xref: /freebsd/sys/kern/kern_fork.c (revision cc16dea626cf2fc80cde667ac4798065108e596c)
1 /*-
2  * Copyright (c) 1982, 1986, 1989, 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)kern_fork.c	8.6 (Berkeley) 4/8/94
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_kdtrace.h"
41 #include "opt_ktrace.h"
42 #include "opt_kstack_pages.h"
43 #include "opt_procdesc.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/sysproto.h>
48 #include <sys/eventhandler.h>
49 #include <sys/fcntl.h>
50 #include <sys/filedesc.h>
51 #include <sys/jail.h>
52 #include <sys/kernel.h>
53 #include <sys/kthread.h>
54 #include <sys/sysctl.h>
55 #include <sys/lock.h>
56 #include <sys/malloc.h>
57 #include <sys/mutex.h>
58 #include <sys/priv.h>
59 #include <sys/proc.h>
60 #include <sys/procdesc.h>
61 #include <sys/pioctl.h>
62 #include <sys/racct.h>
63 #include <sys/resourcevar.h>
64 #include <sys/sched.h>
65 #include <sys/syscall.h>
66 #include <sys/vmmeter.h>
67 #include <sys/vnode.h>
68 #include <sys/acct.h>
69 #include <sys/ktr.h>
70 #include <sys/ktrace.h>
71 #include <sys/unistd.h>
72 #include <sys/sdt.h>
73 #include <sys/sx.h>
74 #include <sys/sysent.h>
75 #include <sys/signalvar.h>
76 
77 #include <security/audit/audit.h>
78 #include <security/mac/mac_framework.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_extern.h>
84 #include <vm/uma.h>
85 
86 #ifdef KDTRACE_HOOKS
87 #include <sys/dtrace_bsd.h>
88 dtrace_fork_func_t	dtrace_fasttrap_fork;
89 #endif
90 
91 SDT_PROVIDER_DECLARE(proc);
92 SDT_PROBE_DEFINE3(proc, kernel, , create, 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);
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 #ifdef PROCDESC
123 	int error, fd;
124 	struct proc *p2;
125 
126 	/*
127 	 * It is necessary to return fd by reference because 0 is a valid file
128 	 * descriptor number, and the child needs to be able to distinguish
129 	 * itself from the parent using the return value.
130 	 */
131 	error = fork1(td, RFFDG | RFPROC | RFPROCDESC, 0, &p2,
132 	    &fd, uap->flags);
133 	if (error == 0) {
134 		td->td_retval[0] = p2->p_pid;
135 		td->td_retval[1] = 0;
136 		error = copyout(&fd, uap->fdp, sizeof(fd));
137 	}
138 	return (error);
139 #else
140 	return (ENOSYS);
141 #endif
142 }
143 
144 /* ARGSUSED */
145 int
146 sys_vfork(struct thread *td, struct vfork_args *uap)
147 {
148 	int error, flags;
149 	struct proc *p2;
150 
151 	flags = RFFDG | RFPROC | RFPPWAIT | RFMEM;
152 	error = fork1(td, flags, 0, &p2, NULL, 0);
153 	if (error == 0) {
154 		td->td_retval[0] = p2->p_pid;
155 		td->td_retval[1] = 0;
156 	}
157 	return (error);
158 }
159 
160 int
161 sys_rfork(struct thread *td, struct rfork_args *uap)
162 {
163 	struct proc *p2;
164 	int error;
165 
166 	/* Don't allow kernel-only flags. */
167 	if ((uap->flags & RFKERNELONLY) != 0)
168 		return (EINVAL);
169 
170 	AUDIT_ARG_FFLAGS(uap->flags);
171 	error = fork1(td, uap->flags, 0, &p2, NULL, 0);
172 	if (error == 0) {
173 		td->td_retval[0] = p2 ? p2->p_pid : 0;
174 		td->td_retval[1] = 0;
175 	}
176 	return (error);
177 }
178 
179 int	nprocs = 1;		/* process 0 */
180 int	lastpid = 0;
181 SYSCTL_INT(_kern, OID_AUTO, lastpid, CTLFLAG_RD, &lastpid, 0,
182     "Last used PID");
183 
184 /*
185  * Random component to lastpid generation.  We mix in a random factor to make
186  * it a little harder to predict.  We sanity check the modulus value to avoid
187  * doing it in critical paths.  Don't let it be too small or we pointlessly
188  * waste randomness entropy, and don't let it be impossibly large.  Using a
189  * modulus that is too big causes a LOT more process table scans and slows
190  * down fork processing as the pidchecked caching is defeated.
