xref: /freebsd/sys/kern/kern_fork.c (revision a4dc509f723944821bcfcc52005ff87c9a5dee5b)
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_ktrace.h"
41 #include "opt_kstack_pages.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/sysproto.h>
46 #include <sys/eventhandler.h>
47 #include <sys/fcntl.h>
48 #include <sys/filedesc.h>
49 #include <sys/jail.h>
50 #include <sys/kernel.h>
51 #include <sys/kthread.h>
52 #include <sys/sysctl.h>
53 #include <sys/lock.h>
54 #include <sys/malloc.h>
55 #include <sys/mutex.h>
56 #include <sys/priv.h>
57 #include <sys/proc.h>
58 #include <sys/procdesc.h>
59 #include <sys/pioctl.h>
60 #include <sys/ptrace.h>
61 #include <sys/racct.h>
62 #include <sys/resourcevar.h>
63 #include <sys/sched.h>
64 #include <sys/syscall.h>
65 #include <sys/vmmeter.h>
66 #include <sys/vnode.h>
67 #include <sys/acct.h>
68 #include <sys/ktr.h>
69 #include <sys/ktrace.h>
70 #include <sys/unistd.h>
71 #include <sys/sdt.h>
72 #include <sys/sx.h>
73 #include <sys/sysent.h>
74 #include <sys/signalvar.h>
75 
76 #include <security/audit/audit.h>
77 #include <security/mac/mac_framework.h>
78 
79 #include <vm/vm.h>
80 #include <vm/pmap.h>
81 #include <vm/vm_map.h>
82 #include <vm/vm_extern.h>
83 #include <vm/uma.h>
84 #include <vm/vm_domain.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_DEFINE3(proc, kernel, , create, "struct proc *",
93     "struct proc *", "int");
94 
95 #ifndef _SYS_SYSPROTO_H_
96 struct fork_args {
97 	int     dummy;
98 };
99 #endif
100 
101 /* ARGSUSED */
102 int
103 sys_fork(struct thread *td, struct fork_args *uap)
104 {
105 	int error;
106 	struct proc *p2;
107 
108 	error = fork1(td, RFFDG | RFPROC, 0, &p2, NULL, 0, NULL);
109 	if (error == 0) {
110 		td->td_retval[0] = p2->p_pid;
111 		td->td_retval[1] = 0;
112 	}
113 	return (error);
114 }
115 
116 /* ARGUSED */
117 int
118 sys_pdfork(td, uap)
119 	struct thread *td;
120 	struct pdfork_args *uap;
121 {
122 	int error, fd;
123 	struct proc *p2;
124 
125 	/*
126 	 * It is necessary to return fd by reference because 0 is a valid file
127 	 * descriptor number, and the child needs to be able to distinguish
128 	 * itself from the parent using the return value.
129 	 */
130 	error = fork1(td, RFFDG | RFPROC | RFPROCDESC, 0, &p2,
131 	    &fd, uap->flags, NULL);
132 	if (error == 0) {
133 		td->td_retval[0] = p2->p_pid;
134 		td->td_retval[1] = 0;
135 		error = copyout(&fd, uap->fdp, sizeof(fd));
136 	}
137 	return (error);
138 }
139 
140 /* ARGSUSED */
141 int
142 sys_vfork(struct thread *td, struct vfork_args *uap)
143 {
144 	int error, flags;
145 	struct proc *p2;
146 
147 	flags = RFFDG | RFPROC | RFPPWAIT | RFMEM;
148 	error = fork1(td, flags, 0, &p2, NULL, 0, NULL);
149 	if (error == 0) {
150 		td->td_retval[0] = p2->p_pid;
151 		td->td_retval[1] = 0;
152 	}
153 	return (error);
154 }
155 
156 int
157 sys_rfork(struct thread *td, struct rfork_args *uap)
158 {
159 	struct proc *p2;
160 	int error;
161 
162 	/* Don't allow kernel-only flags. */
163 	if ((uap->flags & RFKERNELONLY) != 0)
164 		return (EINVAL);
165 
166 	AUDIT_ARG_FFLAGS(uap->flags);
167 	error = fork1(td, uap->flags, 0, &p2, NULL, 0, NULL);
168 	if (error == 0) {
169 		td->td_retval[0] = p2 ? p2->p_pid : 0;
170 		td->td_retval[1] = 0;
171 	}
172 	return (error);
173 }
174 
175 int	nprocs = 1;		/* process 0 */
176 int	lastpid = 0;
177 SYSCTL_INT(_kern, OID_AUTO, lastpid, CTLFLAG_RD, &lastpid, 0,
178     "Last used PID");
179 
180 /*
181  * Random component to lastpid generation.  We mix in a random factor to make
182  * it a little harder to predict.  We sanity check the modulus value to avoid
183  * doing it in critical paths.  Don't let it be too small or we pointlessly
184  * waste randomness entropy, and don't let it be impossibly large.  Using a
185  * modulus that is too big causes a LOT more process table scans and slows
186  * down fork processing as the pidchecked caching is defeated.
