xref: /freebsd/sys/kern/kern_fork.c (revision 595e514d0df2bac5b813d35f83e32875dbf16a83)
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  * 4. 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  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_kdtrace.h"
41 #include "opt_ktrace.h"
42 #include "opt_kstack_pages.h"
43 #include "opt_procdesc.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/sysproto.h>
48 #include <sys/eventhandler.h>
49 #include <sys/fcntl.h>
50 #include <sys/filedesc.h>
51 #include <sys/jail.h>
52 #include <sys/kernel.h>
53 #include <sys/kthread.h>
54 #include <sys/sysctl.h>
55 #include <sys/lock.h>
56 #include <sys/malloc.h>
57 #include <sys/mutex.h>
58 #include <sys/priv.h>
59 #include <sys/proc.h>
60 #include <sys/procdesc.h>
61 #include <sys/pioctl.h>
62 #include <sys/racct.h>
63 #include <sys/resourcevar.h>
64 #include <sys/sched.h>
65 #include <sys/syscall.h>
66 #include <sys/vmmeter.h>
67 #include <sys/vnode.h>
68 #include <sys/acct.h>
69 #include <sys/ktr.h>
70 #include <sys/ktrace.h>
71 #include <sys/unistd.h>
72 #include <sys/sdt.h>
73 #include <sys/sx.h>
74 #include <sys/sysent.h>
75 #include <sys/signalvar.h>
76 
77 #include <security/audit/audit.h>
78 #include <security/mac/mac_framework.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_extern.h>
84 #include <vm/uma.h>
85 
86 #ifdef KDTRACE_HOOKS
87 #include <sys/dtrace_bsd.h>
88 dtrace_fork_func_t	dtrace_fasttrap_fork;
89 #endif
90 
91 SDT_PROVIDER_DECLARE(proc);
92 SDT_PROBE_DEFINE(proc, kernel, , create, create);
93 SDT_PROBE_ARGTYPE(proc, kernel, , create, 0, "struct proc *");
94 SDT_PROBE_ARGTYPE(proc, kernel, , create, 1, "struct proc *");
95 SDT_PROBE_ARGTYPE(proc, kernel, , create, 2, "int");
96 
97 #ifndef _SYS_SYSPROTO_H_
98 struct fork_args {
99 	int     dummy;
100 };
101 #endif
102 
103 /* ARGSUSED */
104 int
105 sys_fork(struct thread *td, struct fork_args *uap)
106 {
107 	int error;
108 	struct proc *p2;
109 
110 	error = fork1(td, RFFDG | RFPROC, 0, &p2, NULL, 0);
111 	if (error == 0) {
112 		td->td_retval[0] = p2->p_pid;
113 		td->td_retval[1] = 0;
114 	}
115 	return (error);
116 }
117 
118 /* ARGUSED */
119 int
120 sys_pdfork(td, uap)
121 	struct thread *td;
122 	struct pdfork_args *uap;
123 {
124 #ifdef PROCDESC
125 	int error, fd;
126 	struct proc *p2;
127 
128 	/*
129 	 * It is necessary to return fd by reference because 0 is a valid file
130 	 * descriptor number, and the child needs to be able to distinguish
131 	 * itself from the parent using the return value.
132 	 */
133 	error = fork1(td, RFFDG | RFPROC | RFPROCDESC, 0, &p2,
134 	    &fd, uap->flags);
135 	if (error == 0) {
136 		td->td_retval[0] = p2->p_pid;
137 		td->td_retval[1] = 0;
138 		error = copyout(&fd, uap->fdp, sizeof(fd));
139 	}
140 	return (error);
141 #else
142 	return (ENOSYS);
143 #endif
144 }
145 
146 /* ARGSUSED */
147 int
148 sys_vfork(struct thread *td, struct vfork_args *uap)
149 {
150 	int error, flags;
151 	struct proc *p2;
152 
153 	flags = RFFDG | RFPROC | RFPPWAIT | RFMEM;
154 	error = fork1(td, flags, 0, &p2, NULL, 0);
155 	if (error == 0) {
156 		td->td_retval[0] = p2->p_pid;
157 		td->td_retval[1] = 0;
158 	}
159 	return (error);
160 }
161 
162 int
163 sys_rfork(struct thread *td, struct rfork_args *uap)
164 {
165 	struct proc *p2;
166 	int error;
167 
168 	/* Don't allow kernel-only flags. */
169 	if ((uap->flags & RFKERNELONLY) != 0)
170 		return (EINVAL);
171 
172 	AUDIT_ARG_FFLAGS(uap->flags);
173 	error = fork1(td, uap->flags, 0, &p2, NULL, 0);
174 	if (error == 0) {
175 		td->td_retval[0] = p2 ? p2->p_pid : 0;
176 		td->td_retval[1] = 0;
177 	}
178 	return (error);
179 }
180 
181 int	nprocs = 1;		/* process 0 */
182 int	lastpid = 0;
183 SYSCTL_INT(_kern, OID_AUTO, lastpid, CTLFLAG_RD, &lastpid, 0,
184     "Last used PID");
185 
186 /*
187  * Random component to lastpid generation.  We mix in a random factor to make
188  * it a little harder to predict.  We sanity check the modulus value to avoid
189  * doing it in critical paths.  Don't let it be too small or we pointlessly
190  * waste randomness entropy, and don't let it be impossibly large.  Using a
191  * modulus that is too big causes a LOT more process table scans and slows
192  * down fork processing as the pidchecked caching is defeated.