191  */
192 static int randompid = 0;
193 
194 static int
195 sysctl_kern_randompid(SYSCTL_HANDLER_ARGS)
196 {
197 	int error, pid;
198 
199 	error = sysctl_wire_old_buffer(req, sizeof(int));
200 	if (error != 0)
201 		return(error);
202 	sx_xlock(&allproc_lock);
203 	pid = randompid;
204 	error = sysctl_handle_int(oidp, &pid, 0, req);
205 	if (error == 0 && req->newptr != NULL) {
206 		if (pid < 0 || pid > pid_max - 100)	/* out of range */
207 			pid = pid_max - 100;
208 		else if (pid < 2)			/* NOP */
209 			pid = 0;
210 		else if (pid < 100)			/* Make it reasonable */
211 			pid = 100;
212 		randompid = pid;
213 	}
214 	sx_xunlock(&allproc_lock);
215 	return (error);
216 }
217 
218 SYSCTL_PROC(_kern, OID_AUTO, randompid, CTLTYPE_INT|CTLFLAG_RW,
219     0, 0, sysctl_kern_randompid, "I", "Random PID modulus");
220 
221 static int
222 fork_findpid(int flags)
223 {
224 	struct proc *p;
225 	int trypid;
226 	static int pidchecked = 0;
227 
228 	/*
229 	 * Requires allproc_lock in order to iterate over the list
230 	 * of processes, and proctree_lock to access p_pgrp.
231 	 */
232 	sx_assert(&allproc_lock, SX_LOCKED);
233 	sx_assert(&proctree_lock, SX_LOCKED);
234 
235 	/*
236 	 * Find an unused process ID.  We remember a range of unused IDs
237 	 * ready to use (from lastpid+1 through pidchecked-1).
238 	 *
239 	 * If RFHIGHPID is set (used during system boot), do not allocate
240 	 * low-numbered pids.
241 	 */
242 	trypid = lastpid + 1;
243 	if (flags & RFHIGHPID) {
244 		if (trypid < 10)
245 			trypid = 10;
246 	} else {
247 		if (randompid)
248 			trypid += arc4random() % randompid;
249 	}
250 retry:
251 	/*
252 	 * If the process ID prototype has wrapped around,
253 	 * restart somewhat above 0, as the low-numbered procs
254 	 * tend to include daemons that don't exit.
255 	 */
256 	if (trypid >= pid_max) {
257 		trypid = trypid % pid_max;
258 		if (trypid < 100)
259 			trypid += 100;
260 		pidchecked = 0;
261 	}
262 	if (trypid >= pidchecked) {
263 		int doingzomb = 0;
264 
265 		pidchecked = PID_MAX;
266 		/*
267 		 * Scan the active and zombie procs to check whether this pid
268 		 * is in use.  Remember the lowest pid that's greater
269 		 * than trypid, so we can avoid checking for a while.