187  */
188 static int randompid = 0;
189 
190 static int
191 sysctl_kern_randompid(SYSCTL_HANDLER_ARGS)
192 {
193 	int error, pid;
194 
195 	error = sysctl_wire_old_buffer(req, sizeof(int));
196 	if (error != 0)
197 		return(error);
198 	sx_xlock(&allproc_lock);
199 	pid = randompid;
200 	error = sysctl_handle_int(oidp, &pid, 0, req);
201 	if (error == 0 && req->newptr != NULL) {
202 		if (pid < 0 || pid > pid_max - 100)	/* out of range */
203 			pid = pid_max - 100;
204 		else if (pid < 2)			/* NOP */
205 			pid = 0;
206 		else if (pid < 100)			/* Make it reasonable */
207 			pid = 100;
208 		randompid = pid;
209 	}
210 	sx_xunlock(&allproc_lock);
211 	return (error);
212 }
213 
214 SYSCTL_PROC(_kern, OID_AUTO, randompid, CTLTYPE_INT|CTLFLAG_RW,
215     0, 0, sysctl_kern_randompid, "I", "Random PID modulus");
216 
217 static int
218 fork_findpid(int flags)
219 {
220 	struct proc *p;
221 	int trypid;
222 	static int pidchecked = 0;
223 
224 	/*
225 	 * Requires allproc_lock in order to iterate over the list
226 	 * of processes, and proctree_lock to access p_pgrp.
227 	 */
228 	sx_assert(&allproc_lock, SX_LOCKED);
229 	sx_assert(&proctree_lock, SX_LOCKED);
230 
231 	/*
232 	 * Find an unused process ID.  We remember a range of unused IDs
233 	 * ready to use (from lastpid+1 through pidchecked-1).
234 	 *
235 	 * If RFHIGHPID is set (used during system boot), do not allocate
236 	 * low-numbered pids.
237 	 */
238 	trypid = lastpid + 1;
239 	if (flags & RFHIGHPID) {
240 		if (trypid < 10)
241 			trypid = 10;
242 	} else {
243 		if (randompid)
244 			trypid += arc4random() % randompid;
245 	}
246 retry:
247 	/*
248 	 * If the process ID prototype has wrapped around,
249 	 * restart somewhat above 0, as the low-numbered procs
250 	 * tend to include daemons that don't exit.
251 	 */
252 	if (trypid >= pid_max) {
253 		trypid = trypid % pid_max;
254 		if (trypid < 100)
255 			trypid += 100;
256 		pidchecked = 0;
257 	}
258 	if (trypid >= pidchecked) {
259 		int doingzomb = 0;
260 
261 		pidchecked = PID_MAX;
262 		/*
263 		 * Scan the active and zombie procs to check whether this pid
264 		 * is in use.  Remember the lowest pid that's greater
265 		 * than trypid, so we can avoid checking for a while.
266 		 *
267 		 * Avoid reuse of the process group id, session id or
268 		 * the reaper subtree id.  Note that for process group
269 		 * and sessions, the amount of reserved pids is
270 		 * limited by process limit.  For the subtree ids, the
271 		 * id is kept reserved only while there is a
272 		 * non-reaped process in the subtree, so amount of
273 		 * reserved pids is limited by process limit times
274 		 * two.
275 		 */
276 		p = LIST_FIRST(&allproc);
277 again:
278 		for (; p != NULL; p = LIST_NEXT(p, p_list)) {
279 			while (p->p_pid == trypid ||
280 			    p->p_reapsubtree == trypid ||
281 			    (p->p_pgrp != NULL &&
282 			    (p->p_pgrp->pg_id == trypid ||
283 			    (p->p_session != NULL &&
284 			    p->p_session->s_sid == trypid)))) {
285 				trypid++;
286 				if (trypid >= pidchecked)
287 					goto retry;
288 			}
289 			if (p->p_pid > trypid && pidchecked > p->p_pid)
290 				pidchecked = p->p_pid;
291 			if (p->p_pgrp != NULL) {
292 				if (p->p_pgrp->pg_id > trypid &&
293 				    pidchecked > p->p_pgrp->pg_id)
294 					pidchecked = p->p_pgrp->pg_id;
295 				if (p->p_session != NULL &&
296 				    p->p_session->s_sid > trypid &&
297 				    pidchecked > p->p_session->s_sid)
298 					pidchecked = p->p_session->s_sid;
299 			}
300 		}
301 		if (!doingzomb) {
302 			doingzomb = 1;
303 			p = LIST_FIRST(&zombproc);
304 			goto again;
305 		}
306 	}
307 
308 	/*
309 	 * RFHIGHPID does not mess with the lastpid counter during boot.
