xref: /freebsd/sys/kern/kern_fork.c (revision b6a05070fa77edc7ce6e60b61623fd806e807be6)
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, , , create, "struct proc *", "struct proc *", "int");
93 
94 #ifndef _SYS_SYSPROTO_H_
95 struct fork_args {
96 	int     dummy;
97 };
98 #endif
99 
100 /* ARGSUSED */
101 int
102 sys_fork(struct thread *td, struct fork_args *uap)
103 {
104 	int error;
105 	struct proc *p2;
106 
107 	error = fork1(td, RFFDG | RFPROC, 0, &p2, NULL, 0, NULL);
108 	if (error == 0) {
109 		td->td_retval[0] = p2->p_pid;
110 		td->td_retval[1] = 0;
111 	}
112 	return (error);
113 }
114 
115 /* ARGUSED */
116 int
117 sys_pdfork(td, uap)
118 	struct thread *td;
119 	struct pdfork_args *uap;
120 {
121 	int error, fd;
122 	struct proc *p2;
123 
124 	/*
125 	 * It is necessary to return fd by reference because 0 is a valid file
126 	 * descriptor number, and the child needs to be able to distinguish
127 	 * itself from the parent using the return value.
128 	 */
129 	error = fork1(td, RFFDG | RFPROC | RFPROCDESC, 0, &p2,
130 	    &fd, uap->flags, NULL);
131 	if (error == 0) {
132 		td->td_retval[0] = p2->p_pid;
133 		td->td_retval[1] = 0;
134 		error = copyout(&fd, uap->fdp, sizeof(fd));
135 	}
136 	return (error);
137 }
138 
139 /* ARGSUSED */
140 int
141 sys_vfork(struct thread *td, struct vfork_args *uap)
142 {
143 	int error, flags;
144 	struct proc *p2;
145 
146 	flags = RFFDG | RFPROC | RFPPWAIT | RFMEM;
147 	error = fork1(td, flags, 0, &p2, NULL, 0, NULL);
148 	if (error == 0) {
149 		td->td_retval[0] = p2->p_pid;
150 		td->td_retval[1] = 0;
151 	}
152 	return (error);
153 }
154 
155 int
156 sys_rfork(struct thread *td, struct rfork_args *uap)
157 {
158 	struct proc *p2;
159 	int error;
160 
161 	/* Don't allow kernel-only flags. */
162 	if ((uap->flags & RFKERNELONLY) != 0)
163 		return (EINVAL);
164 
165 	AUDIT_ARG_FFLAGS(uap->flags);
166 	error = fork1(td, uap->flags, 0, &p2, NULL, 0, NULL);
167 	if (error == 0) {
168 		td->td_retval[0] = p2 ? p2->p_pid : 0;
169 		td->td_retval[1] = 0;
170 	}
171 	return (error);
172 }
173 
174 int	nprocs = 1;		/* process 0 */
175 int	lastpid = 0;
176 SYSCTL_INT(_kern, OID_AUTO, lastpid, CTLFLAG_RD, &lastpid, 0,
177     "Last used PID");
178 
179 /*
180  * Random component to lastpid generation.  We mix in a random factor to make
181  * it a little harder to predict.  We sanity check the modulus value to avoid
182  * doing it in critical paths.  Don't let it be too small or we pointlessly
183  * waste randomness entropy, and don't let it be impossibly large.  Using a
184  * modulus that is too big causes a LOT more process table scans and slows
185  * down fork processing as the pidchecked caching is defeated.
186  */
187 static int randompid = 0;
188 
189 static int
190 sysctl_kern_randompid(SYSCTL_HANDLER_ARGS)
191 {
192 	int error, pid;
193 
194 	error = sysctl_wire_old_buffer(req, sizeof(int));
195 	if (error != 0)
196 		return(error);
197 	sx_xlock(&allproc_lock);
198 	pid = randompid;
199 	error = sysctl_handle_int(oidp, &pid, 0, req);
200 	if (error == 0 && req->newptr != NULL) {
201 		if (pid < 0 || pid > pid_max - 100)	/* out of range */
202 			pid = pid_max - 100;
203 		else if (pid < 2)			/* NOP */
204 			pid = 0;
205 		else if (pid < 100)			/* Make it reasonable */
206 			pid = 100;
207 		randompid = pid;
208 	}
209 	sx_xunlock(&allproc_lock);
210 	return (error);
211 }
212 
213 SYSCTL_PROC(_kern, OID_AUTO, randompid, CTLTYPE_INT|CTLFLAG_RW,
214     0, 0, sysctl_kern_randompid, "I", "Random PID modulus");
215 
216 static int
217 fork_findpid(int flags)
218 {
219 	struct proc *p;
220 	int trypid;
221 	static int pidchecked = 0;
222 
223 	/*
224 	 * Requires allproc_lock in order to iterate over the list
225 	 * of processes, and proctree_lock to access p_pgrp.