270 		 */
271 		p = LIST_FIRST(&allproc);
272 again:
273 		for (; p != NULL; p = LIST_NEXT(p, p_list)) {
274 			while (p->p_pid == trypid ||
275 			    (p->p_pgrp != NULL &&
276 			    (p->p_pgrp->pg_id == trypid ||
277 			    (p->p_session != NULL &&
278 			    p->p_session->s_sid == trypid)))) {
279 				trypid++;
280 				if (trypid >= pidchecked)
281 					goto retry;
282 			}
283 			if (p->p_pid > trypid && pidchecked > p->p_pid)
284 				pidchecked = p->p_pid;
285 			if (p->p_pgrp != NULL) {
286 				if (p->p_pgrp->pg_id > trypid &&
287 				    pidchecked > p->p_pgrp->pg_id)
288 					pidchecked = p->p_pgrp->pg_id;
289 				if (p->p_session != NULL &&
290 				    p->p_session->s_sid > trypid &&
291 				    pidchecked > p->p_session->s_sid)
292 					pidchecked = p->p_session->s_sid;
293 			}
294 		}
295 		if (!doingzomb) {
296 			doingzomb = 1;
297 			p = LIST_FIRST(&zombproc);
298 			goto again;
299 		}
300 	}
301 
302 	/*
303 	 * RFHIGHPID does not mess with the lastpid counter during boot.
304 	 */
305 	if (flags & RFHIGHPID)
306 		pidchecked = 0;
307 	else
308 		lastpid = trypid;
309 
310 	return (trypid);
311 }
312 
313 static int
314 fork_norfproc(struct thread *td, int flags)
315 {
316 	int error;
317 	struct proc *p1;
318 
319 	KASSERT((flags & RFPROC) == 0,
320 	    ("fork_norfproc called with RFPROC set"));
321 	p1 = td->td_proc;
322 
323 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
324 	    (flags & (RFCFDG | RFFDG))) {
325 		PROC_LOCK(p1);
326 		if (thread_single(SINGLE_BOUNDARY)) {
327 			PROC_UNLOCK(p1);
328 			return (ERESTART);
329 		}
330 		PROC_UNLOCK(p1);
331 	}
332 
333 	error = vm_forkproc(td, NULL, NULL, NULL, flags);
334 	if (error)
335 		goto fail;
336 
337 	/*
338 	 * Close all file descriptors.
339 	 */
340 	if (flags & RFCFDG) {
341 		struct filedesc *fdtmp;
342 		fdtmp = fdinit(td->td_proc->p_fd);
343 		fdescfree(td);
344 		p1->p_fd = fdtmp;
345 	}
346 
347 	/*
348 	 * Unshare file descriptors (from parent).
349 	 */
350 	if (flags & RFFDG)
351 		fdunshare(p1, td);
352 
353 fail:
354 	if (((p1->p_flag & (P_HADTHREADS|P_SYSTEM)) == P_HADTHREADS) &&
355 	    (flags & (RFCFDG | RFFDG))) {
356 		PROC_LOCK(p1);
357 		thread_single_end();
358 		PROC_UNLOCK(p1);
359 	}
360 	return (error);
361 }
362 
363 static void
364 do_fork(struct thread *td, int flags, struct proc *p2, struct thread *td2,
365     struct vmspace *vm2, int pdflags)
366 {
367 	struct proc *p1, *pptr;
368 	int p2_held, trypid;
369 	struct filedesc *fd;
370 	struct filedesc_to_leader *fdtol;
371 	struct sigacts *newsigacts;
372 
373 	sx_assert(&proctree_lock, SX_SLOCKED);
374 	sx_assert(&allproc_lock, SX_XLOCKED);
375 
376 	p2_held = 0;
377 	p1 = td->td_proc;
378 
379 	/*
380 	 * Increment the nprocs resource before blocking can occur.  There
381 	 * are hard-limits as to the number of processes that can run.
