xref: /freebsd/sys/kern/kern_fork.c (revision 3bdf775801b218aa5a89564839405b122f4b233e)
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/racct.h>
61 #include <sys/resourcevar.h>
62 #include <sys/sched.h>
63 #include <sys/syscall.h>
64 #include <sys/vmmeter.h>
65 #include <sys/vnode.h>
66 #include <sys/acct.h>
67 #include <sys/ktr.h>
68 #include <sys/ktrace.h>
69 #include <sys/unistd.h>
70 #include <sys/sdt.h>
71 #include <sys/sx.h>
72 #include <sys/sysent.h>
73 #include <sys/signalvar.h>
74 
75 #include <security/audit/audit.h>
76 #include <security/mac/mac_framework.h>
77 
78 #include <vm/vm.h>
79 #include <vm/pmap.h>
80 #include <vm/vm_map.h>
81 #include <vm/vm_extern.h>
82 #include <vm/uma.h>
83 
84 #ifdef KDTRACE_HOOKS
85 #include <sys/dtrace_bsd.h>
86 dtrace_fork_func_t	dtrace_fasttrap_fork;
87 #endif
88 
89 SDT_PROVIDER_DECLARE(proc);
90 SDT_PROBE_DEFINE3(proc, kernel, , create, "struct proc *",
91     "struct proc *", "int");
92 
93 #ifndef _SYS_SYSPROTO_H_
94 struct fork_args {
95 	int     dummy;
96 };
97 #endif
98 
99 /* ARGSUSED */
100 int
101 sys_fork(struct thread *td, struct fork_args *uap)
102 {
103 	int error;
104 	struct proc *p2;
105 
106 	error = fork1(td, RFFDG | RFPROC, 0, &p2, NULL, 0);
107 	if (error == 0) {
108 		td->td_retval[0] = p2->p_pid;
109 		td->td_retval[1] = 0;
110 	}
111 	return (error);
112 }
113 
114 /* ARGUSED */
115 int
116 sys_pdfork(td, uap)
117 	struct thread *td;
118 	struct pdfork_args *uap;
119 {
120 	int error, fd;
121 	struct proc *p2;
122 
123 	/*
124 	 * It is necessary to return fd by reference because 0 is a valid file
125 	 * descriptor number, and the child needs to be able to distinguish
126 	 * itself from the parent using the return value.
127 	 */
128 	error = fork1(td, RFFDG | RFPROC | RFPROCDESC, 0, &p2,
129 	    &fd, uap->flags);
130 	if (error == 0) {
131 		td->td_retval[0] = p2->p_pid;
132 		td->td_retval[1] = 0;
133 		error = copyout(&fd, uap->fdp, sizeof(fd));
134 	}
135 	return (error);
136 }
137 
138 /* ARGSUSED */
139 int
140 sys_vfork(struct thread *td, struct vfork_args *uap)
141 {
142 	int error, flags;
143 	struct proc *p2;
144 
145 	flags = RFFDG | RFPROC | RFPPWAIT | RFMEM;
146 	error = fork1(td, flags, 0, &p2, NULL, 0);
147 	if (error == 0) {
148 		td->td_retval[0] = p2->p_pid;
149 		td->td_retval[1] = 0;
150 	}
151 	return (error);
152 }
153 
154 int
155 sys_rfork(struct thread *td, struct rfork_args *uap)
156 {
157 	struct proc *p2;
158 	int error;
159 
160 	/* Don't allow kernel-only flags. */
161 	if ((uap->flags & RFKERNELONLY) != 0)
162 		return (EINVAL);
163 
164 	AUDIT_ARG_FFLAGS(uap->flags);
165 	error = fork1(td, uap->flags, 0, &p2, NULL, 0);
166 	if (error == 0) {
167 		td->td_retval[0] = p2 ? p2->p_pid : 0;
168 		td->td_retval[1] = 0;
169 	}
170 	return (error);
171 }
172 
173 int	nprocs = 1;		/* process 0 */
174 int	lastpid = 0;
175 SYSCTL_INT(_kern, OID_AUTO, lastpid, CTLFLAG_RD, &lastpid, 0,
176     "Last used PID");
177 
178 /*
179  * Random component to lastpid generation.  We mix in a random factor to make
180  * it a little harder to predict.  We sanity check the modulus value to avoid
181  * doing it in critical paths.  Don't let it be too small or we pointlessly
182  * waste randomness entropy, and don't let it be impossibly large.  Using a
183  * modulus that is too big causes a LOT more process table scans and slows
184  * down fork processing as the pidchecked caching is defeated.