193  */
194 static int randompid = 0;
195 
196 static int
197 sysctl_kern_randompid(SYSCTL_HANDLER_ARGS)
198 {
199 	int error, pid;
200 
201 	error = sysctl_wire_old_buffer(req, sizeof(int));
202 	if (error != 0)
203 		return(error);
204 	sx_xlock(&allproc_lock);
205 	pid = randompid;
206 	error = sysctl_handle_int(oidp, &pid, 0, req);
207 	if (error == 0 && req->newptr != NULL) {
208 		if (pid < 0 || pid > pid_max - 100)	/* out of range */
209 			pid = pid_max - 100;
210 		else if (pid < 2)			/* NOP */
211 			pid = 0;
212 		else if (pid < 100)			/* Make it reasonable */
213 			pid = 100;
214 		randompid = pid;
215 	}
216 	sx_xunlock(&allproc_lock);
217 	return (error);
218 }
219 
220 SYSCTL_PROC(_kern, OID_AUTO, randompid, CTLTYPE_INT|CTLFLAG_RW,
221     0, 0, sysctl_kern_randompid, "I", "Random PID modulus");
222 
223 static int
224 fork_findpid(int flags)
225 {
226 	struct proc *p;
227 	int trypid;
228 	static int pidchecked = 0;
229 
230 	/*
231 	 * Requires allproc_lock in order to iterate over the list
232 	 * of processes, and proctree_lock to access p_pgrp.
233 	 */
234 	sx_assert(&allproc_lock, SX_LOCKED);
235 	sx_assert(&proctree_lock, SX_LOCKED);
236 
237 	/*
238 	 * Find an unused process ID.  We remember a range of unused IDs
239 	 * ready to use (from lastpid+1 through pidchecked-1).
240 	 *
241 	 * If RFHIGHPID is set (used during system boot), do not allocate
242 	 * low-numbered pids.
243 	 */
244 	trypid = lastpid + 1;
245 	if (flags & RFHIGHPID) {
246 		if (trypid < 10)
247 			trypid = 10;
248 	} else {
249 		if (randompid)
250 			trypid += arc4random() % randompid;
251 	}
252 retry:
253 	/*
254 	 * If the process ID prototype has wrapped around,
255 	 * restart somewhat above 0, as the low-numbered procs
256 	 * tend to include daemons that don't exit.
257 	 */
258 	if (trypid >= pid_max) {
259 		trypid = trypid % pid_max;
260 		if (trypid < 100)
261 			trypid += 100;
262 		pidchecked = 0;
263 	}
264 	if (trypid >= pidchecked) {
265 		int doingzomb = 0;
266 
267 		pidchecked = PID_MAX;
268 		/*
269 		 * Scan the active and zombie procs to check whether this pid
270 		 * is in use.  Remember the lowest pid that's greater
271 		 * than trypid, so we can avoid checking for a while.
272 		 */
273 		p = LIST_FIRST(&allproc);
274 again:
275 		for (; p != NULL; p = LIST_NEXT(p, p_list)) {
276 			while (p->p_pid == trypid ||
277 			    (p->p_pgrp != NULL &&
278 			    (p->p_pgrp->pg_id == trypid ||
279 			    (p->p_session != NULL &&
280 			    p->p_session->s_sid == trypid)))) {
281 				trypid++;
282 				if (trypid >= pidchecked)
283 					goto retry;
284 			}
285 			if (p->p_pid > trypid && pidchecked > p->p_pid)
286 				pidchecked = p->p_pid;
287 			if (p->p_pgrp != NULL) {
288 				if (p->p_pgrp->pg_id > trypid &&
289 				    pidchecked > p->p_pgrp->pg_id)
290 					pidchecked = p->p_pgrp->pg_id;
291 				if (p->p_session != NULL &&
292 				    p->p_session->s_sid > trypid &&
293 				    pidchecked > p->p_session->s_sid)
294 					pidchecked = p->p_session->s_sid;
295 			}
296 		}
297 		if (!doingzomb) {
298 			doingzomb = 1;
299 			p = LIST_FIRST(&zombproc);
300 			goto again;
301 		}
302 	}
303 
304 	/*
305 	 * RFHIGHPID does not mess with the lastpid counter during boot.