310 	 */
311 	if (flags & RFHIGHPID)
312 		pidchecked = 0;
313 	else
314 		lastpid = trypid;
315 
316 	return (trypid);
317 }
318 
319 static int
320 fork_norfproc(struct thread *td, int flags)
321 {
322 	int error;
323 	struct proc *p1;
324 
325 	KASSERT((flags & RFPROC) == 0,
326 	    ("fork_norfproc called with RFPROC set"));
327 	p1 = td->td_proc;
328 
329 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
330 	    (flags & (RFCFDG | RFFDG))) {
331 		PROC_LOCK(p1);
332 		if (thread_single(p1, SINGLE_BOUNDARY)) {
333 			PROC_UNLOCK(p1);
334 			return (ERESTART);
335 		}
336 		PROC_UNLOCK(p1);
337 	}
338 
339 	error = vm_forkproc(td, NULL, NULL, NULL, flags);
340 	if (error)
341 		goto fail;
342 
343 	/*
344 	 * Close all file descriptors.
345 	 */
346 	if (flags & RFCFDG) {
347 		struct filedesc *fdtmp;
348 		fdtmp = fdinit(td->td_proc->p_fd, false);
349 		fdescfree(td);
350 		p1->p_fd = fdtmp;
351 	}
352 
353 	/*
354 	 * Unshare file descriptors (from parent).
355 	 */
356 	if (flags & RFFDG)
357 		fdunshare(td);
358 
359 fail:
360 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
361 	    (flags & (RFCFDG | RFFDG))) {
362 		PROC_LOCK(p1);
363 		thread_single_end(p1, SINGLE_BOUNDARY);
364 		PROC_UNLOCK(p1);
365 	}
366 	return (error);
367 }
368 
369 static void
370 do_fork(struct thread *td, int flags, struct proc *p2, struct thread *td2,
371     struct vmspace *vm2, int pdflags)
372 {
373 	struct proc *p1, *pptr;
374 	int p2_held, trypid;
375 	struct filedesc *fd;
376 	struct filedesc_to_leader *fdtol;
377 	struct sigacts *newsigacts;
378 
379 	sx_assert(&proctree_lock, SX_SLOCKED);
380 	sx_assert(&allproc_lock, SX_XLOCKED);
381 
382 	p2_held = 0;
383 	p1 = td->td_proc;
384 
385 	trypid = fork_findpid(flags);
386 
387 	sx_sunlock(&proctree_lock);
388 
389 	p2->p_state = PRS_NEW;		/* protect against others */
390 	p2->p_pid = trypid;
391 	AUDIT_ARG_PID(p2->p_pid);
392 	LIST_INSERT_HEAD(&allproc, p2, p_list);
393 	allproc_gen++;
394 	LIST_INSERT_HEAD(PIDHASH(p2->p_pid), p2, p_hash);
395 	tidhash_add(td2);
396 	PROC_LOCK(p2);
397 	PROC_LOCK(p1);
398 
399 	sx_xunlock(&allproc_lock);
400 
401 	bcopy(&p1->p_startcopy, &p2->p_startcopy,
402 	    __rangeof(struct proc, p_startcopy, p_endcopy));
403 	pargs_hold(p2->p_args);
404 
405 	PROC_UNLOCK(p1);
406 
407 	bzero(&p2->p_startzero,
408 	    __rangeof(struct proc, p_startzero, p_endzero));
409 
410 	/* Tell the prison that we exist. */
411 	prison_proc_hold(p2->p_ucred->cr_prison);
412 
413 	PROC_UNLOCK(p2);
414 
415 	/*
416 	 * Malloc things while we don't hold any locks.
417 	 */
418 	if (flags & RFSIGSHARE)
419 		newsigacts = NULL;
420 	else
421 		newsigacts = sigacts_alloc();
422 
423 	/*
424 	 * Copy filedesc.
425 	 */
426 	if (flags & RFCFDG) {
427 		fd = fdinit(p1->p_fd, false);
428 		fdtol = NULL;
429 	} else if (flags & RFFDG) {
430 		fd = fdcopy(p1->p_fd);
431 		fdtol = NULL;
432 	} else {
433 		fd = fdshare(p1->p_fd);
434 		if (p1->p_fdtol == NULL)
435 			p1->p_fdtol = filedesc_to_leader_alloc(NULL, NULL,
436 			    p1->p_leader);
437 		if ((flags & RFTHREAD) != 0) {
438 			/*
439 			 * Shared file descriptor table, and shared
440 			 * process leaders.
441 			 */
442 			fdtol = p1->p_fdtol;
443 			FILEDESC_XLOCK(p1->p_fd);
444 			fdtol->fdl_refcount++;
445 			FILEDESC_XUNLOCK(p1->p_fd);
446 		} else {
447 			/*
448 			 * Shared file descriptor table, and different
449 			 * process leaders.
450 			 */
451 			fdtol = filedesc_to_leader_alloc(p1->p_fdtol,
452 			    p1->p_fd, p2);
453 		}
454 	}
455 	/*
456 	 * Make a proc table entry for the new process.
457 	 * Start by zeroing the section of proc that is zero-initialized,
458 	 * then copy the section that is copied directly from the parent.