185  */
186 static int randompid = 0;
187 
188 static int
189 sysctl_kern_randompid(SYSCTL_HANDLER_ARGS)
190 {
191 	int error, pid;
192 
193 	error = sysctl_wire_old_buffer(req, sizeof(int));
194 	if (error != 0)
195 		return(error);
196 	sx_xlock(&allproc_lock);
197 	pid = randompid;
198 	error = sysctl_handle_int(oidp, &pid, 0, req);
199 	if (error == 0 && req->newptr != NULL) {
200 		if (pid < 0 || pid > pid_max - 100)	/* out of range */
201 			pid = pid_max - 100;
202 		else if (pid < 2)			/* NOP */
203 			pid = 0;
204 		else if (pid < 100)			/* Make it reasonable */
205 			pid = 100;
206 		randompid = pid;
207 	}
208 	sx_xunlock(&allproc_lock);
209 	return (error);
210 }
211 
212 SYSCTL_PROC(_kern, OID_AUTO, randompid, CTLTYPE_INT|CTLFLAG_RW,
213     0, 0, sysctl_kern_randompid, "I", "Random PID modulus");
214 
215 static int
216 fork_findpid(int flags)
217 {
218 	struct proc *p;
219 	int trypid;
220 	static int pidchecked = 0;
221 
222 	/*
223 	 * Requires allproc_lock in order to iterate over the list
224 	 * of processes, and proctree_lock to access p_pgrp.
225 	 */
226 	sx_assert(&allproc_lock, SX_LOCKED);
227 	sx_assert(&proctree_lock, SX_LOCKED);
228 
229 	/*
230 	 * Find an unused process ID.  We remember a range of unused IDs
231 	 * ready to use (from lastpid+1 through pidchecked-1).
232 	 *
233 	 * If RFHIGHPID is set (used during system boot), do not allocate
234 	 * low-numbered pids.
235 	 */
236 	trypid = lastpid + 1;
237 	if (flags & RFHIGHPID) {
238 		if (trypid < 10)
239 			trypid = 10;
240 	} else {
241 		if (randompid)
242 			trypid += arc4random() % randompid;
243 	}
244 retry:
245 	/*
246 	 * If the process ID prototype has wrapped around,
247 	 * restart somewhat above 0, as the low-numbered procs
248 	 * tend to include daemons that don't exit.
249 	 */
250 	if (trypid >= pid_max) {
251 		trypid = trypid % pid_max;
252 		if (trypid < 100)
253 			trypid += 100;
254 		pidchecked = 0;
255 	}
256 	if (trypid >= pidchecked) {
257 		int doingzomb = 0;
258 
259 		pidchecked = PID_MAX;
260 		/*
261 		 * Scan the active and zombie procs to check whether this pid
262 		 * is in use.  Remember the lowest pid that's greater
263 		 * than trypid, so we can avoid checking for a while.
264 		 */
265 		p = LIST_FIRST(&allproc);
266 again:
267 		for (; p != NULL; p = LIST_NEXT(p, p_list)) {
268 			while (p->p_pid == trypid ||
269 			    (p->p_pgrp != NULL &&
270 			    (p->p_pgrp->pg_id == trypid ||
271 			    (p->p_session != NULL &&
272 			    p->p_session->s_sid == trypid)))) {
273 				trypid++;
274 				if (trypid >= pidchecked)
275 					goto retry;
276 			}
277 			if (p->p_pid > trypid && pidchecked > p->p_pid)
278 				pidchecked = p->p_pid;
279 			if (p->p_pgrp != NULL) {
280 				if (p->p_pgrp->pg_id > trypid &&
281 				    pidchecked > p->p_pgrp->pg_id)
282 					pidchecked = p->p_pgrp->pg_id;
283 				if (p->p_session != NULL &&
284 				    p->p_session->s_sid > trypid &&
285 				    pidchecked > p->p_session->s_sid)
286 					pidchecked = p->p_session->s_sid;
287 			}
288 		}
289 		if (!doingzomb) {
290 			doingzomb = 1;
291 			p = LIST_FIRST(&zombproc);
292 			goto again;
293 		}
294 	}
295 
296 	/*
297 	 * RFHIGHPID does not mess with the lastpid counter during boot.
298 	 */
299 	if (flags & RFHIGHPID)
300 		pidchecked = 0;
301 	else
302 		lastpid = trypid;
303 
304 	return (trypid);
305 }
306 
307 static int
308 fork_norfproc(struct thread *td, int flags)
309 {
310 	int error;
311 	struct proc *p1;
312 
313 	KASSERT((flags & RFPROC) == 0,
314 	    ("fork_norfproc called with RFPROC set"));
315 	p1 = td->td_proc;
316 
317 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
318 	    (flags & (RFCFDG | RFFDG))) {
319 		PROC_LOCK(p1);
320 		if (thread_single(SINGLE_BOUNDARY)) {
321 			PROC_UNLOCK(p1);
322 			return (ERESTART);
323 		}
324 		PROC_UNLOCK(p1);
325 	}
326 
327 	error = vm_forkproc(td, NULL, NULL, NULL, flags);
328 	if (error)
329 		goto fail;
330 
331 	/*
332 	 * Close all file descriptors.