306 	 */
307 	if (flags & RFHIGHPID)
308 		pidchecked = 0;
309 	else
310 		lastpid = trypid;
311 
312 	return (trypid);
313 }
314 
315 static int
316 fork_norfproc(struct thread *td, int flags)
317 {
318 	int error;
319 	struct proc *p1;
320 
321 	KASSERT((flags & RFPROC) == 0,
322 	    ("fork_norfproc called with RFPROC set"));
323 	p1 = td->td_proc;
324 
325 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
326 	    (flags & (RFCFDG | RFFDG))) {
327 		PROC_LOCK(p1);
328 		if (thread_single(SINGLE_BOUNDARY)) {
329 			PROC_UNLOCK(p1);
330 			return (ERESTART);
331 		}
332 		PROC_UNLOCK(p1);
333 	}
334 
335 	error = vm_forkproc(td, NULL, NULL, NULL, flags);
336 	if (error)
337 		goto fail;
338 
339 	/*
340 	 * Close all file descriptors.
341 	 */
342 	if (flags & RFCFDG) {
343 		struct filedesc *fdtmp;
344 		fdtmp = fdinit(td->td_proc->p_fd);
345 		fdescfree(td);
346 		p1->p_fd = fdtmp;
347 	}
348 
349 	/*
350 	 * Unshare file descriptors (from parent).
351 	 */
352 	if (flags & RFFDG)
353 		fdunshare(p1, td);
354 
355 fail:
356 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
357 	    (flags & (RFCFDG | RFFDG))) {
358 		PROC_LOCK(p1);
359 		thread_single_end();
360 		PROC_UNLOCK(p1);
361 	}
362 	return (error);
363 }
364 
365 static void
366 do_fork(struct thread *td, int flags, struct proc *p2, struct thread *td2,
367     struct vmspace *vm2, int pdflags)
368 {
369 	struct proc *p1, *pptr;
370 	int p2_held, trypid;
371 	struct filedesc *fd;
372 	struct filedesc_to_leader *fdtol;
373 	struct sigacts *newsigacts;
374 
375 	sx_assert(&proctree_lock, SX_SLOCKED);
376 	sx_assert(&allproc_lock, SX_XLOCKED);
377 
378 	p2_held = 0;
379 	p1 = td->td_proc;
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 	nprocs++;
386 
387 	trypid = fork_findpid(flags);
388 
389 	sx_sunlock(&proctree_lock);
390 
391 	p2->p_state = PRS_NEW;		/* protect against others */
392 	p2->p_pid = trypid;
393 	AUDIT_ARG_PID(p2->p_pid);
394 	LIST_INSERT_HEAD(&allproc, p2, p_list);
395 	LIST_INSERT_HEAD(PIDHASH(p2->p_pid), p2, p_hash);
396 	tidhash_add(td2);
397 	PROC_LOCK(p2);
398 	PROC_LOCK(p1);
399 
400 	sx_xunlock(&allproc_lock);
401 
402 	bcopy(&p1->p_startcopy, &p2->p_startcopy,
403 	    __rangeof(struct proc, p_startcopy, p_endcopy));
404 	pargs_hold(p2->p_args);
405 	PROC_UNLOCK(p1);
406 
407 	bzero(&p2->p_startzero,
408 	    __rangeof(struct proc, p_startzero, p_endzero));
409 
410 	p2->p_ucred = crhold(td->td_ucred);
411 
412 	/* Tell the prison that we exist. */
413 	prison_proc_hold(p2->p_ucred->cr_prison);
414 
415 	PROC_UNLOCK(p2);
416 
417 	/*
418 	 * Malloc things while we don't hold any locks.
419 	 */
420 	if (flags & RFSIGSHARE)
421 		newsigacts = NULL;
422 	else
423 		newsigacts = sigacts_alloc();
424 
425 	/*
426 	 * Copy filedesc.
427 	 */
428 	if (flags & RFCFDG) {
429 		fd = fdinit(p1->p_fd);
430 		fdtol = NULL;
431 	} else if (flags & RFFDG) {
432 		fd = fdcopy(p1->p_fd);
433 		fdtol = NULL;
434 	} else {
435 		fd = fdshare(p1->p_fd);
436 		if (p1->p_fdtol == NULL)
437 			p1->p_fdtol = filedesc_to_leader_alloc(NULL, NULL,
438 			    p1->p_leader);
439 		if ((flags & RFTHREAD) != 0) {
440 			/*
441 			 * Shared file descriptor table, and shared
442 			 * process leaders.