382 	 */
383 	nprocs++;
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 	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 	PROC_UNLOCK(p1);
404 
405 	bzero(&p2->p_startzero,
406 	    __rangeof(struct proc, p_startzero, p_endzero));
407 
408 	p2->p_ucred = crhold(td->td_ucred);
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);
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_swtick = ticks;
493 	if (p1->p_flag & P_PROFIL)
494 		startprofclock(p2);
495 	td2->td_ucred = crhold(p2->p_ucred);
496 
497 	if (flags & RFSIGSHARE) {
498 		p2->p_sigacts = sigacts_hold(p1->p_sigacts);
499 	} else {
500 		sigacts_copy(newsigacts, p1->p_sigacts);
501 		p2->p_sigacts = newsigacts;
502 	}
503 
504 	if (flags & RFTSIGZMB)
505 	        p2->p_sigparent = RFTSIGNUM(flags);
506 	else if (flags & RFLINUXTHPN)
507 	        p2->p_sigparent = SIGUSR1;
508 	else
509 	        p2->p_sigparent = SIGCHLD;
510 
511 	p2->p_textvp = p1->p_textvp;
512 	p2->p_fd = fd;
513 	p2->p_fdtol = fdtol;
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 #ifdef PROCDESC
654 	/*
655 	 * Associate the process descriptor with the process before anything
656 	 * can happen that might cause that process to need the descriptor.
657 	 * However, don't do this until after fork(2) can no longer fail.
658 	 */
659 	if (flags & RFPROCDESC)
660 		procdesc_new(p2, pdflags);
661 #endif
662 
663 	/*
664 	 * Both processes are set up, now check if any loadable modules want
665 	 * to adjust anything.
666 	 */
667 	EVENTHANDLER_INVOKE(process_fork, p1, p2, flags);
668 
669 	/*
670 	 * Set the child start time and mark the process as being complete.
671 	 */
672 	PROC_LOCK(p2);
673 	PROC_LOCK(p1);
674 	microuptime(&p2->p_stats->p_start);
675 	PROC_SLOCK(p2);
676 	p2->p_state = PRS_NORMAL;
677 	PROC_SUNLOCK(p2);
678 
679 #ifdef KDTRACE_HOOKS
680 	/*
681 	 * Tell the DTrace fasttrap provider about the new process
682 	 * if it has registered an interest. We have to do this only after
683 	 * p_state is PRS_NORMAL since the fasttrap module will use pfind()
684 	 * later on.
685 	 */
686 	if (dtrace_fasttrap_fork)
687 		dtrace_fasttrap_fork(p1, p2);
688 #endif
689 	if ((p1->p_flag & (P_TRACED | P_FOLLOWFORK)) == (P_TRACED |
690 	    P_FOLLOWFORK)) {
691 		/*
692 		 * Arrange for debugger to receive the fork event.
693 		 *
694 		 * We can report PL_FLAG_FORKED regardless of
695 		 * P_FOLLOWFORK settings, but it does not make a sense
696 		 * for runaway child.
697 		 */
698 		td->td_dbgflags |= TDB_FORK;
699 		td->td_dbg_forked = p2->p_pid;
700 		td2->td_dbgflags |= TDB_STOPATFORK;
701 		_PHOLD(p2);
702 		p2_held = 1;
703 	}
704 	if (flags & RFPPWAIT) {
705 		td->td_pflags |= TDP_RFPPWAIT;
706 		td->td_rfppwait_p = p2;
707 	}
708 	PROC_UNLOCK(p2);
709 	if ((flags & RFSTOPPED) == 0) {
710 		/*
711 		 * If RFSTOPPED not requested, make child runnable and
712 		 * add to run queue.
713 		 */
714 		thread_lock(td2);
715 		TD_SET_CAN_RUN(td2);
716 		sched_add(td2, SRQ_BORING);
717 		thread_unlock(td2);
718 	}
719 
720 	/*
721 	 * Now can be swapped.
722 	 */
723 	_PRELE(p1);
724 	PROC_UNLOCK(p1);
725 
726 	/*
727 	 * Tell any interested parties about the new process.
728 	 */
729 	knote_fork(&p1->p_klist, p2->p_pid);
730 	SDT_PROBE(proc, kernel, , create, p2, p1, flags, 0, 0);
731 
732 	/*
733 	 * Wait until debugger is attached to child.