459 	 */
460 
461 	PROC_LOCK(p2);
462 	PROC_LOCK(p1);
463 
464 	bzero(&td2->td_startzero,
465 	    __rangeof(struct thread, td_startzero, td_endzero));
466 
467 	bcopy(&td->td_startcopy, &td2->td_startcopy,
468 	    __rangeof(struct thread, td_startcopy, td_endcopy));
469 
470 	bcopy(&p2->p_comm, &td2->td_name, sizeof(td2->td_name));
471 	td2->td_sigstk = td->td_sigstk;
472 	td2->td_flags = TDF_INMEM;
473 	td2->td_lend_user_pri = PRI_MAX;
474 
475 #ifdef VIMAGE
476 	td2->td_vnet = NULL;
477 	td2->td_vnet_lpush = NULL;
478 #endif
479 
480 	/*
481 	 * Allow the scheduler to initialize the child.
482 	 */
483 	thread_lock(td);
484 	sched_fork(td, td2);
485 	thread_unlock(td);
486 
487 	/*
488 	 * Duplicate sub-structures as needed.
489 	 * Increase reference counts on shared objects.
490 	 */
491 	p2->p_flag = P_INMEM;
492 	p2->p_flag2 = p1->p_flag2 & (P2_NOTRACE | P2_NOTRACE_EXEC);
493 	p2->p_swtick = ticks;
494 	if (p1->p_flag & P_PROFIL)
495 		startprofclock(p2);
496 
497 	/*
498 	 * Whilst the proc lock is held, copy the VM domain data out
499 	 * using the VM domain method.
500 	 */
501 	vm_domain_policy_init(&p2->p_vm_dom_policy);
502 	vm_domain_policy_localcopy(&p2->p_vm_dom_policy,
503 	    &p1->p_vm_dom_policy);
504 
505 	if (flags & RFSIGSHARE) {
506 		p2->p_sigacts = sigacts_hold(p1->p_sigacts);
507 	} else {
508 		sigacts_copy(newsigacts, p1->p_sigacts);
509 		p2->p_sigacts = newsigacts;
510 	}
511 
512 	if (flags & RFTSIGZMB)
513 	        p2->p_sigparent = RFTSIGNUM(flags);
514 	else if (flags & RFLINUXTHPN)
515 	        p2->p_sigparent = SIGUSR1;
516 	else
517 	        p2->p_sigparent = SIGCHLD;
518 
519 	p2->p_textvp = p1->p_textvp;
520 	p2->p_fd = fd;
521 	p2->p_fdtol = fdtol;
522 
523 	if (p1->p_flag2 & P2_INHERIT_PROTECTED) {
524 		p2->p_flag |= P_PROTECTED;
525 		p2->p_flag2 |= P2_INHERIT_PROTECTED;
526 	}
527 
528 	/*
529 	 * p_limit is copy-on-write.  Bump its refcount.
530 	 */
531 	lim_fork(p1, p2);
532 
533 	thread_cow_get_proc(td2, p2);
534 
535 	pstats_fork(p1->p_stats, p2->p_stats);
536 
537 	PROC_UNLOCK(p1);
538 	PROC_UNLOCK(p2);
539 
540 	/* Bump references to the text vnode (for procfs). */
541 	if (p2->p_textvp)
542 		vref(p2->p_textvp);
543 
544 	/*
545 	 * Set up linkage for kernel based threading.
546 	 */
547 	if ((flags & RFTHREAD) != 0) {
548 		mtx_lock(&ppeers_lock);
549 		p2->p_peers = p1->p_peers;
550 		p1->p_peers = p2;
551 		p2->p_leader = p1->p_leader;
552 		mtx_unlock(&ppeers_lock);
553 		PROC_LOCK(p1->p_leader);
554 		if ((p1->p_leader->p_flag & P_WEXIT) != 0) {
555 			PROC_UNLOCK(p1->p_leader);
556 			/*
557 			 * The task leader is exiting, so process p1 is
558 			 * going to be killed shortly.  Since p1 obviously
559 			 * isn't dead yet, we know that the leader is either
560 			 * sending SIGKILL's to all the processes in this
561 			 * task or is sleeping waiting for all the peers to
562 			 * exit.  We let p1 complete the fork, but we need
563 			 * to go ahead and kill the new process p2 since
564 			 * the task leader may not get a chance to send
565 			 * SIGKILL to it.  We leave it on the list so that
566 			 * the task leader will wait for this new process
567 			 * to commit suicide.
568 			 */
569 			PROC_LOCK(p2);
570 			kern_psignal(p2, SIGKILL);
571 			PROC_UNLOCK(p2);
572 		} else
573 			PROC_UNLOCK(p1->p_leader);
574 	} else {
575 		p2->p_peers = NULL;
576 		p2->p_leader = p2;
577 	}
578 
579 	sx_xlock(&proctree_lock);
580 	PGRP_LOCK(p1->p_pgrp);
581 	PROC_LOCK(p2);
582 	PROC_LOCK(p1);
583 
584 	/*
585 	 * Preserve some more flags in subprocess.  P_PROFIL has already
586 	 * been preserved.
