsd-src/sys/kern/kern_fork.c (revision 1fe7e945f3685215c267632222c9bb8c33f052ec)
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/sysproto.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_fork(struct lwp *, struct proc *, int flags,
86 			    const cpumask_t *mask);
87 static int		lwp_create1(struct lwp_params *params,
88 			    const cpumask_t *mask);
89 static struct lock reaper_lock = LOCK_INITIALIZER("reapgl", 0, 0);
90 
91 int forksleep; /* Place for fork1() to sleep on. */
92 
93 /*
94  * Red-Black tree support for LWPs
95  */
96 
97 static int
98 rb_lwp_compare(struct lwp *lp1, struct lwp *lp2)
99 {
100 	if (lp1->lwp_tid < lp2->lwp_tid)
101 		return(-1);
102 	if (lp1->lwp_tid > lp2->lwp_tid)
103 		return(1);
104 	return(0);
105 }
106 
107 RB_GENERATE2(lwp_rb_tree, lwp, u.lwp_rbnode, rb_lwp_compare, lwpid_t, lwp_tid);
108 
109 /*
110  * fork() system call
111  */
112 int
113 sys_fork(struct fork_args *uap)
114 {
115 	struct lwp *lp = curthread->td_lwp;
116 	struct proc *p2;
117 	int error;
118 
119 	error = fork1(lp, RFFDG | RFPROC | RFPGLOCK, &p2);
120 	if (error == 0) {
121 		PHOLD(p2);
122 		start_forked_proc(lp, p2);
123 		uap->sysmsg_fds[0] = p2->p_pid;
124 		uap->sysmsg_fds[1] = 0;
125 		PRELE(p2);
126 	}
127 	return error;
128 }
129 
130 /*
131  * vfork() system call
132  */
133 int
134 sys_vfork(struct vfork_args *uap)
135 {
136 	struct lwp *lp = curthread->td_lwp;
137 	struct proc *p2;
138 	int error;
139 
140 	error = fork1(lp, RFFDG | RFPROC | RFPPWAIT | RFMEM | RFPGLOCK, &p2);
141 	if (error == 0) {
142 		PHOLD(p2);
143 		start_forked_proc(lp, p2);
144 		uap->sysmsg_fds[0] = p2->p_pid;
145 		uap->sysmsg_fds[1] = 0;
146 		PRELE(p2);
147 	}
148 	return error;
149 }
150 
151 /*
152  * Handle rforks.  An rfork may (1) operate on the current process without
153  * creating a new, (2) create a new process that shared the current process's
154  * vmspace, signals, and/or descriptors, or (3) create a new process that does
155  * not share these things (normal fork).
156  *
157  * Note that we only call start_forked_proc() if a new process is actually
158  * created.
159  *
160  * rfork { int flags }
161  */
162 int
163 sys_rfork(struct rfork_args *uap)
164 {
165 	struct lwp *lp = curthread->td_lwp;
166 	struct proc *p2;
167 	int error;
168 
169 	if ((uap->flags & RFKERNELONLY) != 0)
170 		return (EINVAL);
171 
172 	error = fork1(lp, uap->flags | RFPGLOCK, &p2);
173 	if (error == 0) {
174 		if (p2) {
175 			PHOLD(p2);
176 			start_forked_proc(lp, p2);
177 			uap->sysmsg_fds[0] = p2->p_pid;
178 			uap->sysmsg_fds[1] = 0;
179 			PRELE(p2);
180 		} else {
181 			uap->sysmsg_fds[0] = 0;
182 			uap->sysmsg_fds[1] = 0;
183 		}
184 	}
185 	return error;
186 }
187 
188 static int
189 lwp_create1(struct lwp_params *uprm, const cpumask_t *umask)
190 {
191 	struct proc *p = curproc;
192 	struct lwp *lp;
193 	struct lwp_params params;
194 	cpumask_t *mask = NULL, mask0;
195 	int error;
196 
197 	error = copyin(uprm, &params, sizeof(params));
198 	if (error)
199 		goto fail2;
200 
201 	if (umask != NULL) {
202 		error = copyin(umask, &mask0, sizeof(mask0));
203 		if (error)
204 			goto fail2;
205 		CPUMASK_ANDMASK(mask0, smp_active_mask);
206 		if (CPUMASK_TESTNZERO(mask0))
207 			mask = &mask0;
208 	}
209 
210 	lwkt_gettoken(&p->p_token);
211 	plimit_lwp_fork(p);	/* force exclusive access */
212 	lp = lwp_fork(curthread->td_lwp, p, RFPROC | RFMEM, mask);
213 	error = cpu_prepare_lwp(lp, &params);
214 	if (error)
215 		goto fail;
216 	if (params.lwp_tid1 != NULL &&
217 	    (error = copyout(&lp->lwp_tid, params.lwp_tid1, sizeof(lp->lwp_tid))))
218 		goto fail;
219 	if (params.lwp_tid2 != NULL &&
220 	    (error = copyout(&lp->lwp_tid, params.lwp_tid2, sizeof(lp->lwp_tid))))
221 		goto fail;
222 
223 	/*
224 	 * Now schedule the new lwp.