734 	 */
735 	PROC_LOCK(p2);
736 	while ((td2->td_dbgflags & TDB_STOPATFORK) != 0)
737 		cv_wait(&p2->p_dbgwait, &p2->p_mtx);
738 	if (p2_held)
739 		_PRELE(p2);
740 	PROC_UNLOCK(p2);
741 }
742 
743 int
744 fork1(struct thread *td, int flags, int pages, struct proc **procp,
745     int *procdescp, int pdflags)
746 {
747 	struct proc *p1;
748 	struct proc *newproc;
749 	int ok;
750 	struct thread *td2;
751 	struct vmspace *vm2;
752 	vm_ooffset_t mem_charged;
753 	int error;
754 	static int curfail;
755 	static struct timeval lastfail;
756 #ifdef PROCDESC
757 	struct file *fp_procdesc = NULL;
758 #endif
759 
760 	/* Check for the undefined or unimplemented flags. */
761 	if ((flags & ~(RFFLAGS | RFTSIGFLAGS(RFTSIGMASK))) != 0)
762 		return (EINVAL);
763 
764 	/* Signal value requires RFTSIGZMB. */
765 	if ((flags & RFTSIGFLAGS(RFTSIGMASK)) != 0 && (flags & RFTSIGZMB) == 0)
766 		return (EINVAL);
767 
768 	/* Can't copy and clear. */
769 	if ((flags & (RFFDG|RFCFDG)) == (RFFDG|RFCFDG))
770 		return (EINVAL);
771 
772 	/* Check the validity of the signal number. */
773 	if ((flags & RFTSIGZMB) != 0 && (u_int)RFTSIGNUM(flags) > _SIG_MAXSIG)
774 		return (EINVAL);
775 
776 #ifdef PROCDESC
777 	if ((flags & RFPROCDESC) != 0) {
778 		/* Can't not create a process yet get a process descriptor. */
779 		if ((flags & RFPROC) == 0)
780 			return (EINVAL);
781 
782 		/* Must provide a place to put a procdesc if creating one. */
783 		if (procdescp == NULL)
784 			return (EINVAL);
785 	}
786 #endif
787 
788 	p1 = td->td_proc;
789 
790 	/*
791 	 * Here we don't create a new process, but we divorce
792 	 * certain parts of a process from itself.
793 	 */
794 	if ((flags & RFPROC) == 0) {
795 		*procp = NULL;
796 		return (fork_norfproc(td, flags));
797 	}
798 
799 #ifdef PROCDESC
800 	/*
801 	 * If required, create a process descriptor in the parent first; we
802 	 * will abandon it if something goes wrong. We don't finit() until
803 	 * later.
804 	 */
805 	if (flags & RFPROCDESC) {
806 		error = falloc(td, &fp_procdesc, procdescp, 0);
807 		if (error != 0)
808 			return (error);
809 	}
810 #endif
811 
812 	mem_charged = 0;
813 	vm2 = NULL;
814 	if (pages == 0)
815 		pages = KSTACK_PAGES;
816 	/* Allocate new proc. */
817 	newproc = uma_zalloc(proc_zone, M_WAITOK);
818 	td2 = FIRST_THREAD_IN_PROC(newproc);
819 	if (td2 == NULL) {
820 		td2 = thread_alloc(pages);
821 		if (td2 == NULL) {
822 			error = ENOMEM;
823 			goto fail1;
824 		}
825 		proc_linkup(newproc, td2);
826 	} else {
827 		if (td2->td_kstack == 0 || td2->td_kstack_pages != pages) {
828 			if (td2->td_kstack != 0)
829 				vm_thread_dispose(td2);
830 			if (!thread_alloc_stack(td2, pages)) {
831 				error = ENOMEM;
832 				goto fail1;
833 			}
834 		}
835 	}
836 
837 	if ((flags & RFMEM) == 0) {
838 		vm2 = vmspace_fork(p1->p_vmspace, &mem_charged);
839 		if (vm2 == NULL) {
840 			error = ENOMEM;
841 			goto fail1;
842 		}
843 		if (!swap_reserve(mem_charged)) {
844 			/*
845 			 * The swap reservation failed. The accounting
846 			 * from the entries of the copied vm2 will be
847 			 * substracted in vmspace_free(), so force the
848 			 * reservation there.
