xref: /freebsd/sys/kern/sys_procdesc.c (revision e91d723ad446b5429318b24f4578a4f7a160f65e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2009, 2016 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * This software was developed at the University of Cambridge Computer
8  * Laboratory with support from a grant from Google, Inc.
9  *
10  * Portions of this software were developed by BAE Systems, the University of
11  * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
12  * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
13  * Computing (TC) research program.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  */
36 
37 /*-
38  * FreeBSD process descriptor facility.
39  *
40  * Some processes are represented by a file descriptor, which will be used in
41  * preference to signaling and pids for the purposes of process management,
42  * and is, in effect, a form of capability.  When a process descriptor is
43  * used with a process, it ceases to be visible to certain traditional UNIX
44  * process facilities, such as waitpid(2).
45  *
46  * Some semantics:
47  *
48  * - At most one process descriptor will exist for any process, although
49  *   references to that descriptor may be held from many processes (or even
50  *   be in flight between processes over a local domain socket).
51  * - Last close on the process descriptor will terminate the process using
52  *   SIGKILL and reparent it to init so that there's a process to reap it
53  *   when it's done exiting.
54  * - If the process exits before the descriptor is closed, it will not
55  *   generate SIGCHLD on termination, or be picked up by waitpid().
56  * - The pdkill(2) system call may be used to deliver a signal to the process
57  *   using its process descriptor.
58  * - The pdwait4(2) system call may be used to block (or not) on a process
59  *   descriptor to collect termination information.
60  *
61  * Open questions:
62  *
63  * - How to handle ptrace(2)?
64  * - Will we want to add a pidtoprocdesc(2) system call to allow process
65  *   descriptors to be created for processes without pdfork(2)?
66  */
67 
68 #include <sys/cdefs.h>
69 __FBSDID("$FreeBSD$");
70 
71 #include <sys/param.h>
72 #include <sys/capsicum.h>
73 #include <sys/fcntl.h>
74 #include <sys/file.h>
75 #include <sys/filedesc.h>
76 #include <sys/kernel.h>
77 #include <sys/lock.h>
78 #include <sys/mutex.h>
79 #include <sys/poll.h>
80 #include <sys/proc.h>
81 #include <sys/procdesc.h>
82 #include <sys/resourcevar.h>
83 #include <sys/stat.h>
84 #include <sys/sysproto.h>
85 #include <sys/sysctl.h>
86 #include <sys/systm.h>
87 #include <sys/ucred.h>
88 #include <sys/user.h>
89 
90 #include <security/audit/audit.h>
91 
92 #include <vm/uma.h>
93 
94 FEATURE(process_descriptors, "Process Descriptors");
95 
96 static uma_zone_t procdesc_zone;
97 
98 static fo_poll_t	procdesc_poll;
99 static fo_kqfilter_t	procdesc_kqfilter;
100 static fo_stat_t	procdesc_stat;
101 static fo_close_t	procdesc_close;
102 static fo_fill_kinfo_t	procdesc_fill_kinfo;
103 
104 static struct fileops procdesc_ops = {
105 	.fo_read = invfo_rdwr,
106 	.fo_write = invfo_rdwr,
107 	.fo_truncate = invfo_truncate,
108 	.fo_ioctl = invfo_ioctl,
109 	.fo_poll = procdesc_poll,
110 	.fo_kqfilter = procdesc_kqfilter,
111 	.fo_stat = procdesc_stat,
112 	.fo_close = procdesc_close,
113 	.fo_chmod = invfo_chmod,
114 	.fo_chown = invfo_chown,
115 	.fo_sendfile = invfo_sendfile,
116 	.fo_fill_kinfo = procdesc_fill_kinfo,
117 	.fo_flags = DFLAG_PASSABLE,
118 };
119 
120 /*
121  * Initialize with VFS so that process descriptors are available along with
122  * other file descriptor types.  As long as it runs before init(8) starts,
123  * there shouldn't be a problem.