587 	 */
588 	p2->p_flag |= p1->p_flag & P_SUGID;
589 	td2->td_pflags |= (td->td_pflags & TDP_ALTSTACK) | TDP_FORKING;
590 	SESS_LOCK(p1->p_session);
591 	if (p1->p_session->s_ttyvp != NULL && p1->p_flag & P_CONTROLT)
592 		p2->p_flag |= P_CONTROLT;
593 	SESS_UNLOCK(p1->p_session);
594 	if (flags & RFPPWAIT)
595 		p2->p_flag |= P_PPWAIT;
596 
597 	p2->p_pgrp = p1->p_pgrp;
598 	LIST_INSERT_AFTER(p1, p2, p_pglist);
599 	PGRP_UNLOCK(p1->p_pgrp);
600 	LIST_INIT(&p2->p_children);
601 	LIST_INIT(&p2->p_orphans);
602 
603 	callout_init_mtx(&p2->p_itcallout, &p2->p_mtx, 0);
604 
605 	/*
606 	 * If PF_FORK is set, the child process inherits the
607 	 * procfs ioctl flags from its parent.
608 	 */
609 	if (p1->p_pfsflags & PF_FORK) {
610 		p2->p_stops = p1->p_stops;
611 		p2->p_pfsflags = p1->p_pfsflags;
612 	}
613 
614 	/*
615 	 * This begins the section where we must prevent the parent
616 	 * from being swapped.
617 	 */
618 	_PHOLD(p1);
619 	PROC_UNLOCK(p1);
620 
621 	/*
622 	 * Attach the new process to its parent.
623 	 *
624 	 * If RFNOWAIT is set, the newly created process becomes a child
625 	 * of init.  This effectively disassociates the child from the
626 	 * parent.
627 	 */
628 	if ((flags & RFNOWAIT) != 0) {
629 		pptr = p1->p_reaper;
630 		p2->p_reaper = pptr;
631 	} else {
632 		p2->p_reaper = (p1->p_treeflag & P_TREE_REAPER) != 0 ?
633 		    p1 : p1->p_reaper;
634 		pptr = p1;
635 	}
636 	p2->p_pptr = pptr;
637 	LIST_INSERT_HEAD(&pptr->p_children, p2, p_sibling);
638 	LIST_INIT(&p2->p_reaplist);
639 	LIST_INSERT_HEAD(&p2->p_reaper->p_reaplist, p2, p_reapsibling);
640 	if (p2->p_reaper == p1)
641 		p2->p_reapsubtree = p2->p_pid;
642 	sx_xunlock(&proctree_lock);
643 
644 	/* Inform accounting that we have forked. */
645 	p2->p_acflag = AFORK;
646 	PROC_UNLOCK(p2);
647 
648 #ifdef KTRACE
649 	ktrprocfork(p1, p2);
650 #endif
651 
652 	/*
653 	 * Finish creating the child process.  It will return via a different
654 	 * execution path later.  (ie: directly into user mode)
655 	 */
656 	vm_forkproc(td, p2, td2, vm2, flags);
657 
658 	if (flags == (RFFDG | RFPROC)) {
659 		PCPU_INC(cnt.v_forks);
660 		PCPU_ADD(cnt.v_forkpages, p2->p_vmspace->vm_dsize +
661 		    p2->p_vmspace->vm_ssize);
662 	} else if (flags == (RFFDG | RFPROC | RFPPWAIT | RFMEM)) {
663 		PCPU_INC(cnt.v_vforks);
664 		PCPU_ADD(cnt.v_vforkpages, p2->p_vmspace->vm_dsize +
665 		    p2->p_vmspace->vm_ssize);
666 	} else if (p1 == &proc0) {
667 		PCPU_INC(cnt.v_kthreads);
668 		PCPU_ADD(cnt.v_kthreadpages, p2->p_vmspace->vm_dsize +
669 		    p2->p_vmspace->vm_ssize);
670 	} else {
671 		PCPU_INC(cnt.v_rforks);
672 		PCPU_ADD(cnt.v_rforkpages, p2->p_vmspace->vm_dsize +
673 		    p2->p_vmspace->vm_ssize);
674 	}
675 
676 	/*
677 	 * Associate the process descriptor with the process before anything
678 	 * can happen that might cause that process to need the descriptor.
679 	 * However, don't do this until after fork(2) can no longer fail.
680 	 */
681 	if (flags & RFPROCDESC)
682 		procdesc_new(p2, pdflags);
683 
684 	/*
685 	 * Both processes are set up, now check if any loadable modules want
686 	 * to adjust anything.
687 	 */
688 	EVENTHANDLER_INVOKE(process_fork, p1, p2, flags);
689 
690 	/*
691 	 * Set the child start time and mark the process as being complete.
692 	 */
693 	PROC_LOCK(p2);
694 	PROC_LOCK(p1);
695 	microuptime(&p2->p_stats->p_start);
696 	PROC_SLOCK(p2);
697 	p2->p_state = PRS_NORMAL;
698 	PROC_SUNLOCK(p2);
699 
700 #ifdef KDTRACE_HOOKS
701 	/*
702 	 * Tell the DTrace fasttrap provider about the new process so that any
703 	 * tracepoints inherited from the parent can be removed. We have to do
704 	 * this only after p_state is PRS_NORMAL since the fasttrap module will
705 	 * use pfind() later on.