226 	 */
227 	sx_assert(&allproc_lock, SX_LOCKED);
228 	sx_assert(&proctree_lock, SX_LOCKED);
229 
230 	/*
231 	 * Find an unused process ID.  We remember a range of unused IDs
232 	 * ready to use (from lastpid+1 through pidchecked-1).
233 	 *
234 	 * If RFHIGHPID is set (used during system boot), do not allocate
235 	 * low-numbered pids.
236 	 */
237 	trypid = lastpid + 1;
238 	if (flags & RFHIGHPID) {
239 		if (trypid < 10)
240 			trypid = 10;
241 	} else {
242 		if (randompid)
243 			trypid += arc4random() % randompid;
244 	}
245 retry:
246 	/*
247 	 * If the process ID prototype has wrapped around,
248 	 * restart somewhat above 0, as the low-numbered procs
249 	 * tend to include daemons that don't exit.
250 	 */
251 	if (trypid >= pid_max) {
252 		trypid = trypid % pid_max;
253 		if (trypid < 100)
254 			trypid += 100;
255 		pidchecked = 0;
256 	}
257 	if (trypid >= pidchecked) {
258 		int doingzomb = 0;
259 
260 		pidchecked = PID_MAX;
261 		/*
262 		 * Scan the active and zombie procs to check whether this pid
263 		 * is in use.  Remember the lowest pid that's greater
264 		 * than trypid, so we can avoid checking for a while.
265 		 *
266 		 * Avoid reuse of the process group id, session id or
267 		 * the reaper subtree id.  Note that for process group
268 		 * and sessions, the amount of reserved pids is
269 		 * limited by process limit.  For the subtree ids, the
270 		 * id is kept reserved only while there is a
271 		 * non-reaped process in the subtree, so amount of
272 		 * reserved pids is limited by process limit times
273 		 * two.
274 		 */
275 		p = LIST_FIRST(&allproc);
276 again:
277 		for (; p != NULL; p = LIST_NEXT(p, p_list)) {
278 			while (p->p_pid == trypid ||
279 			    p->p_reapsubtree == trypid ||
280 			    (p->p_pgrp != NULL &&
281 			    (p->p_pgrp->pg_id == trypid ||
282 			    (p->p_session != NULL &&
283 			    p->p_session->s_sid == trypid)))) {
284 				trypid++;
285 				if (trypid >= pidchecked)
286 					goto retry;
287 			}
288 			if (p->p_pid > trypid && pidchecked > p->p_pid)
289 				pidchecked = p->p_pid;
290 			if (p->p_pgrp != NULL) {
291 				if (p->p_pgrp->pg_id > trypid &&
292 				    pidchecked > p->p_pgrp->pg_id)
293 					pidchecked = p->p_pgrp->pg_id;
294 				if (p->p_session != NULL &&
295 				    p->p_session->s_sid > trypid &&
296 				    pidchecked > p->p_session->s_sid)
297 					pidchecked = p->p_session->s_sid;
298 			}
299 		}
300 		if (!doingzomb) {
301 			doingzomb = 1;
302 			p = LIST_FIRST(&zombproc);
303 			goto again;
304 		}
305 	}
306 
307 	/*
308 	 * RFHIGHPID does not mess with the lastpid counter during boot.
309 	 */
310 	if (flags & RFHIGHPID)
311 		pidchecked = 0;
312 	else
313 		lastpid = trypid;
314 
315 	return (trypid);
316 }
317 
318 static int
319 fork_norfproc(struct thread *td, int flags)
320 {
321 	int error;
322 	struct proc *p1;
323 
324 	KASSERT((flags & RFPROC) == 0,
325 	    ("fork_norfproc called with RFPROC set"));
326 	p1 = td->td_proc;
327 
328 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
329 	    (flags & (RFCFDG | RFFDG))) {
330 		PROC_LOCK(p1);
331 		if (thread_single(p1, SINGLE_BOUNDARY)) {
332 			PROC_UNLOCK(p1);
333 			return (ERESTART);
334 		}
335 		PROC_UNLOCK(p1);
336 	}
337 
338 	error = vm_forkproc(td, NULL, NULL, NULL, flags);
339 	if (error)
340 		goto fail;
341 
342 	/*
343 	 * Close all file descriptors.
344 	 */
345 	if (flags & RFCFDG) {
346 		struct filedesc *fdtmp;
347 		fdtmp = fdinit(td->td_proc->p_fd, false);
348 		fdescfree(td);
349 		p1->p_fd = fdtmp;
350 	}
351 
352 	/*
353 	 * Unshare file descriptors (from parent).