849 			 */
850 			swap_reserve_force(mem_charged);
851 			error = ENOMEM;
852 			goto fail1;
853 		}
854 	} else
855 		vm2 = NULL;
856 
857 	/*
858 	 * XXX: This is ugly; when we copy resource usage, we need to bump
859 	 *      per-cred resource counters.
860 	 */
861 	newproc->p_ucred = p1->p_ucred;
862 
863 	/*
864 	 * Initialize resource accounting for the child process.
865 	 */
866 	error = racct_proc_fork(p1, newproc);
867 	if (error != 0) {
868 		error = EAGAIN;
869 		goto fail1;
870 	}
871 
872 #ifdef MAC
873 	mac_proc_init(newproc);
874 #endif
875 	knlist_init_mtx(&newproc->p_klist, &newproc->p_mtx);
876 	STAILQ_INIT(&newproc->p_ktr);
877 
878 	/* We have to lock the process tree while we look for a pid. */
879 	sx_slock(&proctree_lock);
880 
881 	/*
882 	 * Although process entries are dynamically created, we still keep
883 	 * a global limit on the maximum number we will create.  Don't allow
884 	 * a nonprivileged user to use the last ten processes; don't let root
885 	 * exceed the limit. The variable nprocs is the current number of
886 	 * processes, maxproc is the limit.
887 	 */
888 	sx_xlock(&allproc_lock);
889 	if ((nprocs >= maxproc - 10 && priv_check_cred(td->td_ucred,
890 	    PRIV_MAXPROC, 0) != 0) || nprocs >= maxproc) {
891 		error = EAGAIN;
892 		goto fail;
893 	}
894 
895 	/*
896 	 * Increment the count of procs running with this uid. Don't allow
897 	 * a nonprivileged user to exceed their current limit.
898 	 *
899 	 * XXXRW: Can we avoid privilege here if it's not needed?
900 	 */
901 	error = priv_check_cred(td->td_ucred, PRIV_PROC_LIMIT, 0);
902 	if (error == 0)
903 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1, 0);
904 	else {
905 		PROC_LOCK(p1);
906 		ok = chgproccnt(td->td_ucred->cr_ruidinfo, 1,
907 		    lim_cur(p1, RLIMIT_NPROC));
908 		PROC_UNLOCK(p1);
909 	}
910 	if (ok) {
911 		do_fork(td, flags, newproc, td2, vm2, pdflags);
912 
913 		/*
914 		 * Return child proc pointer to parent.
915 		 */
916 		*procp = newproc;
917 #ifdef PROCDESC
918 		if (flags & RFPROCDESC) {
919 			procdesc_finit(newproc->p_procdesc, fp_procdesc);
920 			fdrop(fp_procdesc, td);
921 		}
922 #endif
923 		racct_proc_fork_done(newproc);
924 		return (0);
925 	}
926 
927 	error = EAGAIN;
928 fail:
929 	sx_sunlock(&proctree_lock);
930 	if (ppsratecheck(&lastfail, &curfail, 1))
931 		printf("maxproc limit exceeded by uid %u (pid %d); see tuning(7) and login.conf(5)\n",
932 		    td->td_ucred->cr_ruid, p1->p_pid);
933 	sx_xunlock(&allproc_lock);
934 #ifdef MAC
935 	mac_proc_destroy(newproc);
936 #endif
937 	racct_proc_exit(newproc);
938 fail1:
939 	if (vm2 != NULL)
940 		vmspace_free(vm2);
941 	uma_zfree(proc_zone, newproc);
942 #ifdef PROCDESC
943 	if ((flags & RFPROCDESC) != 0 && fp_procdesc != NULL) {
944 		fdclose(td->td_proc->p_fd, fp_procdesc, *procdescp, td);
945 		fdrop(fp_procdesc, td);
946 	}
947 #endif
948 	pause("fork", hz / 2);
949 	return (error);
950 }
951 
952 /*
953  * Handle the return of a child process from fork1().  This function
954  * is called from the MD fork_trampoline() entry point.