706 	 */
707 	if ((flags & RFMEM) == 0 && dtrace_fasttrap_fork)
708 		dtrace_fasttrap_fork(p1, p2);
709 #endif
710 	if ((p1->p_flag & (P_TRACED | P_FOLLOWFORK)) == (P_TRACED |
711 	    P_FOLLOWFORK)) {
712 		/*
713 		 * Arrange for debugger to receive the fork event.
714 		 *
715 		 * We can report PL_FLAG_FORKED regardless of
716 		 * P_FOLLOWFORK settings, but it does not make a sense
717 		 * for runaway child.
718 		 */
719 		td->td_dbgflags |= TDB_FORK;
720 		td->td_dbg_forked = p2->p_pid;
721 		td2->td_dbgflags |= TDB_STOPATFORK;
722 		_PHOLD(p2);
723 		p2_held = 1;
724 	}
725 	if (flags & RFPPWAIT) {
726 		td->td_pflags |= TDP_RFPPWAIT;
727 		td->td_rfppwait_p = p2;
728 	}
729 	PROC_UNLOCK(p2);
730 	if ((flags & RFSTOPPED) == 0) {
731 		/*
732 		 * If RFSTOPPED not requested, make child runnable and
733 		 * add to run queue.
734 		 */
735 		thread_lock(td2);
736 		TD_SET_CAN_RUN(td2);
737 		sched_add(td2, SRQ_BORING);
738 		thread_unlock(td2);
739 	}
740 
741 	/*
742 	 * Now can be swapped.
743 	 */
744 	_PRELE(p1);
745 	PROC_UNLOCK(p1);
746 
747 	/*
748 	 * Tell any interested parties about the new process.
749 	 */
750 	knote_fork(&p1->p_klist, p2->p_pid);
751 	SDT_PROBE3(proc, kernel, , create, p2, p1, flags);
752 
753 	/*
754 	 * Wait until debugger is attached to child.
755 	 */
756 	PROC_LOCK(p2);
757 	while ((td2->td_dbgflags & TDB_STOPATFORK) != 0)
758 		cv_wait(&p2->p_dbgwait, &p2->p_mtx);
759 	if (p2_held)
760 		_PRELE(p2);
761 	PROC_UNLOCK(p2);
762 }
763 
764 int
765 fork1(struct thread *td, int flags, int pages, struct proc **procp,
766     int *procdescp, int pdflags, struct filecaps *fcaps)
767 {
768 	struct proc *p1, *newproc;
769 	struct thread *td2;
770 	struct vmspace *vm2;
771 	struct file *fp_procdesc;
772 	vm_ooffset_t mem_charged;
773 	int error, nprocs_new, ok;
774 	static int curfail;
775 	static struct timeval lastfail;
776 
777 	/* Check for the undefined or unimplemented flags. */
778 	if ((flags & ~(RFFLAGS | RFTSIGFLAGS(RFTSIGMASK))) != 0)
779 		return (EINVAL);
780 
781 	/* Signal value requires RFTSIGZMB. */
782 	if ((flags & RFTSIGFLAGS(RFTSIGMASK)) != 0 && (flags & RFTSIGZMB) == 0)
783 		return (EINVAL);
784 
785 	/* Can't copy and clear. */
786 	if ((flags & (RFFDG|RFCFDG)) == (RFFDG|RFCFDG))
787 		return (EINVAL);
788 
789 	/* Check the validity of the signal number. */
790 	if ((flags & RFTSIGZMB) != 0 && (u_int)RFTSIGNUM(flags) > _SIG_MAXSIG)
791 		return (EINVAL);
792 
793 	if ((flags & RFPROCDESC) != 0) {
794 		/* Can't not create a process yet get a process descriptor. */
795 		if ((flags & RFPROC) == 0)
796 			return (EINVAL);
797 
798 		/* Must provide a place to put a procdesc if creating one. */
799 		if (procdescp == NULL)
800 			return (EINVAL);
801 	}
802 
803 	p1 = td->td_proc;
804 
805 	/*
806 	 * Here we don't create a new process, but we divorce
807 	 * certain parts of a process from itself.
808 	 */
809 	if ((flags & RFPROC) == 0) {
810 		*procp = NULL;
811 		return (fork_norfproc(td, flags));
812 	}
813 
814 	fp_procdesc = NULL;
815 	newproc = NULL;
816 	vm2 = NULL;
817 
818 	/*
819 	 * Increment the nprocs resource before allocations occur.
820 	 * Although process entries are dynamically created, we still
821 	 * keep a global limit on the maximum number we will
822 	 * create. There are hard-limits as to the number of processes
823 	 * that can run, established by the KVA and memory usage for
824 	 * the process data.
825 	 *
826 	 * Don't allow a nonprivileged user to use the last ten
827 	 * processes; don't let root exceed the limit.