354 	 */
355 	if (flags & RFFDG)
356 		fdunshare(td);
357 
358 fail:
359 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
360 	    (flags & (RFCFDG | RFFDG))) {
361 		PROC_LOCK(p1);
362 		thread_single_end(p1, SINGLE_BOUNDARY);
363 		PROC_UNLOCK(p1);
364 	}
365 	return (error);
366 }
367 
368 static void
369 do_fork(struct thread *td, int flags, struct proc *p2, struct thread *td2,
370     struct vmspace *vm2, int pdflags)
371 {
372 	struct proc *p1, *pptr;
373 	int p2_held, trypid;
374 	struct filedesc *fd;
375 	struct filedesc_to_leader *fdtol;
376 	struct sigacts *newsigacts;
377 
378 	sx_assert(&proctree_lock, SX_SLOCKED);
379 	sx_assert(&allproc_lock, SX_XLOCKED);
380 
381 	p2_held = 0;
382 	p1 = td->td_proc;
383 
384 	trypid = fork_findpid(flags);
385 
386 	sx_sunlock(&proctree_lock);
387 
388 	p2->p_state = PRS_NEW;		/* protect against others */
389 	p2->p_pid = trypid;
390 	AUDIT_ARG_PID(p2->p_pid);
391 	LIST_INSERT_HEAD(&allproc, p2, p_list);
392 	allproc_gen++;
393 	LIST_INSERT_HEAD(PIDHASH(p2->p_pid), p2, p_hash);
394 	tidhash_add(td2);
395 	PROC_LOCK(p2);
396 	PROC_LOCK(p1);
397 
398 	sx_xunlock(&allproc_lock);
399 
400 	bcopy(&p1->p_startcopy, &p2->p_startcopy,
401 	    __rangeof(struct proc, p_startcopy, p_endcopy));
402 	pargs_hold(p2->p_args);
403 
404 	PROC_UNLOCK(p1);
405 
406 	bzero(&p2->p_startzero,
407 	    __rangeof(struct proc, p_startzero, p_endzero));
408 
409 	/* Tell the prison that we exist. */
410 	prison_proc_hold(p2->p_ucred->cr_prison);
411 
412 	PROC_UNLOCK(p2);
413 
414 	/*
415 	 * Malloc things while we don't hold any locks.
416 	 */
417 	if (flags & RFSIGSHARE)
418 		newsigacts = NULL;
419 	else
420 		newsigacts = sigacts_alloc();
421 
422 	/*
423 	 * Copy filedesc.
424 	 */
425 	if (flags & RFCFDG) {
426 		fd = fdinit(p1->p_fd, false);
427 		fdtol = NULL;
428 	} else if (flags & RFFDG) {
429 		fd = fdcopy(p1->p_fd);
430 		fdtol = NULL;
431 	} else {
432 		fd = fdshare(p1->p_fd);
433 		if (p1->p_fdtol == NULL)
434 			p1->p_fdtol = filedesc_to_leader_alloc(NULL, NULL,
435 			    p1->p_leader);
436 		if ((flags & RFTHREAD) != 0) {
437 			/*
438 			 * Shared file descriptor table, and shared
439 			 * process leaders.
440 			 */
441 			fdtol = p1->p_fdtol;
442 			FILEDESC_XLOCK(p1->p_fd);
443 			fdtol->fdl_refcount++;
444 			FILEDESC_XUNLOCK(p1->p_fd);
445 		} else {
446 			/*
447 			 * Shared file descriptor table, and different
448 			 * process leaders.
449 			 */
450 			fdtol = filedesc_to_leader_alloc(p1->p_fdtol,
451 			    p1->p_fd, p2);
452 		}
453 	}
454 	/*
455 	 * Make a proc table entry for the new process.
456 	 * Start by zeroing the section of proc that is zero-initialized,
457 	 * then copy the section that is copied directly from the parent.
458 	 */
459 
460 	PROC_LOCK(p2);
461 	PROC_LOCK(p1);
462 
463 	bzero(&td2->td_startzero,
464 	    __rangeof(struct thread, td_startzero, td_endzero));
465 
466 	bcopy(&td->td_startcopy, &td2->td_startcopy,
467 	    __rangeof(struct thread, td_startcopy, td_endcopy));
468 
469 	bcopy(&p2->p_comm, &td2->td_name, sizeof(td2->td_name));
470 	td2->td_sigstk = td->td_sigstk;
471 	td2->td_flags = TDF_INMEM;
472 	td2->td_lend_user_pri = PRI_MAX;
473 
474 #ifdef VIMAGE
475 	td2->td_vnet = NULL;
476 	td2->td_vnet_lpush = NULL;
477 #endif
478 
479 	/*
480 	 * Allow the scheduler to initialize the child.
481 	 */
482 	thread_lock(td);
483 	sched_fork(td, td2);
484 	thread_unlock(td);
485 
486 	/*
487 	 * Duplicate sub-structures as needed.
488 	 * Increase reference counts on shared objects.