333 	 */
334 	if (flags & RFCFDG) {
335 		struct filedesc *fdtmp;
336 		fdtmp = fdinit(td->td_proc->p_fd);
337 		fdescfree(td);
338 		p1->p_fd = fdtmp;
339 	}
340 
341 	/*
342 	 * Unshare file descriptors (from parent).
343 	 */
344 	if (flags & RFFDG)
345 		fdunshare(p1, td);
346 
347 fail:
348 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
349 	    (flags & (RFCFDG | RFFDG))) {
350 		PROC_LOCK(p1);
351 		thread_single_end();
352 		PROC_UNLOCK(p1);
353 	}
354 	return (error);
355 }
356 
357 static void
358 do_fork(struct thread *td, int flags, struct proc *p2, struct thread *td2,
359     struct vmspace *vm2, int pdflags)
360 {
361 	struct proc *p1, *pptr;
362 	int p2_held, trypid;
363 	struct filedesc *fd;
364 	struct filedesc_to_leader *fdtol;
365 	struct sigacts *newsigacts;
366 
367 	sx_assert(&proctree_lock, SX_SLOCKED);
368 	sx_assert(&allproc_lock, SX_XLOCKED);
369 
370 	p2_held = 0;
371 	p1 = td->td_proc;
372 
373 	/*
374 	 * Increment the nprocs resource before blocking can occur.  There
375 	 * are hard-limits as to the number of processes that can run.
376 	 */
377 	nprocs++;
378 
379 	trypid = fork_findpid(flags);
380 
381 	sx_sunlock(&proctree_lock);
382 
383 	p2->p_state = PRS_NEW;		/* protect against others */
384 	p2->p_pid = trypid;
385 	AUDIT_ARG_PID(p2->p_pid);
386 	LIST_INSERT_HEAD(&allproc, p2, p_list);
387 	LIST_INSERT_HEAD(PIDHASH(p2->p_pid), p2, p_hash);
388 	tidhash_add(td2);
389 	PROC_LOCK(p2);
390 	PROC_LOCK(p1);
391 
392 	sx_xunlock(&allproc_lock);
393 
394 	bcopy(&p1->p_startcopy, &p2->p_startcopy,
395 	    __rangeof(struct proc, p_startcopy, p_endcopy));
396 	pargs_hold(p2->p_args);
397 	PROC_UNLOCK(p1);
398 
399 	bzero(&p2->p_startzero,
400 	    __rangeof(struct proc, p_startzero, p_endzero));
401 
402 	p2->p_ucred = crhold(td->td_ucred);
403 
404 	/* Tell the prison that we exist. */
405 	prison_proc_hold(p2->p_ucred->cr_prison);
406 
407 	PROC_UNLOCK(p2);
408 
409 	/*
410 	 * Malloc things while we don't hold any locks.
411 	 */
412 	if (flags & RFSIGSHARE)
413 		newsigacts = NULL;
414 	else
415 		newsigacts = sigacts_alloc();
416 
417 	/*
418 	 * Copy filedesc.
419 	 */
420 	if (flags & RFCFDG) {
421 		fd = fdinit(p1->p_fd);
422 		fdtol = NULL;
423 	} else if (flags & RFFDG) {
424 		fd = fdcopy(p1->p_fd);
425 		fdtol = NULL;
426 	} else {
427 		fd = fdshare(p1->p_fd);
428 		if (p1->p_fdtol == NULL)
429 			p1->p_fdtol = filedesc_to_leader_alloc(NULL, NULL,
430 			    p1->p_leader);
431 		if ((flags & RFTHREAD) != 0) {
432 			/*
433 			 * Shared file descriptor table, and shared
434 			 * process leaders.
435 			 */
436 			fdtol = p1->p_fdtol;
437 			FILEDESC_XLOCK(p1->p_fd);
438 			fdtol->fdl_refcount++;
439 			FILEDESC_XUNLOCK(p1->p_fd);
440 		} else {
441 			/*
442 			 * Shared file descriptor table, and different
443 			 * process leaders.
444 			 */
445 			fdtol = filedesc_to_leader_alloc(p1->p_fdtol,
446 			    p1->p_fd, p2);
447 		}
448 	}
449 	/*
450 	 * Make a proc table entry for the new process.