225 	 */
226 	p->p_usched->resetpriority(lp);
227 	crit_enter();
228 	lp->lwp_stat = LSRUN;
229 	p->p_usched->setrunqueue(lp);
230 	crit_exit();
231 	lwkt_reltoken(&p->p_token);
232 
233 	return (0);
234 
235 fail:
236 	/*
237 	 * Make sure no one is using this lwp, before it is removed from
238 	 * the tree.  If we didn't wait it here, lwp tree iteration with
239 	 * blocking operation would be broken.
240 	 */
241 	while (lp->lwp_lock > 0)
242 		tsleep(lp, 0, "lwpfail", 1);
243 	lwp_rb_tree_RB_REMOVE(&p->p_lwp_tree, lp);
244 	--p->p_nthreads;
245 	/* lwp_dispose expects an exited lwp, and a held proc */
246 	atomic_set_int(&lp->lwp_mpflags, LWP_MP_WEXIT);
247 	lp->lwp_thread->td_flags |= TDF_EXITING;
248 	lwkt_remove_tdallq(lp->lwp_thread);
249 	PHOLD(p);
250 	biosched_done(lp->lwp_thread);
251 	dsched_exit_thread(lp->lwp_thread);
252 	lwp_dispose(lp);
253 	lwkt_reltoken(&p->p_token);
254 fail2:
255 	return (error);
256 }
257 
258 /*
259  * Low level thread create used by pthreads.
260  */
261 int
262 sys_lwp_create(struct lwp_create_args *uap)
263 {
264 
265 	return (lwp_create1(uap->params, NULL));
266 }
267 
268 int
269 sys_lwp_create2(struct lwp_create2_args *uap)
270 {
271 
272 	return (lwp_create1(uap->params, uap->mask));
273 }
274 
275 int	nprocs = 1;		/* process 0 */
276 
277 int
278 fork1(struct lwp *lp1, int flags, struct proc **procp)
279 {
280 	struct proc *p1 = lp1->lwp_proc;
281 	struct proc *p2;
282 	struct proc *pptr;
283 	struct pgrp *p1grp;
284 	struct pgrp *plkgrp;
285 	struct sysreaper *reap;
286 	uid_t uid;
287 	int ok, error;
288 	static int curfail = 0;
289 	static struct timeval lastfail;
290 	struct forklist *ep;
291 	struct filedesc_to_leader *fdtol;
292 
293 	if ((flags & (RFFDG|RFCFDG)) == (RFFDG|RFCFDG))
294 		return (EINVAL);
295 
296 	lwkt_gettoken(&p1->p_token);
297 	plkgrp = NULL;
298 	p2 = NULL;
299 
300 	/*
301 	 * Here we don't create a new process, but we divorce
302 	 * certain parts of a process from itself.
303 	 */
304 	if ((flags & RFPROC) == 0) {
305 		/*
306 		 * This kind of stunt does not work anymore if
307 		 * there are native threads (lwps) running
308 		 */
309 		if (p1->p_nthreads != 1) {
310 			error = EINVAL;
311 			goto done;
312 		}
313 
314 		vm_fork(p1, 0, flags);
315 
316 		/*
317 		 * Close all file descriptors.
318 		 */
319 		if (flags & RFCFDG) {
320 			struct filedesc *fdtmp;
321 			fdtmp = fdinit(p1);
322 			fdfree(p1, fdtmp);
323 		}
324 
325 		/*
326 		 * Unshare file descriptors (from parent.)
327 		 */
328 		if (flags & RFFDG) {
329 			if (p1->p_fd->fd_refcnt > 1) {
330 				struct filedesc *newfd;
331 				error = fdcopy(p1, &newfd);
332 				if (error != 0) {
333 					error = ENOMEM;
334 					goto done;
335 				}
336 				fdfree(p1, newfd);
337 			}
338 		}
339 		*procp = NULL;
340 		error = 0;
341 		goto done;
342 	}
343 
344 	/*
345 	 * Interlock against process group signal delivery.  If signals
346 	 * are pending after the interlock is obtained we have to restart
347 	 * the system call to process the signals.  If we don't the child
348 	 * can miss a pgsignal (such as ^C) sent during the fork.
349 	 *
350 	 * We can't use CURSIG() here because it will process any STOPs
351 	 * and cause the process group lock to be held indefinitely.  If
352 	 * a STOP occurs, the fork will be restarted after the CONT.
353 	 */
354 	p1grp = p1->p_pgrp;
355 	if ((flags & RFPGLOCK) && (plkgrp = p1->p_pgrp) != NULL) {
356 		pgref(plkgrp);
357 		lockmgr(&plkgrp->pg_lock, LK_SHARED);
358 		if (CURSIG_NOBLOCK(lp1)) {
359 			error = ERESTART;
360 			goto done;
361 		}
362 	}
363 
364 	/*
365 	 * Although process entries are dynamically created, we still keep
366 	 * a global limit on the maximum number we will create.  Don't allow
367 	 * a nonprivileged user to use the last ten processes; don't let root
368 	 * exceed the limit. The variable nprocs is the current number of
369 	 * processes, maxproc is the limit.