489 	 */
490 	p2->p_flag = P_INMEM;
491 	p2->p_flag2 = p1->p_flag2 & (P2_NOTRACE | P2_NOTRACE_EXEC);
492 	p2->p_swtick = ticks;
493 	if (p1->p_flag & P_PROFIL)
494 		startprofclock(p2);
495 
496 	/*
497 	 * Whilst the proc lock is held, copy the VM domain data out
498 	 * using the VM domain method.
499 	 */
500 	vm_domain_policy_init(&p2->p_vm_dom_policy);
501 	vm_domain_policy_localcopy(&p2->p_vm_dom_policy,
502 	    &p1->p_vm_dom_policy);
503 
504 	if (flags & RFSIGSHARE) {
505 		p2->p_sigacts = sigacts_hold(p1->p_sigacts);
506 	} else {
507 		sigacts_copy(newsigacts, p1->p_sigacts);
508 		p2->p_sigacts = newsigacts;
509 	}
510 
511 	if (flags & RFTSIGZMB)
512 	        p2->p_sigparent = RFTSIGNUM(flags);
513 	else if (flags & RFLINUXTHPN)
514 	        p2->p_sigparent = SIGUSR1;
515 	else
516 	        p2->p_sigparent = SIGCHLD;
517 
518 	p2->p_textvp = p1->p_textvp;
519 	p2->p_fd = fd;
520 	p2->p_fdtol = fdtol;
521 
522 	if (p1->p_flag2 & P2_INHERIT_PROTECTED) {
523 		p2->p_flag |= P_PROTECTED;
524 		p2->p_flag2 |= P2_INHERIT_PROTECTED;
525 	}
526 
527 	/*
528 	 * p_limit is copy-on-write.  Bump its refcount.
529 	 */
530 	lim_fork(p1, p2);
531 
532 	thread_cow_get_proc(td2, p2);
533 
534 	pstats_fork(p1->p_stats, p2->p_stats);
535 
536 	PROC_UNLOCK(p1);
537 	PROC_UNLOCK(p2);
538 
539 	/* Bump references to the text vnode (for procfs). */
540 	if (p2->p_textvp)
541 		vref(p2->p_textvp);
542 
543 	/*
544 	 * Set up linkage for kernel based threading.
545 	 */
546 	if ((flags & RFTHREAD) != 0) {
547 		mtx_lock(&ppeers_lock);
548 		p2->p_peers = p1->p_peers;
549 		p1->p_peers = p2;
550 		p2->p_leader = p1->p_leader;
551 		mtx_unlock(&ppeers_lock);
552 		PROC_LOCK(p1->p_leader);
553 		if ((p1->p_leader->p_flag & P_WEXIT) != 0) {
554 			PROC_UNLOCK(p1->p_leader);
555 			/*
556 			 * The task leader is exiting, so process p1 is
557 			 * going to be killed shortly.  Since p1 obviously
558 			 * isn't dead yet, we know that the leader is either
559 			 * sending SIGKILL's to all the processes in this
560 			 * task or is sleeping waiting for all the peers to
561 			 * exit.  We let p1 complete the fork, but we need
562 			 * to go ahead and kill the new process p2 since
563 			 * the task leader may not get a chance to send
564 			 * SIGKILL to it.  We leave it on the list so that
565 			 * the task leader will wait for this new process
566 			 * to commit suicide.
567 			 */
568 			PROC_LOCK(p2);
569 			kern_psignal(p2, SIGKILL);
570 			PROC_UNLOCK(p2);
571 		} else
572 			PROC_UNLOCK(p1->p_leader);
573 	} else {
574 		p2->p_peers = NULL;
575 		p2->p_leader = p2;
576 	}
577 
578 	sx_xlock(&proctree_lock);
579 	PGRP_LOCK(p1->p_pgrp);
580 	PROC_LOCK(p2);
581 	PROC_LOCK(p1);
582 
583 	/*
584 	 * Preserve some more flags in subprocess.  P_PROFIL has already
585 	 * been preserved.
586 	 */
587 	p2->p_flag |= p1->p_flag & P_SUGID;
588 	td2->td_pflags |= (td->td_pflags & TDP_ALTSTACK) | TDP_FORKING;
589 	SESS_LOCK(p1->p_session);
590 	if (p1->p_session->s_ttyvp != NULL && p1->p_flag & P_CONTROLT)
591 		p2->p_flag |= P_CONTROLT;
592 	SESS_UNLOCK(p1->p_session);
593 	if (flags & RFPPWAIT)
594 		p2->p_flag |= P_PPWAIT;
595 
596 	p2->p_pgrp = p1->p_pgrp;
597 	LIST_INSERT_AFTER(p1, p2, p_pglist);
598 	PGRP_UNLOCK(p1->p_pgrp);
599 	LIST_INIT(&p2->p_children);
600 	LIST_INIT(&p2->p_orphans);
601 
602 	callout_init_mtx(&p2->p_itcallout, &p2->p_mtx, 0);
603 
604 	/*
605 	 * If PF_FORK is set, the child process inherits the
606 	 * procfs ioctl flags from its parent.