451 	 * Start by zeroing the section of proc that is zero-initialized,
452 	 * then copy the section that is copied directly from the parent.
453 	 */
454 
455 	PROC_LOCK(p2);
456 	PROC_LOCK(p1);
457 
458 	bzero(&td2->td_startzero,
459 	    __rangeof(struct thread, td_startzero, td_endzero));
460 
461 	bcopy(&td->td_startcopy, &td2->td_startcopy,
462 	    __rangeof(struct thread, td_startcopy, td_endcopy));
463 
464 	bcopy(&p2->p_comm, &td2->td_name, sizeof(td2->td_name));
465 	td2->td_sigstk = td->td_sigstk;
466 	td2->td_flags = TDF_INMEM;
467 	td2->td_lend_user_pri = PRI_MAX;
468 
469 #ifdef VIMAGE
470 	td2->td_vnet = NULL;
471 	td2->td_vnet_lpush = NULL;
472 #endif
473 
474 	/*
475 	 * Allow the scheduler to initialize the child.
476 	 */
477 	thread_lock(td);
478 	sched_fork(td, td2);
479 	thread_unlock(td);
480 
481 	/*
482 	 * Duplicate sub-structures as needed.
483 	 * Increase reference counts on shared objects.
484 	 */
485 	p2->p_flag = P_INMEM;
486 	p2->p_flag2 = 0;
487 	p2->p_swtick = ticks;
488 	if (p1->p_flag & P_PROFIL)
489 		startprofclock(p2);
490 	td2->td_ucred = crhold(p2->p_ucred);
491 
492 	if (flags & RFSIGSHARE) {
493 		p2->p_sigacts = sigacts_hold(p1->p_sigacts);
494 	} else {
495 		sigacts_copy(newsigacts, p1->p_sigacts);
496 		p2->p_sigacts = newsigacts;
497 	}
498 
499 	if (flags & RFTSIGZMB)
500 	        p2->p_sigparent = RFTSIGNUM(flags);
501 	else if (flags & RFLINUXTHPN)
502 	        p2->p_sigparent = SIGUSR1;
503 	else
504 	        p2->p_sigparent = SIGCHLD;
505 
506 	p2->p_textvp = p1->p_textvp;
507 	p2->p_fd = fd;
508 	p2->p_fdtol = fdtol;
509 
510 	if (p1->p_flag2 & P2_INHERIT_PROTECTED) {
511 		p2->p_flag |= P_PROTECTED;
512 		p2->p_flag2 |= P2_INHERIT_PROTECTED;
513 	}
514 
515 	/*
516 	 * p_limit is copy-on-write.  Bump its refcount.
517 	 */
518 	lim_fork(p1, p2);
519 
520 	pstats_fork(p1->p_stats, p2->p_stats);
521 
522 	PROC_UNLOCK(p1);
523 	PROC_UNLOCK(p2);
524 
525 	/* Bump references to the text vnode (for procfs). */
526 	if (p2->p_textvp)
527 		vref(p2->p_textvp);
528 
529 	/*
530 	 * Set up linkage for kernel based threading.
531 	 */
532 	if ((flags & RFTHREAD) != 0) {
533 		mtx_lock(&ppeers_lock);
534 		p2->p_peers = p1->p_peers;
535 		p1->p_peers = p2;
536 		p2->p_leader = p1->p_leader;
537 		mtx_unlock(&ppeers_lock);
538 		PROC_LOCK(p1->p_leader);
539 		if ((p1->p_leader->p_flag & P_WEXIT) != 0) {
540 			PROC_UNLOCK(p1->p_leader);
541 			/*
542 			 * The task leader is exiting, so process p1 is
543 			 * going to be killed shortly.  Since p1 obviously
544 			 * isn't dead yet, we know that the leader is either
545 			 * sending SIGKILL's to all the processes in this
546 			 * task or is sleeping waiting for all the peers to
547 			 * exit.  We let p1 complete the fork, but we need
548 			 * to go ahead and kill the new process p2 since
549 			 * the task leader may not get a chance to send
550 			 * SIGKILL to it.  We leave it on the list so that
551 			 * the task leader will wait for this new process
552 			 * to commit suicide.
553 			 */
554 			PROC_LOCK(p2);
555 			kern_psignal(p2, SIGKILL);
556 			PROC_UNLOCK(p2);
557 		} else
558 			PROC_UNLOCK(p1->p_leader);
559 	} else {
560 		p2->p_peers = NULL;
561 		p2->p_leader = p2;
562 	}
563 
564 	sx_xlock(&proctree_lock);
565 	PGRP_LOCK(p1->p_pgrp);
566 	PROC_LOCK(p2);
567 	PROC_LOCK(p1);
568 
569 	/*
570 	 * Preserve some more flags in subprocess.  P_PROFIL has already
571 	 * been preserved.