370 	 */
371 	uid = lp1->lwp_thread->td_ucred->cr_ruid;
372 	if ((nprocs >= maxproc - 10 && uid != 0) || nprocs >= maxproc) {
373 		if (ppsratecheck(&lastfail, &curfail, 1))
374 			kprintf("maxproc limit exceeded by uid %d, please "
375 			       "see tuning(7) and login.conf(5).\n", uid);
376 		tsleep(&forksleep, 0, "fork", hz / 2);
377 		error = EAGAIN;
378 		goto done;
379 	}
380 
381 	/*
382 	 * Increment the nprocs resource before blocking can occur.  There
383 	 * are hard-limits as to the number of processes that can run.
384 	 */
385 	atomic_add_int(&nprocs, 1);
386 
387 	/*
388 	 * Increment the count of procs running with this uid.  This also
389 	 * applies to root.
390 	 */
391 	ok = chgproccnt(lp1->lwp_thread->td_ucred->cr_ruidinfo, 1,
392 			plimit_getadjvalue(RLIMIT_NPROC));
393 	if (!ok) {
394 		/*
395 		 * Back out the process count
396 		 */
397 		atomic_add_int(&nprocs, -1);
398 		if (ppsratecheck(&lastfail, &curfail, 1)) {
399 			kprintf("maxproc limit of %jd "
400 				"exceeded by \"%s\" uid %d, "
401 				"please see tuning(7) and login.conf(5).\n",
402 				plimit_getadjvalue(RLIMIT_NPROC),
403 				p1->p_comm,
404 				uid);
405 		}
406 		tsleep(&forksleep, 0, "fork", hz / 2);
407 		error = EAGAIN;
408 		goto done;
409 	}
410 
411 	/*
412 	 * Allocate a new process, don't get fancy: zero the structure.
413 	 */
414 	p2 = kmalloc(sizeof(struct proc), M_PROC, M_WAITOK|M_ZERO);
415 
416 	/*
417 	 * Core initialization.  SIDL is a safety state that protects the
418 	 * partially initialized process once it starts getting hooked
419 	 * into system structures and becomes addressable.
420 	 *
421 	 * We must be sure to acquire p2->p_token as well, we must hold it
422 	 * once the process is on the allproc list to avoid things such
423 	 * as competing modifications to p_flags.
424 	 */
425 	mycpu->gd_forkid += ncpus;
426 	p2->p_forkid = mycpu->gd_forkid + mycpu->gd_cpuid;
427 	p2->p_lasttid = 0;	/* first tid will be 1 */
428 	p2->p_stat = SIDL;
429 
430 	/*
431 	 * NOTE: Process 0 will not have a reaper, but process 1 (init) and
432 	 *	 all other processes always will.
433 	 */
434 	if ((reap = p1->p_reaper) != NULL) {
435 		reaper_hold(reap);
436 		p2->p_reaper = reap;
437 	} else {
438 		p2->p_reaper = NULL;
439 	}
440 
441 	RB_INIT(&p2->p_lwp_tree);
442 	spin_init(&p2->p_spin, "procfork1");
443 	lwkt_token_init(&p2->p_token, "proc");
444 	lwkt_gettoken(&p2->p_token);
445 
446 	/*
447 	 * Setup linkage for kernel based threading XXX lwp.  Also add the
448 	 * process to the allproclist.
449 	 *
450 	 * The process structure is addressable after this point.
451 	 */
452 	if (flags & RFTHREAD) {
453 		p2->p_peers = p1->p_peers;
454 		p1->p_peers = p2;
455 		p2->p_leader = p1->p_leader;
456 	} else {
457 		p2->p_leader = p2;
458 	}
459 	proc_add_allproc(p2);
460 
461 	/*
462 	 * Initialize the section which is copied verbatim from the parent.
463 	 */
464 	bcopy(&p1->p_startcopy, &p2->p_startcopy,
465 	      ((caddr_t)&p2->p_endcopy - (caddr_t)&p2->p_startcopy));
466 
467 	/*
468 	 * Duplicate sub-structures as needed.  Increase reference counts
469 	 * on shared objects.
470 	 *
471 	 * NOTE: because we are now on the allproc list it is possible for
472 	 *	 other consumers to gain temporary references to p2
473 	 *	 (p2->p_lock can change).