828 	 */
829 	nprocs_new = atomic_fetchadd_int(&nprocs, 1) + 1;
830 	if ((nprocs_new >= maxproc - 10 && priv_check_cred(td->td_ucred,
831 	    PRIV_MAXPROC, 0) != 0) || nprocs_new >= maxproc) {
832 		error = EAGAIN;
833 		sx_xlock(&allproc_lock);
834 		if (ppsratecheck(&lastfail, &curfail, 1)) {
835 			printf("maxproc limit exceeded by uid %u (pid %d); "
836 			    "see tuning(7) and login.conf(5)\n",
837 			    td->td_ucred->cr_ruid, p1->p_pid);
838 		}
839 		sx_xunlock(&allproc_lock);
840 		goto fail2;
841 	}
842 
843 	/*
844 	 * If required, create a process descriptor in the parent first; we
845 	 * will abandon it if something goes wrong. We don't finit() until
846 	 * later.
847 	 */
848 	if (flags & RFPROCDESC) {
849 		error = falloc_caps(td, &fp_procdesc, procdescp, 0, fcaps);
850 		if (error != 0)
851 			goto fail2;
852 	}
853 
854 	mem_charged = 0;
855 	if (pages == 0)
856 		pages = kstack_pages;
857 	/* Allocate new proc. */
858 	newproc = uma_zalloc(proc_zone, M_WAITOK);
859 	td2 = FIRST_THREAD_IN_PROC(newproc);
860 	if (td2 == NULL) {
861 		td2 = thread_alloc(pages);
862 		if (td2 == NULL) {
863 			error = ENOMEM;
864 			goto fail2;
865 		}
866 		proc_linkup(newproc, td2);
867 	} else {
868 		if (td2->td_kstack == 0 || td2->td_kstack_pages != pages) {
869 			if (td2->td_kstack != 0)
870 				vm_thread_dispose(td2);
871 			if (!thread_alloc_stack(td2, pages)) {
872 				error = ENOMEM;
873 				goto fail2;
874 			}
875 		}
876 	}
877 
878 	if ((flags & RFMEM) == 0) {
879 		vm2 = vmspace_fork(p1->p_vmspace, &mem_charged);
880 		if (vm2 == NULL) {
881 			error = ENOMEM;
882 			goto fail2;
883 		}
884 		if (!swap_reserve(mem_charged)) {
885 			/*
886 			 * The swap reservation failed. The accounting
887 			 * from the entries of the copied vm2 will be
888 			 * substracted in vmspace_free(), so force the
889 			 * reservation there.
890 			 */
891 			swap_reserve_force(mem_charged);
892 			error = ENOMEM;
893 			goto fail2;
894 		}
895 	} else
896 		vm2 = NULL;
897 
898 	/*
899 	 * XXX: This is ugly; when we copy resource usage, we need to bump
900 	 *      per-cred resource counters.
901 	 */
902 	proc_set_cred_init(newproc, crhold(td->td_ucred));
903 
904 	/*
905 	 * Initialize resource accounting for the child process.
906 	 */
907 	error = racct_proc_fork(p1, newproc);
908 	if (error != 0) {
909 		error = EAGAIN;
910 		goto fail1;
911 	}
912 
913 #ifdef MAC
914 	mac_proc_init(newproc);
915 #endif
916 	knlist_init_mtx(&newproc->p_klist, &newproc->p_mtx);
917 	STAILQ_INIT(&newproc->p_ktr);
918 
919 	/* We have to lock the process tree while we look for a pid. */
920 	sx_slock(&proctree_lock);
921 	sx_xlock(&allproc_lock);
922 
923 	/*
924 	 * Increment the count of procs running with this uid. Don't allow
925 	 * a nonprivileged user to exceed their current limit.
926 	 *
927 	 * XXXRW: Can we avoid privilege here if it's not needed?
928 	 */
929 	error = priv_check_cred(td->td_ucred, PRIV_PROC_LIMIT, 0);
930 	if (error == 0)
931 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1, 0);
932 	else {
933 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1,
934 		    lim_cur(td, RLIMIT_NPROC));
935 	}
936 	if (ok) {
937 		do_fork(td, flags, newproc, td2, vm2, pdflags);
938 
939 		/*
940 		 * Return child proc pointer to parent.
941 		 */
942 		*procp = newproc;
943 		if (flags & RFPROCDESC) {
944 			procdesc_finit(newproc->p_procdesc, fp_procdesc);
945 			fdrop(fp_procdesc, td);
946 		}
947 		racct_proc_fork_done(newproc);
948 		return (0);
949 	}
950 
951 	error = EAGAIN;
952 	sx_sunlock(&proctree_lock);
953 	sx_xunlock(&allproc_lock);
954 #ifdef MAC
955 	mac_proc_destroy(newproc);
956 #endif
957 	racct_proc_exit(newproc);
958 fail1:
959 	crfree(newproc->p_ucred);
960 	newproc->p_ucred = NULL;
961 fail2:
962 	if (vm2 != NULL)
963 		vmspace_free(vm2);
964 	uma_zfree(proc_zone, newproc);
965 	if ((flags & RFPROCDESC) != 0 && fp_procdesc != NULL) {
966 		fdclose(td, fp_procdesc, *procdescp);
967 		fdrop(fp_procdesc, td);
968 	}
969 	atomic_add_int(&nprocs, -1);
970 	pause("fork", hz / 2);
971 	return (error);
972 }
973 
974 /*
975  * Handle the return of a child process from fork1().  This function
976  * is called from the MD fork_trampoline() entry point.