572 	 */
573 	p2->p_flag |= p1->p_flag & P_SUGID;
574 	td2->td_pflags |= td->td_pflags & TDP_ALTSTACK;
575 	SESS_LOCK(p1->p_session);
576 	if (p1->p_session->s_ttyvp != NULL && p1->p_flag & P_CONTROLT)
577 		p2->p_flag |= P_CONTROLT;
578 	SESS_UNLOCK(p1->p_session);
579 	if (flags & RFPPWAIT)
580 		p2->p_flag |= P_PPWAIT;
581 
582 	p2->p_pgrp = p1->p_pgrp;
583 	LIST_INSERT_AFTER(p1, p2, p_pglist);
584 	PGRP_UNLOCK(p1->p_pgrp);
585 	LIST_INIT(&p2->p_children);
586 	LIST_INIT(&p2->p_orphans);
587 
588 	callout_init_mtx(&p2->p_itcallout, &p2->p_mtx, 0);
589 
590 	/*
591 	 * If PF_FORK is set, the child process inherits the
592 	 * procfs ioctl flags from its parent.
593 	 */
594 	if (p1->p_pfsflags & PF_FORK) {
595 		p2->p_stops = p1->p_stops;
596 		p2->p_pfsflags = p1->p_pfsflags;
597 	}
598 
599 	/*
600 	 * This begins the section where we must prevent the parent
601 	 * from being swapped.
602 	 */
603 	_PHOLD(p1);
604 	PROC_UNLOCK(p1);
605 
606 	/*
607 	 * Attach the new process to its parent.
608 	 *
609 	 * If RFNOWAIT is set, the newly created process becomes a child
610 	 * of init.  This effectively disassociates the child from the
611 	 * parent.
612 	 */
613 	if (flags & RFNOWAIT)
614 		pptr = initproc;
615 	else
616 		pptr = p1;
617 	p2->p_pptr = pptr;
618 	LIST_INSERT_HEAD(&pptr->p_children, p2, p_sibling);
619 	sx_xunlock(&proctree_lock);
620 
621 	/* Inform accounting that we have forked. */
622 	p2->p_acflag = AFORK;
623 	PROC_UNLOCK(p2);
624 
625 #ifdef KTRACE
626 	ktrprocfork(p1, p2);
627 #endif
628 
629 	/*
630 	 * Finish creating the child process.  It will return via a different
631 	 * execution path later.  (ie: directly into user mode)
632 	 */
633 	vm_forkproc(td, p2, td2, vm2, flags);
634 
635 	if (flags == (RFFDG | RFPROC)) {
636 		PCPU_INC(cnt.v_forks);
637 		PCPU_ADD(cnt.v_forkpages, p2->p_vmspace->vm_dsize +
638 		    p2->p_vmspace->vm_ssize);
639 	} else if (flags == (RFFDG | RFPROC | RFPPWAIT | RFMEM)) {
640 		PCPU_INC(cnt.v_vforks);
641 		PCPU_ADD(cnt.v_vforkpages, p2->p_vmspace->vm_dsize +
642 		    p2->p_vmspace->vm_ssize);
643 	} else if (p1 == &proc0) {
644 		PCPU_INC(cnt.v_kthreads);
645 		PCPU_ADD(cnt.v_kthreadpages, p2->p_vmspace->vm_dsize +
646 		    p2->p_vmspace->vm_ssize);
647 	} else {
648 		PCPU_INC(cnt.v_rforks);
649 		PCPU_ADD(cnt.v_rforkpages, p2->p_vmspace->vm_dsize +
650 		    p2->p_vmspace->vm_ssize);
651 	}
652 
653 	/*
654 	 * Associate the process descriptor with the process before anything
655 	 * can happen that might cause that process to need the descriptor.
656 	 * However, don't do this until after fork(2) can no longer fail.
657 	 */
658 	if (flags & RFPROCDESC)
659 		procdesc_new(p2, pdflags);
660 
661 	/*
662 	 * Both processes are set up, now check if any loadable modules want
663 	 * to adjust anything.
664 	 */
665 	EVENTHANDLER_INVOKE(process_fork, p1, p2, flags);
666 
667 	/*
668 	 * Set the child start time and mark the process as being complete.
669 	 */
670 	PROC_LOCK(p2);
671 	PROC_LOCK(p1);
672 	microuptime(&p2->p_stats->p_start);
673 	PROC_SLOCK(p2);
674 	p2->p_state = PRS_NORMAL;
675 	PROC_SUNLOCK(p2);
676 
677 #ifdef KDTRACE_HOOKS
678 	/*
679 	 * Tell the DTrace fasttrap provider about the new process so that any
680 	 * tracepoints inherited from the parent can be removed. We have to do
681 	 * this only after p_state is PRS_NORMAL since the fasttrap module will
682 	 * use pfind() later on.