443 			 */
444 			fdtol = p1->p_fdtol;
445 			FILEDESC_XLOCK(p1->p_fd);
446 			fdtol->fdl_refcount++;
447 			FILEDESC_XUNLOCK(p1->p_fd);
448 		} else {
449 			/*
450 			 * Shared file descriptor table, and different
451 			 * process leaders.
452 			 */
453 			fdtol = filedesc_to_leader_alloc(p1->p_fdtol,
454 			    p1->p_fd, p2);
455 		}
456 	}
457 	/*
458 	 * Make a proc table entry for the new process.
459 	 * Start by zeroing the section of proc that is zero-initialized,
460 	 * then copy the section that is copied directly from the parent.
461 	 */
462 
463 	PROC_LOCK(p2);
464 	PROC_LOCK(p1);
465 
466 	bzero(&td2->td_startzero,
467 	    __rangeof(struct thread, td_startzero, td_endzero));
468 
469 	bcopy(&td->td_startcopy, &td2->td_startcopy,
470 	    __rangeof(struct thread, td_startcopy, td_endcopy));
471 
472 	bcopy(&p2->p_comm, &td2->td_name, sizeof(td2->td_name));
473 	td2->td_sigstk = td->td_sigstk;
474 	td2->td_flags = TDF_INMEM;
475 	td2->td_lend_user_pri = PRI_MAX;
476 
477 #ifdef VIMAGE
478 	td2->td_vnet = NULL;
479 	td2->td_vnet_lpush = NULL;
480 #endif
481 
482 	/*
483 	 * Allow the scheduler to initialize the child.
484 	 */
485 	thread_lock(td);
486 	sched_fork(td, td2);
487 	thread_unlock(td);
488 
489 	/*
490 	 * Duplicate sub-structures as needed.
491 	 * Increase reference counts on shared objects.
492 	 */
493 	p2->p_flag = P_INMEM;
494 	p2->p_swtick = ticks;
495 	if (p1->p_flag & P_PROFIL)
496 		startprofclock(p2);
497 	td2->td_ucred = crhold(p2->p_ucred);
498 
499 	if (flags & RFSIGSHARE) {
500 		p2->p_sigacts = sigacts_hold(p1->p_sigacts);
501 	} else {
502 		sigacts_copy(newsigacts, p1->p_sigacts);
503 		p2->p_sigacts = newsigacts;
504 	}
505 
506 	if (flags & RFTSIGZMB)
507 	        p2->p_sigparent = RFTSIGNUM(flags);
508 	else if (flags & RFLINUXTHPN)
509 	        p2->p_sigparent = SIGUSR1;
510 	else
511 	        p2->p_sigparent = SIGCHLD;
512 
513 	p2->p_textvp = p1->p_textvp;
514 	p2->p_fd = fd;
515 	p2->p_fdtol = fdtol;
516 
517 	/*
518 	 * p_limit is copy-on-write.  Bump its refcount.
519 	 */
520 	lim_fork(p1, p2);
521 
522 	pstats_fork(p1->p_stats, p2->p_stats);
523 
524 	PROC_UNLOCK(p1);
525 	PROC_UNLOCK(p2);
526 
527 	/* Bump references to the text vnode (for procfs). */
528 	if (p2->p_textvp)
529 		vref(p2->p_textvp);
530 
531 	/*
532 	 * Set up linkage for kernel based threading.
533 	 */
534 	if ((flags & RFTHREAD) != 0) {
535 		mtx_lock(&ppeers_lock);
536 		p2->p_peers = p1->p_peers;
537 		p1->p_peers = p2;
538 		p2->p_leader = p1->p_leader;
539 		mtx_unlock(&ppeers_lock);
540 		PROC_LOCK(p1->p_leader);
541 		if ((p1->p_leader->p_flag & P_WEXIT) != 0) {
542 			PROC_UNLOCK(p1->p_leader);
543 			/*
544 			 * The task leader is exiting, so process p1 is
545 			 * going to be killed shortly.  Since p1 obviously
546 			 * isn't dead yet, we know that the leader is either
547 			 * sending SIGKILL's to all the processes in this
548 			 * task or is sleeping waiting for all the peers to
549 			 * exit.  We let p1 complete the fork, but we need
550 			 * to go ahead and kill the new process p2 since
551 			 * the task leader may not get a chance to send
552 			 * SIGKILL to it.  We leave it on the list so that
553 			 * the task leader will wait for this new process
554 			 * to commit suicide.
555 			 */
556 			PROC_LOCK(p2);
557 			kern_psignal(p2, SIGKILL);
558 			PROC_UNLOCK(p2);
559 		} else
560 			PROC_UNLOCK(p1->p_leader);
561 	} else {
562 		p2->p_peers = NULL;
563 		p2->p_leader = p2;
564 	}
565 
566 	sx_xlock(&proctree_lock);
567 	PGRP_LOCK(p1->p_pgrp);
568 	PROC_LOCK(p2);
569 	PROC_LOCK(p1);
570 
571 	/*
572 	 * Preserve some more flags in subprocess.  P_PROFIL has already
573 	 * been preserved.