607 	 */
608 	if (p1->p_pfsflags & PF_FORK) {
609 		p2->p_stops = p1->p_stops;
610 		p2->p_pfsflags = p1->p_pfsflags;
611 	}
612 
613 	/*
614 	 * This begins the section where we must prevent the parent
615 	 * from being swapped.
616 	 */
617 	_PHOLD(p1);
618 	PROC_UNLOCK(p1);
619 
620 	/*
621 	 * Attach the new process to its parent.
622 	 *
623 	 * If RFNOWAIT is set, the newly created process becomes a child
624 	 * of init.  This effectively disassociates the child from the
625 	 * parent.
626 	 */
627 	if ((flags & RFNOWAIT) != 0) {
628 		pptr = p1->p_reaper;
629 		p2->p_reaper = pptr;
630 	} else {
631 		p2->p_reaper = (p1->p_treeflag & P_TREE_REAPER) != 0 ?
632 		    p1 : p1->p_reaper;
633 		pptr = p1;
634 	}
635 	p2->p_pptr = pptr;
636 	LIST_INSERT_HEAD(&pptr->p_children, p2, p_sibling);
637 	LIST_INIT(&p2->p_reaplist);
638 	LIST_INSERT_HEAD(&p2->p_reaper->p_reaplist, p2, p_reapsibling);
639 	if (p2->p_reaper == p1)
640 		p2->p_reapsubtree = p2->p_pid;
641 	sx_xunlock(&proctree_lock);
642 
643 	/* Inform accounting that we have forked. */
644 	p2->p_acflag = AFORK;
645 	PROC_UNLOCK(p2);
646 
647 #ifdef KTRACE
648 	ktrprocfork(p1, p2);
649 #endif
650 
651 	/*
652 	 * Finish creating the child process.  It will return via a different
653 	 * execution path later.  (ie: directly into user mode)
654 	 */
655 	vm_forkproc(td, p2, td2, vm2, flags);
656 
657 	if (flags == (RFFDG | RFPROC)) {
658 		PCPU_INC(cnt.v_forks);
659 		PCPU_ADD(cnt.v_forkpages, p2->p_vmspace->vm_dsize +
660 		    p2->p_vmspace->vm_ssize);
661 	} else if (flags == (RFFDG | RFPROC | RFPPWAIT | RFMEM)) {
662 		PCPU_INC(cnt.v_vforks);
663 		PCPU_ADD(cnt.v_vforkpages, p2->p_vmspace->vm_dsize +
664 		    p2->p_vmspace->vm_ssize);
665 	} else if (p1 == &proc0) {
666 		PCPU_INC(cnt.v_kthreads);
667 		PCPU_ADD(cnt.v_kthreadpages, p2->p_vmspace->vm_dsize +
668 		    p2->p_vmspace->vm_ssize);
669 	} else {
670 		PCPU_INC(cnt.v_rforks);
671 		PCPU_ADD(cnt.v_rforkpages, p2->p_vmspace->vm_dsize +
672 		    p2->p_vmspace->vm_ssize);
673 	}
674 
675 	/*
676 	 * Associate the process descriptor with the process before anything
677 	 * can happen that might cause that process to need the descriptor.
678 	 * However, don't do this until after fork(2) can no longer fail.
679 	 */
680 	if (flags & RFPROCDESC)
681 		procdesc_new(p2, pdflags);
682 
683 	/*
684 	 * Both processes are set up, now check if any loadable modules want
685 	 * to adjust anything.
686 	 */
687 	EVENTHANDLER_INVOKE(process_fork, p1, p2, flags);
688 
689 	/*
690 	 * Set the child start time and mark the process as being complete.
691 	 */
692 	PROC_LOCK(p2);
693 	PROC_LOCK(p1);
694 	microuptime(&p2->p_stats->p_start);
695 	PROC_SLOCK(p2);
696 	p2->p_state = PRS_NORMAL;
697 	PROC_SUNLOCK(p2);
698 
699 #ifdef KDTRACE_HOOKS
700 	/*
701 	 * Tell the DTrace fasttrap provider about the new process so that any
702 	 * tracepoints inherited from the parent can be removed. We have to do
703 	 * this only after p_state is PRS_NORMAL since the fasttrap module will
704 	 * use pfind() later on.
705 	 */
706 	if ((flags & RFMEM) == 0 && dtrace_fasttrap_fork)
707 		dtrace_fasttrap_fork(p1, p2);
708 #endif
709 	if ((p1->p_flag & (P_TRACED | P_FOLLOWFORK)) == (P_TRACED |
710 	    P_FOLLOWFORK)) {
711 		/*
712 		 * Arrange for debugger to receive the fork event.
713 		 *
714 		 * We can report PL_FLAG_FORKED regardless of
715 		 * P_FOLLOWFORK settings, but it does not make a sense
716 		 * for runaway child.