977  */
978 void
979 fork_exit(void (*callout)(void *, struct trapframe *), void *arg,
980     struct trapframe *frame)
981 {
982 	struct proc *p;
983 	struct thread *td;
984 	struct thread *dtd;
985 
986 	td = curthread;
987 	p = td->td_proc;
988 	KASSERT(p->p_state == PRS_NORMAL, ("executing process is still new"));
989 
990 	CTR4(KTR_PROC, "fork_exit: new thread %p (td_sched %p, pid %d, %s)",
991 		td, td->td_sched, p->p_pid, td->td_name);
992 
993 	sched_fork_exit(td);
994 	/*
995 	* Processes normally resume in mi_switch() after being
996 	* cpu_switch()'ed to, but when children start up they arrive here
997 	* instead, so we must do much the same things as mi_switch() would.
998 	*/
999 	if ((dtd = PCPU_GET(deadthread))) {
1000 		PCPU_SET(deadthread, NULL);
1001 		thread_stash(dtd);
1002 	}
1003 	thread_unlock(td);
1004 
1005 	/*
1006 	 * cpu_set_fork_handler intercepts this function call to
1007 	 * have this call a non-return function to stay in kernel mode.
1008 	 * initproc has its own fork handler, but it does return.
1009 	 */
1010 	KASSERT(callout != NULL, ("NULL callout in fork_exit"));
1011 	callout(arg, frame);
1012 
1013 	/*
1014 	 * Check if a kernel thread misbehaved and returned from its main
1015 	 * function.
1016 	 */
1017 	if (p->p_flag & P_KTHREAD) {
1018 		printf("Kernel thread \"%s\" (pid %d) exited prematurely.\n",
1019 		    td->td_name, p->p_pid);
1020 		kproc_exit(0);
1021 	}
1022 	mtx_assert(&Giant, MA_NOTOWNED);
1023 
1024 	if (p->p_sysent->sv_schedtail != NULL)
1025 		(p->p_sysent->sv_schedtail)(td);
1026 	td->td_pflags &= ~TDP_FORKING;
1027 }
1028 
1029 /*
1030  * Simplified back end of syscall(), used when returning from fork()
1031  * directly into user mode.  Giant is not held on entry, and must not
1032  * be held on return.  This function is passed in to fork_exit() as the
1033  * first parameter and is called when returning to a new userland process.
1034  */
1035 void
1036 fork_return(struct thread *td, struct trapframe *frame)
1037 {
1038 	struct proc *p, *dbg;
1039 
1040 	p = td->td_proc;
1041 	if (td->td_dbgflags & TDB_STOPATFORK) {
1042 		sx_xlock(&proctree_lock);
1043 		PROC_LOCK(p);
1044 		if ((p->p_pptr->p_flag & (P_TRACED | P_FOLLOWFORK)) ==
1045 		    (P_TRACED | P_FOLLOWFORK)) {
1046 			/*
1047 			 * If debugger still wants auto-attach for the
1048 			 * parent's children, do it now.
1049 			 */
1050 			dbg = p->p_pptr->p_pptr;
1051 			p->p_flag |= P_TRACED;
1052 			p->p_oppid = p->p_pptr->p_pid;
1053 			CTR2(KTR_PTRACE,
1054 		    "fork_return: attaching to new child pid %d: oppid %d",
1055 			    p->p_pid, p->p_oppid);
1056 			proc_reparent(p, dbg);
1057 			sx_xunlock(&proctree_lock);
1058 			td->td_dbgflags |= TDB_CHILD | TDB_SCX;
1059 			ptracestop(td, SIGSTOP);
1060 			td->td_dbgflags &= ~(TDB_CHILD | TDB_SCX);
1061 		} else {
1062 			/*
1063 			 * ... otherwise clear the request.
1064 			 */
1065 			sx_xunlock(&proctree_lock);
1066 			td->td_dbgflags &= ~TDB_STOPATFORK;
1067 			cv_broadcast(&p->p_dbgwait);
1068 		}
1069 		PROC_UNLOCK(p);
1070 	} else if (p->p_flag & P_TRACED) {
1071  		/*
1072 		 * This is the start of a new thread in a traced
1073 		 * process.  Report a system call exit event.
1074 		 */
1075 		PROC_LOCK(p);
1076 		td->td_dbgflags |= TDB_SCX;
1077 		_STOPEVENT(p, S_SCX, td->td_dbg_sc_code);
1078 		if ((p->p_stops & S_PT_SCX) != 0)
1079 			ptracestop(td, SIGTRAP);
1080 		td->td_dbgflags &= ~TDB_SCX;
1081 		PROC_UNLOCK(p);
1082 	}
1083 
1084 	userret(td, frame);
1085 
1086 #ifdef KTRACE
1087 	if (KTRPOINT(td, KTR_SYSRET))
1088 		ktrsysret(SYS_fork, 0, 0);
1089 #endif
1090 }
1091