574 	 */
575 	p2->p_flag |= p1->p_flag & P_SUGID;
576 	td2->td_pflags |= td->td_pflags & TDP_ALTSTACK;
577 	SESS_LOCK(p1->p_session);
578 	if (p1->p_session->s_ttyvp != NULL && p1->p_flag & P_CONTROLT)
579 		p2->p_flag |= P_CONTROLT;
580 	SESS_UNLOCK(p1->p_session);
581 	if (flags & RFPPWAIT)
582 		p2->p_flag |= P_PPWAIT;
583 
584 	p2->p_pgrp = p1->p_pgrp;
585 	LIST_INSERT_AFTER(p1, p2, p_pglist);
586 	PGRP_UNLOCK(p1->p_pgrp);
587 	LIST_INIT(&p2->p_children);
588 	LIST_INIT(&p2->p_orphans);
589 
590 	callout_init_mtx(&p2->p_itcallout, &p2->p_mtx, 0);
591 
592 	/*
593 	 * If PF_FORK is set, the child process inherits the
594 	 * procfs ioctl flags from its parent.
595 	 */
596 	if (p1->p_pfsflags & PF_FORK) {
597 		p2->p_stops = p1->p_stops;
598 		p2->p_pfsflags = p1->p_pfsflags;
599 	}
600 
601 	/*
602 	 * This begins the section where we must prevent the parent
603 	 * from being swapped.
604 	 */
605 	_PHOLD(p1);
606 	PROC_UNLOCK(p1);
607 
608 	/*
609 	 * Attach the new process to its parent.
610 	 *
611 	 * If RFNOWAIT is set, the newly created process becomes a child
612 	 * of init.  This effectively disassociates the child from the
613 	 * parent.
614 	 */
615 	if (flags & RFNOWAIT)
616 		pptr = initproc;
617 	else
618 		pptr = p1;
619 	p2->p_pptr = pptr;
620 	LIST_INSERT_HEAD(&pptr->p_children, p2, p_sibling);
621 	sx_xunlock(&proctree_lock);
622 
623 	/* Inform accounting that we have forked. */
624 	p2->p_acflag = AFORK;
625 	PROC_UNLOCK(p2);
626 
627 #ifdef KTRACE
628 	ktrprocfork(p1, p2);
629 #endif
630 
631 	/*
632 	 * Finish creating the child process.  It will return via a different
633 	 * execution path later.  (ie: directly into user mode)
634 	 */
635 	vm_forkproc(td, p2, td2, vm2, flags);
636 
637 	if (flags == (RFFDG | RFPROC)) {
638 		PCPU_INC(cnt.v_forks);
639 		PCPU_ADD(cnt.v_forkpages, p2->p_vmspace->vm_dsize +
640 		    p2->p_vmspace->vm_ssize);
641 	} else if (flags == (RFFDG | RFPROC | RFPPWAIT | RFMEM)) {
642 		PCPU_INC(cnt.v_vforks);
643 		PCPU_ADD(cnt.v_vforkpages, p2->p_vmspace->vm_dsize +
644 		    p2->p_vmspace->vm_ssize);
645 	} else if (p1 == &proc0) {
646 		PCPU_INC(cnt.v_kthreads);
647 		PCPU_ADD(cnt.v_kthreadpages, p2->p_vmspace->vm_dsize +
648 		    p2->p_vmspace->vm_ssize);
649 	} else {
650 		PCPU_INC(cnt.v_rforks);
651 		PCPU_ADD(cnt.v_rforkpages, p2->p_vmspace->vm_dsize +
652 		    p2->p_vmspace->vm_ssize);
653 	}
654 
655 #ifdef PROCDESC
656 	/*
657 	 * Associate the process descriptor with the process before anything
658 	 * can happen that might cause that process to need the descriptor.
659 	 * However, don't do this until after fork(2) can no longer fail.
660 	 */
661 	if (flags & RFPROCDESC)
662 		procdesc_new(p2, pdflags);
663 #endif
664 
665 	/*
666 	 * Both processes are set up, now check if any loadable modules want
667 	 * to adjust anything.
668 	 */
669 	EVENTHANDLER_INVOKE(process_fork, p1, p2, flags);
670 
671 	/*
672 	 * Set the child start time and mark the process as being complete.
673 	 */
674 	PROC_LOCK(p2);
675 	PROC_LOCK(p1);
676 	microuptime(&p2->p_stats->p_start);
677 	PROC_SLOCK(p2);
678 	p2->p_state = PRS_NORMAL;
679 	PROC_SUNLOCK(p2);
680 
681 #ifdef KDTRACE_HOOKS
682 	/*
683 	 * Tell the DTrace fasttrap provider about the new process
684 	 * if it has registered an interest. We have to do this only after
685 	 * p_state is PRS_NORMAL since the fasttrap module will use pfind()
686 	 * later on.