683 	 */
684 	if ((flags & RFMEM) == 0 && dtrace_fasttrap_fork)
685 		dtrace_fasttrap_fork(p1, p2);
686 #endif
687 	if ((p1->p_flag & (P_TRACED | P_FOLLOWFORK)) == (P_TRACED |
688 	    P_FOLLOWFORK)) {
689 		/*
690 		 * Arrange for debugger to receive the fork event.
691 		 *
692 		 * We can report PL_FLAG_FORKED regardless of
693 		 * P_FOLLOWFORK settings, but it does not make a sense
694 		 * for runaway child.
695 		 */
696 		td->td_dbgflags |= TDB_FORK;
697 		td->td_dbg_forked = p2->p_pid;
698 		td2->td_dbgflags |= TDB_STOPATFORK;
699 		_PHOLD(p2);
700 		p2_held = 1;
701 	}
702 	if (flags & RFPPWAIT) {
703 		td->td_pflags |= TDP_RFPPWAIT;
704 		td->td_rfppwait_p = p2;
705 	}
706 	PROC_UNLOCK(p2);
707 	if ((flags & RFSTOPPED) == 0) {
708 		/*
709 		 * If RFSTOPPED not requested, make child runnable and
710 		 * add to run queue.
711 		 */
712 		thread_lock(td2);
713 		TD_SET_CAN_RUN(td2);
714 		sched_add(td2, SRQ_BORING);
715 		thread_unlock(td2);
716 	}
717 
718 	/*
719 	 * Now can be swapped.
720 	 */
721 	_PRELE(p1);
722 	PROC_UNLOCK(p1);
723 
724 	/*
725 	 * Tell any interested parties about the new process.
726 	 */
727 	knote_fork(&p1->p_klist, p2->p_pid);
728 	SDT_PROBE(proc, kernel, , create, p2, p1, flags, 0, 0);
729 
730 	/*
731 	 * Wait until debugger is attached to child.
732 	 */
733 	PROC_LOCK(p2);
734 	while ((td2->td_dbgflags & TDB_STOPATFORK) != 0)
735 		cv_wait(&p2->p_dbgwait, &p2->p_mtx);
736 	if (p2_held)
737 		_PRELE(p2);
738 	PROC_UNLOCK(p2);
739 }
740 
741 int
742 fork1(struct thread *td, int flags, int pages, struct proc **procp,
743     int *procdescp, int pdflags)
744 {
745 	struct proc *p1;
746 	struct proc *newproc;
747 	int ok;
748 	struct thread *td2;
749 	struct vmspace *vm2;
750 	vm_ooffset_t mem_charged;
751 	int error;
752 	static int curfail;
753 	static struct timeval lastfail;
754 	struct file *fp_procdesc = NULL;
755 
756 	/* Check for the undefined or unimplemented flags. */
757 	if ((flags & ~(RFFLAGS | RFTSIGFLAGS(RFTSIGMASK))) != 0)
758 		return (EINVAL);
759 
760 	/* Signal value requires RFTSIGZMB. */
761 	if ((flags & RFTSIGFLAGS(RFTSIGMASK)) != 0 && (flags & RFTSIGZMB) == 0)
762 		return (EINVAL);
763 
764 	/* Can't copy and clear. */
765 	if ((flags & (RFFDG|RFCFDG)) == (RFFDG|RFCFDG))
766 		return (EINVAL);
767 
768 	/* Check the validity of the signal number. */
769 	if ((flags & RFTSIGZMB) != 0 && (u_int)RFTSIGNUM(flags) > _SIG_MAXSIG)
770 		return (EINVAL);
771 
772 	if ((flags & RFPROCDESC) != 0) {
773 		/* Can't not create a process yet get a process descriptor. */
774 		if ((flags & RFPROC) == 0)
775 			return (EINVAL);
776 
777 		/* Must provide a place to put a procdesc if creating one. */
778 		if (procdescp == NULL)
779 			return (EINVAL);
780 	}
781 
782 	p1 = td->td_proc;
783 
784 	/*
785 	 * Here we don't create a new process, but we divorce
786 	 * certain parts of a process from itself.
787 	 */
788 	if ((flags & RFPROC) == 0) {
789 		*procp = NULL;
790 		return (fork_norfproc(td, flags));
791 	}
792 
793 	/*
794 	 * If required, create a process descriptor in the parent first; we
795 	 * will abandon it if something goes wrong. We don't finit() until
796 	 * later.