474 	 */
475 	if (p1->p_flags & P_PROFIL)
476 		startprofclock(p2);
477 	p2->p_ucred = crhold(lp1->lwp_thread->td_ucred);
478 
479 	if (jailed(p2->p_ucred))
480 		p2->p_flags |= P_JAILED;
481 
482 	if (p2->p_args)
483 		refcount_acquire(&p2->p_args->ar_ref);
484 
485 	p2->p_usched = p1->p_usched;
486 	/* XXX: verify copy of the secondary iosched stuff */
487 	dsched_enter_proc(p2);
488 
489 	if (flags & RFSIGSHARE) {
490 		p2->p_sigacts = p1->p_sigacts;
491 		refcount_acquire(&p2->p_sigacts->ps_refcnt);
492 	} else {
493 		p2->p_sigacts = kmalloc(sizeof(*p2->p_sigacts),
494 					M_SUBPROC, M_WAITOK);
495 		bcopy(p1->p_sigacts, p2->p_sigacts, sizeof(*p2->p_sigacts));
496 		refcount_init(&p2->p_sigacts->ps_refcnt, 1);
497 	}
498 	if (flags & RFLINUXTHPN)
499 	        p2->p_sigparent = SIGUSR1;
500 	else
501 	        p2->p_sigparent = SIGCHLD;
502 
503 	/* bump references to the text vnode (for procfs) */
504 	p2->p_textvp = p1->p_textvp;
505 	if (p2->p_textvp)
506 		vref(p2->p_textvp);
507 
508 	/* copy namecache handle to the text file */
509 	if (p1->p_textnch.mount)
510 		cache_copy(&p1->p_textnch, &p2->p_textnch);
511 
512 	/*
513 	 * Handle file descriptors
514 	 */
515 	if (flags & RFCFDG) {
516 		p2->p_fd = fdinit(p1);
517 		fdtol = NULL;
518 	} else if (flags & RFFDG) {
519 		error = fdcopy(p1, &p2->p_fd);
520 		if (error != 0) {
521 			error = ENOMEM;
522 			goto done;
523 		}
524 		fdtol = NULL;
525 	} else {
526 		p2->p_fd = fdshare(p1);
527 		if (p1->p_fdtol == NULL) {
528 			p1->p_fdtol = filedesc_to_leader_alloc(NULL,
529 							       p1->p_leader);
530 		}
531 		if ((flags & RFTHREAD) != 0) {
532 			/*
533 			 * Shared file descriptor table and
534 			 * shared process leaders.
535 			 */
536 			fdtol = p1->p_fdtol;
537 			fdtol->fdl_refcount++;
538 		} else {
539 			/*
540 			 * Shared file descriptor table, and
541 			 * different process leaders
542 			 */
543 			fdtol = filedesc_to_leader_alloc(p1->p_fdtol, p2);
544 		}
545 	}
546 	p2->p_fdtol = fdtol;
547 	p2->p_limit = plimit_fork(p1);
548 
549 	/*
550 	 * Adjust depth for resource downscaling
551 	 */
552 	if ((p2->p_depth & 31) != 31)
553 		++p2->p_depth;
554 
555 	/*
556 	 * Preserve some more flags in subprocess.  P_PROFIL has already
557 	 * been preserved.
558 	 */
559 	p2->p_flags |= p1->p_flags & P_SUGID;
560 	if (p1->p_session->s_ttyvp != NULL && (p1->p_flags & P_CONTROLT))
561 		p2->p_flags |= P_CONTROLT;
562 	if (flags & RFPPWAIT) {
563 		p2->p_flags |= P_PPWAIT;
564 		if (p1->p_upmap)
565 			atomic_add_int(&p1->p_upmap->invfork, 1);
566 	}
567 
568 	/*
569 	 * Inherit the virtual kernel structure (allows a virtual kernel
570 	 * to fork to simulate multiple cpus).
571 	 */
572 	if (p1->p_vkernel)
573 		vkernel_inherit(p1, p2);
574 
575 	/*
576 	 * Once we are on a pglist we may receive signals.  XXX we might
577 	 * race a ^C being sent to the process group by not receiving it
578 	 * at all prior to this line.
579 	 */
580 	pgref(p1grp);
581 	lwkt_gettoken(&p1grp->pg_token);
582 	LIST_INSERT_AFTER(p1, p2, p_pglist);
583 	lwkt_reltoken(&p1grp->pg_token);
584 
585 	/*
586 	 * Attach the new process to its parent.
587 	 *
588 	 * If RFNOWAIT is set, the newly created process becomes a child
589 	 * of the reaper (typically init).  This effectively disassociates
590 	 * the child from the parent.
591 	 *
592 	 * Temporarily hold pptr for the RFNOWAIT case to avoid ripouts.
593 	 */
594 	if (flags & RFNOWAIT) {
595 		pptr = reaper_get(reap);
596 		if (pptr == NULL) {
597 			pptr = initproc;
598 			PHOLD(pptr);
599 		}
600 	} else {
601 		pptr = p1;
602 	}
603 	p2->p_pptr = pptr;
604 	LIST_INIT(&p2->p_children);
605 
606 	lwkt_gettoken(&pptr->p_token);
607 	LIST_INSERT_HEAD(&pptr->p_children, p2, p_sibling);
608 	lwkt_reltoken(&pptr->p_token);
609 
610 	if (flags & RFNOWAIT)
611 		PRELE(pptr);
612 
613 	varsymset_init(&p2->p_varsymset, &p1->p_varsymset);
614 	callout_init_mp(&p2->p_ithandle);
615 
616 #ifdef KTRACE
617 	/*
618 	 * Copy traceflag and tracefile if enabled.  If not inherited,
619 	 * these were zeroed above but we still could have a trace race
620 	 * so make sure p2's p_tracenode is NULL.
621 	 */
622 	if ((p1->p_traceflag & KTRFAC_INHERIT) && p2->p_tracenode == NULL) {
623 		p2->p_traceflag = p1->p_traceflag;
624 		p2->p_tracenode = ktrinherit(p1->p_tracenode);
625 	}
626 #endif
627 
628 	/*
629 	 * This begins the section where we must prevent the parent
630 	 * from being swapped.
631 	 *
632 	 * Gets PRELE'd in the caller in start_forked_proc().