717 		 */
718 		td->td_dbgflags |= TDB_FORK;
719 		td->td_dbg_forked = p2->p_pid;
720 		td2->td_dbgflags |= TDB_STOPATFORK;
721 		_PHOLD(p2);
722 		p2_held = 1;
723 	}
724 	if (flags & RFPPWAIT) {
725 		td->td_pflags |= TDP_RFPPWAIT;
726 		td->td_rfppwait_p = p2;
727 	}
728 	PROC_UNLOCK(p2);
729 	if ((flags & RFSTOPPED) == 0) {
730 		/*
731 		 * If RFSTOPPED not requested, make child runnable and
732 		 * add to run queue.
733 		 */
734 		thread_lock(td2);
735 		TD_SET_CAN_RUN(td2);
736 		sched_add(td2, SRQ_BORING);
737 		thread_unlock(td2);
738 	}
739 
740 	/*
741 	 * Now can be swapped.
742 	 */
743 	_PRELE(p1);
744 	PROC_UNLOCK(p1);
745 
746 	/*
747 	 * Tell any interested parties about the new process.
748 	 */
749 	knote_fork(&p1->p_klist, p2->p_pid);
750 	SDT_PROBE3(proc, , , create, p2, p1, flags);
751 
752 	/*
753 	 * Wait until debugger is attached to child.
754 	 */
755 	PROC_LOCK(p2);
756 	while ((td2->td_dbgflags & TDB_STOPATFORK) != 0)
757 		cv_wait(&p2->p_dbgwait, &p2->p_mtx);
758 	if (p2_held)
759 		_PRELE(p2);
760 	PROC_UNLOCK(p2);
761 }
762 
763 int
764 fork1(struct thread *td, int flags, int pages, struct proc **procp,
765     int *procdescp, int pdflags, struct filecaps *fcaps)
766 {
767 	struct proc *p1, *newproc;
768 	struct thread *td2;
769 	struct vmspace *vm2;
770 	struct file *fp_procdesc;
771 	vm_ooffset_t mem_charged;
772 	int error, nprocs_new, ok;
773 	static int curfail;
774 	static struct timeval lastfail;
775 
776 	/* Check for the undefined or unimplemented flags. */
777 	if ((flags & ~(RFFLAGS | RFTSIGFLAGS(RFTSIGMASK))) != 0)
778 		return (EINVAL);
779 
780 	/* Signal value requires RFTSIGZMB. */
781 	if ((flags & RFTSIGFLAGS(RFTSIGMASK)) != 0 && (flags & RFTSIGZMB) == 0)
782 		return (EINVAL);
783 
784 	/* Can't copy and clear. */
785 	if ((flags & (RFFDG|RFCFDG)) == (RFFDG|RFCFDG))
786 		return (EINVAL);
787 
788 	/* Check the validity of the signal number. */
789 	if ((flags & RFTSIGZMB) != 0 && (u_int)RFTSIGNUM(flags) > _SIG_MAXSIG)
790 		return (EINVAL);
791 
792 	if ((flags & RFPROCDESC) != 0) {
793 		/* Can't not create a process yet get a process descriptor. */
794 		if ((flags & RFPROC) == 0)
795 			return (EINVAL);
796 
797 		/* Must provide a place to put a procdesc if creating one. */
798 		if (procdescp == NULL)
799 			return (EINVAL);
800 	}
801 
802 	p1 = td->td_proc;
803 
804 	/*
805 	 * Here we don't create a new process, but we divorce
806 	 * certain parts of a process from itself.
807 	 */
808 	if ((flags & RFPROC) == 0) {
809 		*procp = NULL;
810 		return (fork_norfproc(td, flags));
811 	}
812 
813 	fp_procdesc = NULL;
814 	newproc = NULL;
815 	vm2 = NULL;
816 
817 	/*
818 	 * Increment the nprocs resource before allocations occur.
819 	 * Although process entries are dynamically created, we still
820 	 * keep a global limit on the maximum number we will
821 	 * create. There are hard-limits as to the number of processes
822 	 * that can run, established by the KVA and memory usage for
823 	 * the process data.
824 	 *
825 	 * Don't allow a nonprivileged user to use the last ten
826 	 * processes; don't let root exceed the limit.
827 	 */
828 	nprocs_new = atomic_fetchadd_int(&nprocs, 1) + 1;
829 	if ((nprocs_new >= maxproc - 10 && priv_check_cred(td->td_ucred,
830 	    PRIV_MAXPROC, 0) != 0) || nprocs_new >= maxproc) {
831 		error = EAGAIN;
832 		sx_xlock(&allproc_lock);
833 		if (ppsratecheck(&lastfail, &curfail, 1)) {
834 			printf("maxproc limit exceeded by uid %u (pid %d); "
835 			    "see tuning(7) and login.conf(5)\n",
836 			    td->td_ucred->cr_ruid, p1->p_pid);
837 		}
838 		sx_xunlock(&allproc_lock);
839 		goto fail2;
840 	}
841 
842 	/*
843 	 * If required, create a process descriptor in the parent first; we
844 	 * will abandon it if something goes wrong. We don't finit() until
845 	 * later.