124  */
125 static void
126 procdesc_init(void *dummy __unused)
127 {
128 
129 	procdesc_zone = uma_zcreate("procdesc", sizeof(struct procdesc),
130 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
131 	if (procdesc_zone == NULL)
132 		panic("procdesc_init: procdesc_zone not initialized");
133 }
134 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_ANY, procdesc_init, NULL);
135 
136 /*
137  * Return a locked process given a process descriptor, or ESRCH if it has
138  * died.
139  */
140 int
141 procdesc_find(struct thread *td, int fd, cap_rights_t *rightsp,
142     struct proc **p)
143 {
144 	struct procdesc *pd;
145 	struct file *fp;
146 	int error;
147 
148 	error = fget(td, fd, rightsp, &fp);
149 	if (error)
150 		return (error);
151 	if (fp->f_type != DTYPE_PROCDESC) {
152 		error = EBADF;
153 		goto out;
154 	}
155 	pd = fp->f_data;
156 	sx_slock(&proctree_lock);
157 	if (pd->pd_proc != NULL) {
158 		*p = pd->pd_proc;
159 		PROC_LOCK(*p);
160 	} else
161 		error = ESRCH;
162 	sx_sunlock(&proctree_lock);
163 out:
164 	fdrop(fp, td);
165 	return (error);
166 }
167 
168 /*
169  * Function to be used by procstat(1) sysctls when returning procdesc
170  * information.
171  */
172 pid_t
173 procdesc_pid(struct file *fp_procdesc)
174 {
175 	struct procdesc *pd;
176 
177 	KASSERT(fp_procdesc->f_type == DTYPE_PROCDESC,
178 	   ("procdesc_pid: !procdesc"));
179 
180 	pd = fp_procdesc->f_data;
181 	return (pd->pd_pid);
182 }
183 
184 /*
185  * Retrieve the PID associated with a process descriptor.
186  */
187 int
188 kern_pdgetpid(struct thread *td, int fd, cap_rights_t *rightsp, pid_t *pidp)
189 {
190 	struct file *fp;
191 	int error;
192 
193 	error = fget(td, fd, rightsp, &fp);
194 	if (error)
195 		return (error);
196 	if (fp->f_type != DTYPE_PROCDESC) {
197 		error = EBADF;
198 		goto out;
199 	}
200 	*pidp = procdesc_pid(fp);
201 out:
202 	fdrop(fp, td);
203 	return (error);
204 }
205 
206 /*
207  * System call to return the pid of a process given its process descriptor.
208  */
209 int
210 sys_pdgetpid(struct thread *td, struct pdgetpid_args *uap)
211 {
212 	pid_t pid;
213 	int error;
214 
215 	AUDIT_ARG_FD(uap->fd);
216 	error = kern_pdgetpid(td, uap->fd, &cap_pdgetpid_rights, &pid);
217 	if (error == 0)
218 		error = copyout(&pid, uap->pidp, sizeof(pid));
219 	return (error);
220 }
221 
222 /*
223  * When a new process is forked by pdfork(), a file descriptor is allocated
224  * by the fork code first, then the process is forked, and then we get a
225  * chance to set up the process descriptor.  Failure is not permitted at this
226  * point, so procdesc_new() must succeed.
227  */
228 void
229 procdesc_new(struct proc *p, int flags)
230 {
231 	struct procdesc *pd;
232 
233 	pd = uma_zalloc(procdesc_zone, M_WAITOK | M_ZERO);
234 	pd->pd_proc = p;
235 	pd->pd_pid = p->p_pid;
236 	p->p_procdesc = pd;
237 	pd->pd_flags = 0;
238 	if (flags & PD_DAEMON)
239 		pd->pd_flags |= PDF_DAEMON;
240 	PROCDESC_LOCK_INIT(pd);
241 	knlist_init_mtx(&pd->pd_selinfo.si_note, &pd->pd_lock);
242 
243 	/*
244 	 * Process descriptors start out with two references: one from their
245 	 * struct file, and the other from their struct proc.
246 	 */
247 	refcount_init(&pd->pd_refcount, 2);
248 }
249 
250 /*
251  * Create a new process decriptor for the process that refers to it.