955  */
956 void
957 fork_exit(void (*callout)(void *, struct trapframe *), void *arg,
958     struct trapframe *frame)
959 {
960 	struct proc *p;
961 	struct thread *td;
962 	struct thread *dtd;
963 
964 	td = curthread;
965 	p = td->td_proc;
966 	KASSERT(p->p_state == PRS_NORMAL, ("executing process is still new"));
967 
968 	CTR4(KTR_PROC, "fork_exit: new thread %p (td_sched %p, pid %d, %s)",
969 		td, td->td_sched, p->p_pid, td->td_name);
970 
971 	sched_fork_exit(td);
972 	/*
973 	* Processes normally resume in mi_switch() after being
974 	* cpu_switch()'ed to, but when children start up they arrive here
975 	* instead, so we must do much the same things as mi_switch() would.
976 	*/
977 	if ((dtd = PCPU_GET(deadthread))) {
978 		PCPU_SET(deadthread, NULL);
979 		thread_stash(dtd);
980 	}
981 	thread_unlock(td);
982 
983 	/*
984 	 * cpu_set_fork_handler intercepts this function call to
985 	 * have this call a non-return function to stay in kernel mode.
986 	 * initproc has its own fork handler, but it does return.
987 	 */
988 	KASSERT(callout != NULL, ("NULL callout in fork_exit"));
989 	callout(arg, frame);
990 
991 	/*
992 	 * Check if a kernel thread misbehaved and returned from its main
993 	 * function.
994 	 */
995 	if (p->p_flag & P_KTHREAD) {
996 		printf("Kernel thread \"%s\" (pid %d) exited prematurely.\n",
997 		    td->td_name, p->p_pid);
998 		kproc_exit(0);
999 	}
1000 	mtx_assert(&Giant, MA_NOTOWNED);
1001 
1002 	if (p->p_sysent->sv_schedtail != NULL)
1003 		(p->p_sysent->sv_schedtail)(td);
1004 }
1005 
1006 /*
1007  * Simplified back end of syscall(), used when returning from fork()
1008  * directly into user mode.  Giant is not held on entry, and must not
1009  * be held on return.  This function is passed in to fork_exit() as the
1010  * first parameter and is called when returning to a new userland process.
1011  */
1012 void
1013 fork_return(struct thread *td, struct trapframe *frame)
1014 {
1015 	struct proc *p, *dbg;
1016 
1017 	if (td->td_dbgflags & TDB_STOPATFORK) {
1018 		p = td->td_proc;
1019 		sx_xlock(&proctree_lock);
1020 		PROC_LOCK(p);
1021 		if ((p->p_pptr->p_flag & (P_TRACED | P_FOLLOWFORK)) ==
1022 		    (P_TRACED | P_FOLLOWFORK)) {
1023 			/*
1024 			 * If debugger still wants auto-attach for the
1025 			 * parent's children, do it now.
1026 			 */
1027 			dbg = p->p_pptr->p_pptr;
1028 			p->p_flag |= P_TRACED;
1029 			p->p_oppid = p->p_pptr->p_pid;
1030 			proc_reparent(p, dbg);
1031 			sx_xunlock(&proctree_lock);
1032 			td->td_dbgflags |= TDB_CHILD;
1033 			ptracestop(td, SIGSTOP);
1034 			td->td_dbgflags &= ~TDB_CHILD;
1035 		} else {
1036 			/*
1037 			 * ... otherwise clear the request.
1038 			 */
1039 			sx_xunlock(&proctree_lock);
1040 			td->td_dbgflags &= ~TDB_STOPATFORK;
1041 			cv_broadcast(&p->p_dbgwait);
1042 		}
1043 		PROC_UNLOCK(p);
1044 	}
1045 
1046 	userret(td, frame);
1047 
1048 #ifdef KTRACE
1049 	if (KTRPOINT(td, KTR_SYSRET))
1050 		ktrsysret(SYS_fork, 0, 0);
1051 #endif
1052 }
1053