633 	 */
634 	PHOLD(p1);
635 
636 	vm_fork(p1, p2, flags);
637 
638 	/*
639 	 * Create the first lwp associated with the new proc.
640 	 * It will return via a different execution path later, directly
641 	 * into userland, after it was put on the runq by
642 	 * start_forked_proc().
643 	 */
644 	lwp_fork(lp1, p2, flags, NULL);
645 
646 	if (flags == (RFFDG | RFPROC | RFPGLOCK)) {
647 		mycpu->gd_cnt.v_forks++;
648 		mycpu->gd_cnt.v_forkpages += p2->p_vmspace->vm_dsize +
649 					     p2->p_vmspace->vm_ssize;
650 	} else if (flags == (RFFDG | RFPROC | RFPPWAIT | RFMEM | RFPGLOCK)) {
651 		mycpu->gd_cnt.v_vforks++;
652 		mycpu->gd_cnt.v_vforkpages += p2->p_vmspace->vm_dsize +
653 					      p2->p_vmspace->vm_ssize;
654 	} else if (p1 == &proc0) {
655 		mycpu->gd_cnt.v_kthreads++;
656 		mycpu->gd_cnt.v_kthreadpages += p2->p_vmspace->vm_dsize +
657 						p2->p_vmspace->vm_ssize;
658 	} else {
659 		mycpu->gd_cnt.v_rforks++;
660 		mycpu->gd_cnt.v_rforkpages += p2->p_vmspace->vm_dsize +
661 					      p2->p_vmspace->vm_ssize;
662 	}
663 
664 	/*
665 	 * Both processes are set up, now check if any loadable modules want
666 	 * to adjust anything.
667 	 *   What if they have an error? XXX
668 	 */
669 	TAILQ_FOREACH(ep, &fork_list, next) {
670 		(*ep->function)(p1, p2, flags);
671 	}
672 
673 	/*
674 	 * Set the start time.  Note that the process is not runnable.  The
675 	 * caller is responsible for making it runnable.
676 	 */
677 	microtime(&p2->p_start);
678 	p2->p_acflag = AFORK;
679 
680 	/*
681 	 * tell any interested parties about the new process
682 	 */
683 	KNOTE(&p1->p_klist, NOTE_FORK | p2->p_pid);
684 
685 	/*
686 	 * Return child proc pointer to parent.
687 	 */
688 	*procp = p2;
689 	error = 0;
690 done:
691 	if (p2)
692 		lwkt_reltoken(&p2->p_token);
693 	lwkt_reltoken(&p1->p_token);
694 	if (plkgrp) {
695 		lockmgr(&plkgrp->pg_lock, LK_RELEASE);
696 		pgrel(plkgrp);
697 	}
698 	return (error);
699 }
700 
701 static struct lwp *
702 lwp_fork(struct lwp *origlp, struct proc *destproc, int flags,
703     const cpumask_t *mask)
704 {
705 	globaldata_t gd = mycpu;
706 	struct lwp *lp;
707 	struct thread *td;
708 
709 	lp = kmalloc(sizeof(struct lwp), M_LWP, M_WAITOK|M_ZERO);
710 
711 	lp->lwp_proc = destproc;
712 	lp->lwp_vmspace = destproc->p_vmspace;
713 	lp->lwp_stat = LSRUN;
714 	bcopy(&origlp->lwp_startcopy, &lp->lwp_startcopy,
715 	    (unsigned) ((caddr_t)&lp->lwp_endcopy -
716 			(caddr_t)&lp->lwp_startcopy));
717 	if (mask != NULL)
718 		lp->lwp_cpumask = *mask;
719 
720 	/*
721 	 * Reset the sigaltstack if memory is shared, otherwise inherit
722 	 * it.
723 	 */
724 	if (flags & RFMEM) {
725 		lp->lwp_sigstk.ss_flags = SS_DISABLE;
726 		lp->lwp_sigstk.ss_size = 0;
727 		lp->lwp_sigstk.ss_sp = NULL;
728 		lp->lwp_flags &= ~LWP_ALTSTACK;
729 	} else {
730 		lp->lwp_flags |= origlp->lwp_flags & LWP_ALTSTACK;
731 	}
732 
733 	/*
734 	 * Set cpbase to the last timeout that occured (not the upcoming
735 	 * timeout).
736 	 *
737 	 * A critical section is required since a timer IPI can update
738 	 * scheduler specific data.
739 	 */
740 	crit_enter();
741 	lp->lwp_cpbase = gd->gd_schedclock.time - gd->gd_schedclock.periodic;
742 	destproc->p_usched->heuristic_forking(origlp, lp);
743 	crit_exit();
744 	CPUMASK_ANDMASK(lp->lwp_cpumask, usched_mastermask);
745 	lwkt_token_init(&lp->lwp_token, "lwp_token");
746 	spin_init(&lp->lwp_spin, "lwptoken");
747 
748 	/*
749 	 * Assign the thread to the current cpu to begin with so we
750 	 * can manipulate it.