687 	 */
688 	if (dtrace_fasttrap_fork)
689 		dtrace_fasttrap_fork(p1, p2);
690 #endif
691 	if ((p1->p_flag & (P_TRACED | P_FOLLOWFORK)) == (P_TRACED |
692 	    P_FOLLOWFORK)) {
693 		/*
694 		 * Arrange for debugger to receive the fork event.
695 		 *
696 		 * We can report PL_FLAG_FORKED regardless of
697 		 * P_FOLLOWFORK settings, but it does not make a sense
698 		 * for runaway child.
699 		 */
700 		td->td_dbgflags |= TDB_FORK;
701 		td->td_dbg_forked = p2->p_pid;
702 		td2->td_dbgflags |= TDB_STOPATFORK;
703 		_PHOLD(p2);
704 		p2_held = 1;
705 	}
706 	if (flags & RFPPWAIT) {
707 		td->td_pflags |= TDP_RFPPWAIT;
708 		td->td_rfppwait_p = p2;
709 	}
710 	PROC_UNLOCK(p2);
711 	if ((flags & RFSTOPPED) == 0) {
712 		/*
713 		 * If RFSTOPPED not requested, make child runnable and
714 		 * add to run queue.
715 		 */
716 		thread_lock(td2);
717 		TD_SET_CAN_RUN(td2);
718 		sched_add(td2, SRQ_BORING);
719 		thread_unlock(td2);
720 	}
721 
722 	/*
723 	 * Now can be swapped.
724 	 */
725 	_PRELE(p1);
726 	PROC_UNLOCK(p1);
727 
728 	/*
729 	 * Tell any interested parties about the new process.
730 	 */
731 	knote_fork(&p1->p_klist, p2->p_pid);
732 	SDT_PROBE(proc, kernel, , create, p2, p1, flags, 0, 0);
733 
734 	/*
735 	 * Wait until debugger is attached to child.
736 	 */
737 	PROC_LOCK(p2);
738 	while ((td2->td_dbgflags & TDB_STOPATFORK) != 0)
739 		cv_wait(&p2->p_dbgwait, &p2->p_mtx);
740 	if (p2_held)
741 		_PRELE(p2);
742 	PROC_UNLOCK(p2);
743 }
744 
745 int
746 fork1(struct thread *td, int flags, int pages, struct proc **procp,
747     int *procdescp, int pdflags)
748 {
749 	struct proc *p1;
750 	struct proc *newproc;
751 	int ok;
752 	struct thread *td2;
753 	struct vmspace *vm2;
754 	vm_ooffset_t mem_charged;
755 	int error;
756 	static int curfail;
757 	static struct timeval lastfail;
758 #ifdef PROCDESC
759 	struct file *fp_procdesc = NULL;
760 #endif
761 
762 	/* Check for the undefined or unimplemented flags. */
763 	if ((flags & ~(RFFLAGS | RFTSIGFLAGS(RFTSIGMASK))) != 0)
764 		return (EINVAL);
765 
766 	/* Signal value requires RFTSIGZMB. */
767 	if ((flags & RFTSIGFLAGS(RFTSIGMASK)) != 0 && (flags & RFTSIGZMB) == 0)
768 		return (EINVAL);
769 
770 	/* Can't copy and clear. */
771 	if ((flags & (RFFDG|RFCFDG)) == (RFFDG|RFCFDG))
772 		return (EINVAL);
773 
774 	/* Check the validity of the signal number. */
775 	if ((flags & RFTSIGZMB) != 0 && (u_int)RFTSIGNUM(flags) > _SIG_MAXSIG)
776 		return (EINVAL);
777 
778 #ifdef PROCDESC
779 	if ((flags & RFPROCDESC) != 0) {
780 		/* Can't not create a process yet get a process descriptor. */
781 		if ((flags & RFPROC) == 0)
782 			return (EINVAL);
783 
784 		/* Must provide a place to put a procdesc if creating one. */
785 		if (procdescp == NULL)
786 			return (EINVAL);
787 	}
788 #endif
789 
790 	p1 = td->td_proc;
791 
792 	/*
793 	 * Here we don't create a new process, but we divorce
794 	 * certain parts of a process from itself.
795 	 */
796 	if ((flags & RFPROC) == 0) {
797 		*procp = NULL;
798 		return (fork_norfproc(td, flags));
799 	}
800 
801 #ifdef PROCDESC
802 	/*
803 	 * If required, create a process descriptor in the parent first; we
804 	 * will abandon it if something goes wrong. We don't finit() until
805 	 * later.