846 	 */
847 	if (flags & RFPROCDESC) {
848 		error = falloc_caps(td, &fp_procdesc, procdescp, 0, fcaps);
849 		if (error != 0)
850 			goto fail2;
851 	}
852 
853 	mem_charged = 0;
854 	if (pages == 0)
855 		pages = kstack_pages;
856 	/* Allocate new proc. */
857 	newproc = uma_zalloc(proc_zone, M_WAITOK);
858 	td2 = FIRST_THREAD_IN_PROC(newproc);
859 	if (td2 == NULL) {
860 		td2 = thread_alloc(pages);
861 		if (td2 == NULL) {
862 			error = ENOMEM;
863 			goto fail2;
864 		}
865 		proc_linkup(newproc, td2);
866 	} else {
867 		if (td2->td_kstack == 0 || td2->td_kstack_pages != pages) {
868 			if (td2->td_kstack != 0)
869 				vm_thread_dispose(td2);
870 			if (!thread_alloc_stack(td2, pages)) {
871 				error = ENOMEM;
872 				goto fail2;
873 			}
874 		}
875 	}
876 
877 	if ((flags & RFMEM) == 0) {
878 		vm2 = vmspace_fork(p1->p_vmspace, &mem_charged);
879 		if (vm2 == NULL) {
880 			error = ENOMEM;
881 			goto fail2;
882 		}
883 		if (!swap_reserve(mem_charged)) {
884 			/*
885 			 * The swap reservation failed. The accounting
886 			 * from the entries of the copied vm2 will be
887 			 * substracted in vmspace_free(), so force the
888 			 * reservation there.
889 			 */
890 			swap_reserve_force(mem_charged);
891 			error = ENOMEM;
892 			goto fail2;
893 		}
894 	} else
895 		vm2 = NULL;
896 
897 	/*
898 	 * XXX: This is ugly; when we copy resource usage, we need to bump
899 	 *      per-cred resource counters.
900 	 */
901 	proc_set_cred_init(newproc, crhold(td->td_ucred));
902 
903 	/*
904 	 * Initialize resource accounting for the child process.
905 	 */
906 	error = racct_proc_fork(p1, newproc);
907 	if (error != 0) {
908 		error = EAGAIN;
909 		goto fail1;
910 	}
911 
912 #ifdef MAC
913 	mac_proc_init(newproc);
914 #endif
915 	knlist_init_mtx(&newproc->p_klist, &newproc->p_mtx);
916 	STAILQ_INIT(&newproc->p_ktr);
917 
918 	/* We have to lock the process tree while we look for a pid. */
919 	sx_slock(&proctree_lock);
920 	sx_xlock(&allproc_lock);
921 
922 	/*
923 	 * Increment the count of procs running with this uid. Don't allow
924 	 * a nonprivileged user to exceed their current limit.
925 	 *
926 	 * XXXRW: Can we avoid privilege here if it's not needed?
927 	 */
928 	error = priv_check_cred(td->td_ucred, PRIV_PROC_LIMIT, 0);
929 	if (error == 0)
930 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1, 0);
931 	else {
932 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1,
933 		    lim_cur(td, RLIMIT_NPROC));
934 	}
935 	if (ok) {
936 		do_fork(td, flags, newproc, td2, vm2, pdflags);
937 
938 		/*
939 		 * Return child proc pointer to parent.
940 		 */
941 		*procp = newproc;
942 		if (flags & RFPROCDESC) {
943 			procdesc_finit(newproc->p_procdesc, fp_procdesc);
944 			fdrop(fp_procdesc, td);
945 		}
946 		racct_proc_fork_done(newproc);
947 		return (0);
948 	}
949 
950 	error = EAGAIN;
951 	sx_sunlock(&proctree_lock);
952 	sx_xunlock(&allproc_lock);
953 #ifdef MAC
954 	mac_proc_destroy(newproc);
955 #endif
956 	racct_proc_exit(newproc);
957 fail1:
958 	crfree(newproc->p_ucred);
959 	newproc->p_ucred = NULL;
960 fail2:
961 	if (vm2 != NULL)
962 		vmspace_free(vm2);
963 	uma_zfree(proc_zone, newproc);
964 	if ((flags & RFPROCDESC) != 0 && fp_procdesc != NULL) {
965 		fdclose(td, fp_procdesc, *procdescp);
966 		fdrop(fp_procdesc, td);
967 	}
968 	atomic_add_int(&nprocs, -1);
969 	pause("fork", hz / 2);
970 	return (error);
971 }
972 
973 /*
974  * Handle the return of a child process from fork1().  This function
975  * is called from the MD fork_trampoline() entry point.