751 	 */
752 	td = lwkt_alloc_thread(NULL, LWKT_THREAD_STACK, gd->gd_cpuid, 0);
753 	lp->lwp_thread = td;
754 	td->td_ucred = crhold(destproc->p_ucred);
755 	td->td_proc = destproc;
756 	td->td_lwp = lp;
757 	td->td_switch = cpu_heavy_switch;
758 #ifdef NO_LWKT_SPLIT_USERPRI
759 	lwkt_setpri(td, TDPRI_USER_NORM);
760 #else
761 	lwkt_setpri(td, TDPRI_KERN_USER);
762 #endif
763 	lwkt_set_comm(td, "%s", destproc->p_comm);
764 
765 	/*
766 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
767 	 * and make the child ready to run.
768 	 */
769 	cpu_fork(origlp, lp, flags);
770 	kqueue_init(&lp->lwp_kqueue, destproc->p_fd);
771 
772 	/*
773 	 * Assign a TID to the lp.  Loop until the insert succeeds (returns
774 	 * NULL).
775 	 *
776 	 * If we are in a vfork assign the same TID as the lwp that did the
777 	 * vfork().  This way if the user program messes around with
778 	 * pthread calls inside the vfork(), it will operate like an
779 	 * extension of the (blocked) parent.  Also note that since the
780 	 * address space is being shared, insofar as pthreads is concerned,
781 	 * the code running in the vfork() is part of the original process.
782 	 */
783 	if (flags & RFPPWAIT) {
784 		lp->lwp_tid = origlp->lwp_tid - 1;
785 	} else {
786 		lp->lwp_tid = destproc->p_lasttid;
787 	}
788 
789 	do {
790 		if (++lp->lwp_tid <= 0)
791 			lp->lwp_tid = 1;
792 	} while (lwp_rb_tree_RB_INSERT(&destproc->p_lwp_tree, lp) != NULL);
793 
794 	destproc->p_lasttid = lp->lwp_tid;
795 	destproc->p_nthreads++;
796 
797 	/*
798 	 * This flag is set and never cleared.  It means that the process
799 	 * was threaded at some point.  Used to improve exit performance.
800 	 */
801 	destproc->p_flags |= P_MAYBETHREADED;
802 
803 	return (lp);
804 }
805 
806 /*
807  * The next two functionms are general routines to handle adding/deleting
808  * items on the fork callout list.
809  *
810  * at_fork():
811  * Take the arguments given and put them onto the fork callout list,
812  * However first make sure that it's not already there.
813  * Returns 0 on success or a standard error number.
814  */
815 int
816 at_fork(forklist_fn function)
817 {
818 	struct forklist *ep;
819 
820 #ifdef INVARIANTS
821 	/* let the programmer know if he's been stupid */
822 	if (rm_at_fork(function)) {
823 		kprintf("WARNING: fork callout entry (%p) already present\n",
824 		    function);
825 	}
826 #endif
827 	ep = kmalloc(sizeof(*ep), M_ATFORK, M_WAITOK|M_ZERO);
828 	ep->function = function;
829 	TAILQ_INSERT_TAIL(&fork_list, ep, next);
830 	return (0);
831 }
832 
833 /*
834  * Scan the exit callout list for the given item and remove it..
835  * Returns the number of items removed (0 or 1)
836  */
837 int
838 rm_at_fork(forklist_fn function)
839 {
840 	struct forklist *ep;
841 
842 	TAILQ_FOREACH(ep, &fork_list, next) {
843 		if (ep->function == function) {
844 			TAILQ_REMOVE(&fork_list, ep, next);
845 			kfree(ep, M_ATFORK);
846 			return(1);
847 		}
848 	}
849 	return (0);
850 }
851 
852 /*
853  * Add a forked process to the run queue after any remaining setup, such
854  * as setting the fork handler, has been completed.
855  *
856  * p2 is held by the caller.
857  */
858 void
859 start_forked_proc(struct lwp *lp1, struct proc *p2)
860 {
861 	struct lwp *lp2 = ONLY_LWP_IN_PROC(p2);
862 	int pflags;
863 
864 	/*
865 	 * Move from SIDL to RUN queue, and activate the process's thread.
866 	 * Activation of the thread effectively makes the process "a"
867 	 * current process, so we do not setrunqueue().
868 	 *
869 	 * YYY setrunqueue works here but we should clean up the trampoline
870 	 * code so we just schedule the LWKT thread and let the trampoline
871 	 * deal with the userland scheduler on return to userland.
872 	 */
873 	KASSERT(p2->p_stat == SIDL,
874 	    ("cannot start forked process, bad status: %p", p2));
875 	p2->p_usched->resetpriority(lp2);
876 	crit_enter();
877 	p2->p_stat = SACTIVE;
878 	lp2->lwp_stat = LSRUN;
879 	p2->p_usched->setrunqueue(lp2);
880 	crit_exit();
881 
882 	/*
883 	 * Now can be swapped.
884 	 */
885 	PRELE(lp1->lwp_proc);
886 
887 	/*
888 	 * Preserve synchronization semantics of vfork.  P_PPWAIT is set in
889 	 * the child until it has retired the parent's resources.  The parent
890 	 * must wait for the flag to be cleared by the child.
891 	 *
892 	 * Interlock the flag/tsleep with atomic ops to avoid unnecessary
893 	 * p_token conflicts.