797 	 */
798 	if (flags & RFPROCDESC) {
799 		error = falloc(td, &fp_procdesc, procdescp, 0);
800 		if (error != 0)
801 			return (error);
802 	}
803 
804 	mem_charged = 0;
805 	vm2 = NULL;
806 	if (pages == 0)
807 		pages = KSTACK_PAGES;
808 	/* Allocate new proc. */
809 	newproc = uma_zalloc(proc_zone, M_WAITOK);
810 	td2 = FIRST_THREAD_IN_PROC(newproc);
811 	if (td2 == NULL) {
812 		td2 = thread_alloc(pages);
813 		if (td2 == NULL) {
814 			error = ENOMEM;
815 			goto fail1;
816 		}
817 		proc_linkup(newproc, td2);
818 	} else {
819 		if (td2->td_kstack == 0 || td2->td_kstack_pages != pages) {
820 			if (td2->td_kstack != 0)
821 				vm_thread_dispose(td2);
822 			if (!thread_alloc_stack(td2, pages)) {
823 				error = ENOMEM;
824 				goto fail1;
825 			}
826 		}
827 	}
828 
829 	if ((flags & RFMEM) == 0) {
830 		vm2 = vmspace_fork(p1->p_vmspace, &mem_charged);
831 		if (vm2 == NULL) {
832 			error = ENOMEM;
833 			goto fail1;
834 		}
835 		if (!swap_reserve(mem_charged)) {
836 			/*
837 			 * The swap reservation failed. The accounting
838 			 * from the entries of the copied vm2 will be
839 			 * substracted in vmspace_free(), so force the
840 			 * reservation there.
841 			 */
842 			swap_reserve_force(mem_charged);
843 			error = ENOMEM;
844 			goto fail1;
845 		}
846 	} else
847 		vm2 = NULL;
848 
849 	/*
850 	 * XXX: This is ugly; when we copy resource usage, we need to bump
851 	 *      per-cred resource counters.
852 	 */
853 	newproc->p_ucred = p1->p_ucred;
854 
855 	/*
856 	 * Initialize resource accounting for the child process.
857 	 */
858 	error = racct_proc_fork(p1, newproc);
859 	if (error != 0) {
860 		error = EAGAIN;
861 		goto fail1;
862 	}
863 
864 #ifdef MAC
865 	mac_proc_init(newproc);
866 #endif
867 	knlist_init_mtx(&newproc->p_klist, &newproc->p_mtx);
868 	STAILQ_INIT(&newproc->p_ktr);
869 
870 	/* We have to lock the process tree while we look for a pid. */
871 	sx_slock(&proctree_lock);
872 
873 	/*
874 	 * Although process entries are dynamically created, we still keep
875 	 * a global limit on the maximum number we will create.  Don't allow
876 	 * a nonprivileged user to use the last ten processes; don't let root
877 	 * exceed the limit. The variable nprocs is the current number of
878 	 * processes, maxproc is the limit.
879 	 */
880 	sx_xlock(&allproc_lock);
881 	if ((nprocs >= maxproc - 10 && priv_check_cred(td->td_ucred,
882 	    PRIV_MAXPROC, 0) != 0) || nprocs >= maxproc) {
883 		error = EAGAIN;
884 		goto fail;
885 	}
886 
887 	/*
888 	 * Increment the count of procs running with this uid. Don't allow
889 	 * a nonprivileged user to exceed their current limit.
890 	 *
891 	 * XXXRW: Can we avoid privilege here if it's not needed?
892 	 */
893 	error = priv_check_cred(td->td_ucred, PRIV_PROC_LIMIT, 0);
894 	if (error == 0)
895 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1, 0);
896 	else {
897 		PROC_LOCK(p1);
898 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1,
899 		    lim_cur(p1, RLIMIT_NPROC));
900 		PROC_UNLOCK(p1);
901 	}
902 	if (ok) {
903 		do_fork(td, flags, newproc, td2, vm2, pdflags);
904 
905 		/*
906 		 * Return child proc pointer to parent.