976  */
977 void
978 fork_exit(void (*callout)(void *, struct trapframe *), void *arg,
979     struct trapframe *frame)
980 {
981 	struct proc *p;
982 	struct thread *td;
983 	struct thread *dtd;
984 
985 	td = curthread;
986 	p = td->td_proc;
987 	KASSERT(p->p_state == PRS_NORMAL, ("executing process is still new"));
988 
989 	CTR4(KTR_PROC, "fork_exit: new thread %p (td_sched %p, pid %d, %s)",
990 		td, td->td_sched, p->p_pid, td->td_name);
991 
992 	sched_fork_exit(td);
993 	/*
994 	* Processes normally resume in mi_switch() after being
995 	* cpu_switch()'ed to, but when children start up they arrive here
996 	* instead, so we must do much the same things as mi_switch() would.
997 	*/
998 	if ((dtd = PCPU_GET(deadthread))) {
999 		PCPU_SET(deadthread, NULL);
1000 		thread_stash(dtd);
1001 	}
1002 	thread_unlock(td);
1003 
1004 	/*
1005 	 * cpu_set_fork_handler intercepts this function call to
1006 	 * have this call a non-return function to stay in kernel mode.
1007 	 * initproc has its own fork handler, but it does return.
1008 	 */
1009 	KASSERT(callout != NULL, ("NULL callout in fork_exit"));
1010 	callout(arg, frame);
1011 
1012 	/*
1013 	 * Check if a kernel thread misbehaved and returned from its main
1014 	 * function.
1015 	 */
1016 	if (p->p_flag & P_KTHREAD) {
1017 		printf("Kernel thread \"%s\" (pid %d) exited prematurely.\n",
1018 		    td->td_name, p->p_pid);
1019 		kproc_exit(0);
1020 	}
1021 	mtx_assert(&Giant, MA_NOTOWNED);
1022 
1023 	if (p->p_sysent->sv_schedtail != NULL)
1024 		(p->p_sysent->sv_schedtail)(td);
1025 	td->td_pflags &= ~TDP_FORKING;
1026 }
1027 
1028 /*
1029  * Simplified back end of syscall(), used when returning from fork()
1030  * directly into user mode.  Giant is not held on entry, and must not
1031  * be held on return.  This function is passed in to fork_exit() as the
1032  * first parameter and is called when returning to a new userland process.
1033  */
1034 void
1035 fork_return(struct thread *td, struct trapframe *frame)
1036 {
1037 	struct proc *p, *dbg;
1038 
1039 	p = td->td_proc;
1040 	if (td->td_dbgflags & TDB_STOPATFORK) {
1041 		sx_xlock(&proctree_lock);
1042 		PROC_LOCK(p);
1043 		if ((p->p_pptr->p_flag & (P_TRACED | P_FOLLOWFORK)) ==
1044 		    (P_TRACED | P_FOLLOWFORK)) {
1045 			/*
1046 			 * If debugger still wants auto-attach for the
1047 			 * parent's children, do it now.
1048 			 */
1049 			dbg = p->p_pptr->p_pptr;
1050 			p->p_flag |= P_TRACED;
1051 			p->p_oppid = p->p_pptr->p_pid;
1052 			CTR2(KTR_PTRACE,
1053 		    "fork_return: attaching to new child pid %d: oppid %d",
1054 			    p->p_pid, p->p_oppid);
1055 			proc_reparent(p, dbg);
1056 			sx_xunlock(&proctree_lock);
1057 			td->td_dbgflags |= TDB_CHILD | TDB_SCX;
1058 			ptracestop(td, SIGSTOP);
1059 			td->td_dbgflags &= ~(TDB_CHILD | TDB_SCX);
1060 		} else {
1061 			/*
1062 			 * ... otherwise clear the request.
1063 			 */
1064 			sx_xunlock(&proctree_lock);
1065 			td->td_dbgflags &= ~TDB_STOPATFORK;
1066 			cv_broadcast(&p->p_dbgwait);
1067 		}
1068 		PROC_UNLOCK(p);
1069 	} else if (p->p_flag & P_TRACED) {
1070  		/*
1071 		 * This is the start of a new thread in a traced
1072 		 * process.  Report a system call exit event.
1073 		 */
1074 		PROC_LOCK(p);
1075 		td->td_dbgflags |= TDB_SCX;
1076 		_STOPEVENT(p, S_SCX, td->td_dbg_sc_code);
1077 		if ((p->p_stops & S_PT_SCX) != 0)
1078 			ptracestop(td, SIGTRAP);
1079 		td->td_dbgflags &= ~TDB_SCX;
1080 		PROC_UNLOCK(p);
1081 	}
1082 
1083 	userret(td, frame);
1084 
1085 #ifdef KTRACE
1086 	if (KTRPOINT(td, KTR_SYSRET))
1087 		ktrsysret(SYS_fork, 0, 0);
1088 #endif
1089 }
1090