894 	 *
895 	 * XXX Is this use of an atomic op on a field that is not normally
896 	 *     manipulated with atomic ops ok?
897 	 */
898 	while ((pflags = p2->p_flags) & P_PPWAIT) {
899 		cpu_ccfence();
900 		tsleep_interlock(lp1->lwp_proc, 0);
901 		if (atomic_cmpset_int(&p2->p_flags, pflags, pflags))
902 			tsleep(lp1->lwp_proc, PINTERLOCKED, "ppwait", 0);
903 	}
904 }
905 
906 /*
907  * procctl (idtype_t idtype, id_t id, int cmd, void *arg)
908  */
909 int
910 sys_procctl(struct procctl_args *uap)
911 {
912 	struct proc *p = curproc;
913 	struct proc *p2;
914 	struct sysreaper *reap;
915 	union reaper_info udata;
916 	int error;
917 
918 	if (uap->idtype != P_PID || uap->id != (id_t)p->p_pid)
919 		return EINVAL;
920 
921 	switch(uap->cmd) {
922 	case PROC_REAP_ACQUIRE:
923 		lwkt_gettoken(&p->p_token);
924 		reap = kmalloc(sizeof(*reap), M_REAPER, M_WAITOK|M_ZERO);
925 		if (p->p_reaper == NULL || p->p_reaper->p != p) {
926 			reaper_init(p, reap);
927 			error = 0;
928 		} else {
929 			kfree(reap, M_REAPER);
930 			error = EALREADY;
931 		}
932 		lwkt_reltoken(&p->p_token);
933 		break;
934 	case PROC_REAP_RELEASE:
935 		lwkt_gettoken(&p->p_token);
936 release_again:
937 		reap = p->p_reaper;
938 		KKASSERT(reap != NULL);
939 		if (reap->p == p) {
940 			reaper_hold(reap);	/* in case of thread race */
941 			lockmgr(&reap->lock, LK_EXCLUSIVE);
942 			if (reap->p != p) {
943 				lockmgr(&reap->lock, LK_RELEASE);
944 				reaper_drop(reap);
945 				goto release_again;
946 			}
947 			reap->p = NULL;
948 			p->p_reaper = reap->parent;
949 			if (p->p_reaper)
950 				reaper_hold(p->p_reaper);
951 			lockmgr(&reap->lock, LK_RELEASE);
952 			reaper_drop(reap);	/* our ref */
953 			reaper_drop(reap);	/* old p_reaper ref */
954 			error = 0;
955 		} else {
956 			error = ENOTCONN;
957 		}
958 		lwkt_reltoken(&p->p_token);
959 		break;
960 	case PROC_REAP_STATUS:
961 		bzero(&udata, sizeof(udata));
962 		lwkt_gettoken_shared(&p->p_token);
963 		if ((reap = p->p_reaper) != NULL && reap->p == p) {
964 			udata.status.flags = reap->flags;
965 			udata.status.refs = reap->refs - 1; /* minus ours */
966 		}
967 		p2 = LIST_FIRST(&p->p_children);
968 		udata.status.pid_head = p2 ? p2->p_pid : -1;
969 		lwkt_reltoken(&p->p_token);
970 
971 		if (uap->data) {
972 			error = copyout(&udata, uap->data,
973 					sizeof(udata.status));
974 		} else {
975 			error = 0;
976 		}
977 		break;
978 	default:
979 		error = EINVAL;
980 		break;
981 	}
982 	return error;
983 }
984 
985 /*
986  * Bump ref on reaper, preventing destruction
987  */
988 void
989 reaper_hold(struct sysreaper *reap)
990 {
991 	KKASSERT(reap->refs > 0);
992 	refcount_acquire(&reap->refs);
993 }
994 
995 /*
996  * Drop ref on reaper, destroy the structure on the 1->0
997  * transition and loop on the parent.
998  */
999 void
1000 reaper_drop(struct sysreaper *next)
1001 {
1002 	struct sysreaper *reap;
1003 
1004 	while ((reap = next) != NULL) {
1005 		if (refcount_release(&reap->refs)) {
1006 			next = reap->parent;
1007 			KKASSERT(reap->p == NULL);
1008 			lockmgr(&reaper_lock, LK_EXCLUSIVE);
1009 			reap->parent = NULL;
1010 			kfree(reap, M_REAPER);
1011 			lockmgr(&reaper_lock, LK_RELEASE);
1012 		} else {
1013 			next = NULL;
1014 		}
1015 	}
1016 }
1017 
1018 /*
1019  * Initialize a static or newly allocated reaper structure
1020  */
1021 void
1022 reaper_init(struct proc *p, struct sysreaper *reap)
1023 {
1024 	reap->parent = p->p_reaper;
1025 	reap->p = p;
1026 	if (p == initproc) {
1027 		reap->flags = REAPER_STAT_OWNED | REAPER_STAT_REALINIT;
1028 		reap->refs = 2;
1029 	} else {
1030 		reap->flags = REAPER_STAT_OWNED;
1031 		reap->refs = 1;
1032 	}
1033 	lockinit(&reap->lock, "subrp", 0, 0);
1034 	cpu_sfence();
1035 	p->p_reaper = reap;
1036 }
1037 
1038 /*
1039  * Called with p->p_token held during exit.