907 		 */
908 		*procp = newproc;
909 		if (flags & RFPROCDESC) {
910 			procdesc_finit(newproc->p_procdesc, fp_procdesc);
911 			fdrop(fp_procdesc, td);
912 		}
913 		racct_proc_fork_done(newproc);
914 		return (0);
915 	}
916 
917 	error = EAGAIN;
918 fail:
919 	sx_sunlock(&proctree_lock);
920 	if (ppsratecheck(&lastfail, &curfail, 1))
921 		printf("maxproc limit exceeded by uid %u (pid %d); see tuning(7) and login.conf(5)\n",
922 		    td->td_ucred->cr_ruid, p1->p_pid);
923 	sx_xunlock(&allproc_lock);
924 #ifdef MAC
925 	mac_proc_destroy(newproc);
926 #endif
927 	racct_proc_exit(newproc);
928 fail1:
929 	if (vm2 != NULL)
930 		vmspace_free(vm2);
931 	uma_zfree(proc_zone, newproc);
932 	if ((flags & RFPROCDESC) != 0 && fp_procdesc != NULL) {
933 		fdclose(td->td_proc->p_fd, fp_procdesc, *procdescp, td);
934 		fdrop(fp_procdesc, td);
935 	}
936 	pause("fork", hz / 2);
937 	return (error);
938 }
939 
940 /*
941  * Handle the return of a child process from fork1().  This function
942  * is called from the MD fork_trampoline() entry point.
943  */
944 void
945 fork_exit(void (*callout)(void *, struct trapframe *), void *arg,
946     struct trapframe *frame)
947 {
948 	struct proc *p;
949 	struct thread *td;
950 	struct thread *dtd;
951 
952 	td = curthread;
953 	p = td->td_proc;
954 	KASSERT(p->p_state == PRS_NORMAL, ("executing process is still new"));
955 
956 	CTR4(KTR_PROC, "fork_exit: new thread %p (td_sched %p, pid %d, %s)",
957 		td, td->td_sched, p->p_pid, td->td_name);
958 
959 	sched_fork_exit(td);
960 	/*
961 	* Processes normally resume in mi_switch() after being
962 	* cpu_switch()'ed to, but when children start up they arrive here
963 	* instead, so we must do much the same things as mi_switch() would.
964 	*/
965 	if ((dtd = PCPU_GET(deadthread))) {
966 		PCPU_SET(deadthread, NULL);
967 		thread_stash(dtd);
968 	}
969 	thread_unlock(td);
970 
971 	/*
972 	 * cpu_set_fork_handler intercepts this function call to
973 	 * have this call a non-return function to stay in kernel mode.
974 	 * initproc has its own fork handler, but it does return.
975 	 */
976 	KASSERT(callout != NULL, ("NULL callout in fork_exit"));
977 	callout(arg, frame);
978 
979 	/*
980 	 * Check if a kernel thread misbehaved and returned from its main
981 	 * function.
982 	 */
983 	if (p->p_flag & P_KTHREAD) {
984 		printf("Kernel thread \"%s\" (pid %d) exited prematurely.\n",
985 		    td->td_name, p->p_pid);
986 		kproc_exit(0);
987 	}
988 	mtx_assert(&Giant, MA_NOTOWNED);
989 
990 	if (p->p_sysent->sv_schedtail != NULL)
991 		(p->p_sysent->sv_schedtail)(td);
992 }
993 
994 /*
995  * Simplified back end of syscall(), used when returning from fork()
996  * directly into user mode.  Giant is not held on entry, and must not
997  * be held on return.  This function is passed in to fork_exit() as the
998  * first parameter and is called when returning to a new userland process.
999  */
1000 void
1001 fork_return(struct thread *td, struct trapframe *frame)
1002 {
1003 	struct proc *p, *dbg;
1004 
1005 	if (td->td_dbgflags & TDB_STOPATFORK) {
1006 		p = td->td_proc;
1007 		sx_xlock(&proctree_lock);
1008 		PROC_LOCK(p);
1009 		if ((p->p_pptr->p_flag & (P_TRACED | P_FOLLOWFORK)) ==
1010 		    (P_TRACED | P_FOLLOWFORK)) {
1011 			/*
1012 			 * If debugger still wants auto-attach for the
1013 			 * parent's children, do it now.
1014 			 */
1015 			dbg = p->p_pptr->p_pptr;
1016 			p->p_flag |= P_TRACED;
1017 			p->p_oppid = p->p_pptr->p_pid;
1018 			proc_reparent(p, dbg);
1019 			sx_xunlock(&proctree_lock);
1020 			td->td_dbgflags |= TDB_CHILD;
1021 			ptracestop(td, SIGSTOP);
1022 			td->td_dbgflags &= ~TDB_CHILD;
1023 		} else {
1024 			/*
1025 			 * ... otherwise clear the request.
1026 			 */
1027 			sx_xunlock(&proctree_lock);
1028 			td->td_dbgflags &= ~TDB_STOPATFORK;
1029 			cv_broadcast(&p->p_dbgwait);
1030 		}
1031 		PROC_UNLOCK(p);
1032 	}
1033 
1034 	userret(td, frame);
1035 
1036 #ifdef KTRACE
1037 	if (KTRPOINT(td, KTR_SYSRET))
1038 		ktrsysret(SYS_fork, 0, 0);
1039 #endif
1040 }
1041