806 	 */
807 	if (flags & RFPROCDESC) {
808 		error = falloc(td, &fp_procdesc, procdescp, 0);
809 		if (error != 0)
810 			return (error);
811 	}
812 #endif
813 
814 	mem_charged = 0;
815 	vm2 = NULL;
816 	if (pages == 0)
817 		pages = KSTACK_PAGES;
818 	/* Allocate new proc. */
819 	newproc = uma_zalloc(proc_zone, M_WAITOK);
820 	td2 = FIRST_THREAD_IN_PROC(newproc);
821 	if (td2 == NULL) {
822 		td2 = thread_alloc(pages);
823 		if (td2 == NULL) {
824 			error = ENOMEM;
825 			goto fail1;
826 		}
827 		proc_linkup(newproc, td2);
828 	} else {
829 		if (td2->td_kstack == 0 || td2->td_kstack_pages != pages) {
830 			if (td2->td_kstack != 0)
831 				vm_thread_dispose(td2);
832 			if (!thread_alloc_stack(td2, pages)) {
833 				error = ENOMEM;
834 				goto fail1;
835 			}
836 		}
837 	}
838 
839 	if ((flags & RFMEM) == 0) {
840 		vm2 = vmspace_fork(p1->p_vmspace, &mem_charged);
841 		if (vm2 == NULL) {
842 			error = ENOMEM;
843 			goto fail1;
844 		}
845 		if (!swap_reserve(mem_charged)) {
846 			/*
847 			 * The swap reservation failed. The accounting
848 			 * from the entries of the copied vm2 will be
849 			 * substracted in vmspace_free(), so force the
850 			 * reservation there.
851 			 */
852 			swap_reserve_force(mem_charged);
853 			error = ENOMEM;
854 			goto fail1;
855 		}
856 	} else
857 		vm2 = NULL;
858 
859 	/*
860 	 * XXX: This is ugly; when we copy resource usage, we need to bump
861 	 *      per-cred resource counters.
862 	 */
863 	newproc->p_ucred = p1->p_ucred;
864 
865 	/*
866 	 * Initialize resource accounting for the child process.
867 	 */
868 	error = racct_proc_fork(p1, newproc);
869 	if (error != 0) {
870 		error = EAGAIN;
871 		goto fail1;
872 	}
873 
874 #ifdef MAC
875 	mac_proc_init(newproc);
876 #endif
877 	knlist_init_mtx(&newproc->p_klist, &newproc->p_mtx);
878 	STAILQ_INIT(&newproc->p_ktr);
879 
880 	/* We have to lock the process tree while we look for a pid. */
881 	sx_slock(&proctree_lock);
882 
883 	/*
884 	 * Although process entries are dynamically created, we still keep
885 	 * a global limit on the maximum number we will create.  Don't allow
886 	 * a nonprivileged user to use the last ten processes; don't let root
887 	 * exceed the limit. The variable nprocs is the current number of
888 	 * processes, maxproc is the limit.
889 	 */
890 	sx_xlock(&allproc_lock);
891 	if ((nprocs >= maxproc - 10 && priv_check_cred(td->td_ucred,
892 	    PRIV_MAXPROC, 0) != 0) || nprocs >= maxproc) {
893 		error = EAGAIN;
894 		goto fail;
895 	}
896 
897 	/*
898 	 * Increment the count of procs running with this uid. Don't allow
899 	 * a nonprivileged user to exceed their current limit.
900 	 *
901 	 * XXXRW: Can we avoid privilege here if it's not needed?
902 	 */
903 	error = priv_check_cred(td->td_ucred, PRIV_PROC_LIMIT, 0);
904 	if (error == 0)
905 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1, 0);
906 	else {
907 		PROC_LOCK(p1);
908 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1,
909 		    lim_cur(p1, RLIMIT_NPROC));
910 		PROC_UNLOCK(p1);
911 	}
912 	if (ok) {
913 		do_fork(td, flags, newproc, td2, vm2, pdflags);
914 
915 		/*
916 		 * Return child proc pointer to parent.