1040  *
1041  * This is a bit simpler than RELEASE because there are no threads remaining
1042  * to race.  We only release if we own the reaper, the exit code will handle
1043  * the final p_reaper release.
1044  */
1045 struct sysreaper *
1046 reaper_exit(struct proc *p)
1047 {
1048 	struct sysreaper *reap;
1049 
1050 	/*
1051 	 * Release acquired reaper
1052 	 */
1053 	if ((reap = p->p_reaper) != NULL && reap->p == p) {
1054 		lockmgr(&reap->lock, LK_EXCLUSIVE);
1055 		p->p_reaper = reap->parent;
1056 		if (p->p_reaper)
1057 			reaper_hold(p->p_reaper);
1058 		reap->p = NULL;
1059 		lockmgr(&reap->lock, LK_RELEASE);
1060 		reaper_drop(reap);
1061 	}
1062 
1063 	/*
1064 	 * Return and clear reaper (caller is holding p_token for us)
1065 	 * (reap->p does not equal p).  Caller must drop it.
1066 	 */
1067 	if ((reap = p->p_reaper) != NULL) {
1068 		p->p_reaper = NULL;
1069 	}
1070 	return reap;
1071 }
1072 
1073 /*
1074  * Return a held (PHOLD) process representing the reaper for process (p).
1075  * NULL should not normally be returned.  Caller should PRELE() the returned
1076  * reaper process when finished.
1077  *
1078  * Remove dead internal nodes while we are at it.
1079  *
1080  * Process (p)'s token must be held on call.
1081  * The returned process's token is NOT acquired by this routine.
1082  */
1083 struct proc *
1084 reaper_get(struct sysreaper *reap)
1085 {
1086 	struct sysreaper *next;
1087 	struct proc *reproc;
1088 
1089 	if (reap == NULL)
1090 		return NULL;
1091 
1092 	/*
1093 	 * Extra hold for loop
1094 	 */
1095 	reaper_hold(reap);
1096 
1097 	while (reap) {
1098 		lockmgr(&reap->lock, LK_SHARED);
1099 		if (reap->p) {
1100 			/*
1101 			 * Probable reaper
1102 			 */
1103 			if (reap->p) {
1104 				reproc = reap->p;
1105 				PHOLD(reproc);
1106 				lockmgr(&reap->lock, LK_RELEASE);
1107 				reaper_drop(reap);
1108 				return reproc;
1109 			}
1110 
1111 			/*
1112 			 * Raced, try again
1113 			 */
1114 			lockmgr(&reap->lock, LK_RELEASE);
1115 			continue;
1116 		}
1117 
1118 		/*
1119 		 * Traverse upwards in the reaper topology, destroy
1120 		 * dead internal nodes when possible.
1121 		 *
1122 		 * NOTE: Our ref on next means that a dead node should
1123 		 *	 have 2 (ours and reap->parent's).
1124 		 */
1125 		next = reap->parent;
1126 		while (next) {
1127 			reaper_hold(next);
1128 			if (next->refs == 2 && next->p == NULL) {
1129 				lockmgr(&reap->lock, LK_RELEASE);
1130 				lockmgr(&reap->lock, LK_EXCLUSIVE);
1131 				if (next->refs == 2 &&
1132 				    reap->parent == next &&
1133 				    next->p == NULL) {
1134 					/*
1135 					 * reap->parent inherits ref from next.
1136 					 */
1137 					reap->parent = next->parent;
1138 					next->parent = NULL;
1139 					reaper_drop(next);	/* ours */
1140 					reaper_drop(next);	/* old parent */
1141 					next = reap->parent;
1142 					continue;	/* possible chain */
1143 				}
1144 			}
1145 			break;
1146 		}
1147 		lockmgr(&reap->lock, LK_RELEASE);
1148 		reaper_drop(reap);
1149 		reap = next;
1150 	}
1151 	return NULL;
1152 }
1153 
1154 /*
1155  * Test that the sender is allowed to send a signal to the target.
1156  * The sender process is assumed to have a stable reaper.  The
1157  * target can be e.g. from a scan callback.
1158  *
1159  * Target cannot be the reaper process itself unless reaper_ok is specified,
1160  * or sender == target.
1161  */
1162 int
1163 reaper_sigtest(struct proc *sender, struct proc *target, int reaper_ok)
1164 {
1165 	struct sysreaper *sreap;
1166 	struct sysreaper *reap;
1167 	int r;
1168 
1169 	sreap = sender->p_reaper;
1170 	if (sreap == NULL)
1171 		return 1;
1172 
1173 	if (sreap == target->p_reaper) {
1174 		if (sreap->p == target && sreap->p != sender && reaper_ok == 0)
1175 			return 0;
1176 		return 1;
1177 	}
1178 	lockmgr(&reaper_lock, LK_SHARED);
1179 	r = 0;
1180 	for (reap = target->p_reaper; reap; reap = reap->parent) {
1181 		if (sreap == reap) {
1182 			if (sreap->p != target || reaper_ok)
1183 				r = 1;
1184 			break;
1185 		}
1186 	}
1187 	lockmgr(&reaper_lock, LK_RELEASE);
1188 
1189 	return r;
1190 }
1191