917 		 */
918 		*procp = newproc;
919 #ifdef PROCDESC
920 		if (flags & RFPROCDESC) {
921 			procdesc_finit(newproc->p_procdesc, fp_procdesc);
922 			fdrop(fp_procdesc, td);
923 		}
924 #endif
925 		racct_proc_fork_done(newproc);
926 		return (0);
927 	}
928 
929 	error = EAGAIN;
930 fail:
931 	sx_sunlock(&proctree_lock);
932 	if (ppsratecheck(&lastfail, &curfail, 1))
933 		printf("maxproc limit exceeded by uid %u (pid %d); see tuning(7) and login.conf(5)\n",
934 		    td->td_ucred->cr_ruid, p1->p_pid);
935 	sx_xunlock(&allproc_lock);
936 #ifdef MAC
937 	mac_proc_destroy(newproc);
938 #endif
939 	racct_proc_exit(newproc);
940 fail1:
941 	if (vm2 != NULL)
942 		vmspace_free(vm2);
943 	uma_zfree(proc_zone, newproc);
944 #ifdef PROCDESC
945 	if ((flags & RFPROCDESC) != 0 && fp_procdesc != NULL) {
946 		fdclose(td->td_proc->p_fd, fp_procdesc, *procdescp, td);
947 		fdrop(fp_procdesc, td);
948 	}
949 #endif
950 	pause("fork", hz / 2);
951 	return (error);
952 }
953 
954 /*
955  * Handle the return of a child process from fork1().  This function
956  * is called from the MD fork_trampoline() entry point.
957  */
958 void
959 fork_exit(void (*callout)(void *, struct trapframe *), void *arg,
960     struct trapframe *frame)
961 {
962 	struct proc *p;
963 	struct thread *td;
964 	struct thread *dtd;
965 
966 	td = curthread;
967 	p = td->td_proc;
968 	KASSERT(p->p_state == PRS_NORMAL, ("executing process is still new"));
969 
970 	CTR4(KTR_PROC, "fork_exit: new thread %p (td_sched %p, pid %d, %s)",
971 		td, td->td_sched, p->p_pid, td->td_name);
972 
973 	sched_fork_exit(td);
974 	/*
975 	* Processes normally resume in mi_switch() after being
976 	* cpu_switch()'ed to, but when children start up they arrive here
977 	* instead, so we must do much the same things as mi_switch() would.
978 	*/
979 	if ((dtd = PCPU_GET(deadthread))) {
980 		PCPU_SET(deadthread, NULL);
981 		thread_stash(dtd);
982 	}
983 	thread_unlock(td);
984 
985 	/*
986 	 * cpu_set_fork_handler intercepts this function call to
987 	 * have this call a non-return function to stay in kernel mode.
988 	 * initproc has its own fork handler, but it does return.
989 	 */
990 	KASSERT(callout != NULL, ("NULL callout in fork_exit"));
991 	callout(arg, frame);
992 
993 	/*
994 	 * Check if a kernel thread misbehaved and returned from its main
995 	 * function.
996 	 */
997 	if (p->p_flag & P_KTHREAD) {
998 		printf("Kernel thread \"%s\" (pid %d) exited prematurely.\n",
999 		    td->td_name, p->p_pid);
1000 		kproc_exit(0);
1001 	}
1002 	mtx_assert(&Giant, MA_NOTOWNED);
1003 
1004 	if (p->p_sysent->sv_schedtail != NULL)
1005 		(p->p_sysent->sv_schedtail)(td);
1006 }
1007 
1008 /*
1009  * Simplified back end of syscall(), used when returning from fork()
1010  * directly into user mode.  Giant is not held on entry, and must not
1011  * be held on return.  This function is passed in to fork_exit() as the
1012  * first parameter and is called when returning to a new userland process.
1013  */
1014 void
1015 fork_return(struct thread *td, struct trapframe *frame)
1016 {
1017 	struct proc *p, *dbg;
1018 
1019 	if (td->td_dbgflags & TDB_STOPATFORK) {
1020 		p = td->td_proc;
1021 		sx_xlock(&proctree_lock);
1022 		PROC_LOCK(p);
1023 		if ((p->p_pptr->p_flag & (P_TRACED | P_FOLLOWFORK)) ==
1024 		    (P_TRACED | P_FOLLOWFORK)) {
1025 			/*
1026 			 * If debugger still wants auto-attach for the
1027 			 * parent's children, do it now.
1028 			 */
1029 			dbg = p->p_pptr->p_pptr;
1030 			p->p_flag |= P_TRACED;
1031 			p->p_oppid = p->p_pptr->p_pid;
1032 			proc_reparent(p, dbg);
1033 			sx_xunlock(&proctree_lock);
1034 			td->td_dbgflags |= TDB_CHILD;
1035 			ptracestop(td, SIGSTOP);
1036 			td->td_dbgflags &= ~TDB_CHILD;
1037 		} else {
1038 			/*
1039 			 * ... otherwise clear the request.
1040 			 */
1041 			sx_xunlock(&proctree_lock);
1042 			td->td_dbgflags &= ~TDB_STOPATFORK;
1043 			cv_broadcast(&p->p_dbgwait);
1044 		}
1045 		PROC_UNLOCK(p);
1046 	}
1047 
1048 	userret(td, frame);
1049 
1050 #ifdef KTRACE
1051 	if (KTRPOINT(td, KTR_SYSRET))
1052 		ktrsysret(SYS_fork, 0, 0);
1053 #endif
1054 }
1055