252  */
253 int
254 procdesc_falloc(struct thread *td, struct file **resultfp, int *resultfd,
255     int flags, struct filecaps *fcaps)
256 {
257 	int fflags;
258 
259 	fflags = 0;
260 	if (flags & PD_CLOEXEC)
261 		fflags = O_CLOEXEC;
262 
263 	return (falloc_caps(td, resultfp, resultfd, fflags, fcaps));
264 }
265 
266 /*
267  * Initialize a file with a process descriptor.
268  */
269 void
270 procdesc_finit(struct procdesc *pdp, struct file *fp)
271 {
272 
273 	finit(fp, FREAD | FWRITE, DTYPE_PROCDESC, pdp, &procdesc_ops);
274 }
275 
276 static void
277 procdesc_free(struct procdesc *pd)
278 {
279 
280 	/*
281 	 * When the last reference is released, we assert that the descriptor
282 	 * has been closed, but not that the process has exited, as we will
283 	 * detach the descriptor before the process dies if the descript is
284 	 * closed, as we can't wait synchronously.
285 	 */
286 	if (refcount_release(&pd->pd_refcount)) {
287 		KASSERT(pd->pd_proc == NULL,
288 		    ("procdesc_free: pd_proc != NULL"));
289 		KASSERT((pd->pd_flags & PDF_CLOSED),
290 		    ("procdesc_free: !PDF_CLOSED"));
291 
292 		knlist_destroy(&pd->pd_selinfo.si_note);
293 		PROCDESC_LOCK_DESTROY(pd);
294 		uma_zfree(procdesc_zone, pd);
295 	}
296 }
297 
298 /*
299  * procdesc_exit() - notify a process descriptor that its process is exiting.
300  * We use the proctree_lock to ensure that process exit either happens
301  * strictly before or strictly after a concurrent call to procdesc_close().
302  */
303 int
304 procdesc_exit(struct proc *p)
305 {
306 	struct procdesc *pd;
307 
308 	sx_assert(&proctree_lock, SA_XLOCKED);
309 	PROC_LOCK_ASSERT(p, MA_OWNED);
310 	KASSERT(p->p_procdesc != NULL, ("procdesc_exit: p_procdesc NULL"));
311 
312 	pd = p->p_procdesc;
313 
314 	PROCDESC_LOCK(pd);
315 	KASSERT((pd->pd_flags & PDF_CLOSED) == 0 || p->p_pptr == p->p_reaper,
316 	    ("procdesc_exit: closed && parent not reaper"));
317 
318 	pd->pd_flags |= PDF_EXITED;
319 	pd->pd_xstat = KW_EXITCODE(p->p_xexit, p->p_xsig);
320 
321 	/*
322 	 * If the process descriptor has been closed, then we have nothing
323 	 * to do; return 1 so that init will get SIGCHLD and do the reaping.
324 	 * Clean up the procdesc now rather than letting it happen during
325 	 * that reap.
326 	 */
327 	if (pd->pd_flags & PDF_CLOSED) {
328 		PROCDESC_UNLOCK(pd);
329 		pd->pd_proc = NULL;
330 		p->p_procdesc = NULL;
331 		procdesc_free(pd);
332 		return (1);
333 	}
334 	if (pd->pd_flags & PDF_SELECTED) {
335 		pd->pd_flags &= ~PDF_SELECTED;
336 		selwakeup(&pd->pd_selinfo);
337 	}
338 	KNOTE_LOCKED(&pd->pd_selinfo.si_note, NOTE_EXIT);
339 	PROCDESC_UNLOCK(pd);
340 	return (0);
341 }
342 
343 /*
344  * When a process descriptor is reaped, perhaps as a result of close() or
345  * pdwait4(), release the process's reference on the process descriptor.
346  */
347 void
348 procdesc_reap(struct proc *p)
349 {
350 	struct procdesc *pd;
351 
352 	sx_assert(&proctree_lock, SA_XLOCKED);
353 	KASSERT(p->p_procdesc != NULL, ("procdesc_reap: p_procdesc == NULL"));
354 
355 	pd = p->p_procdesc;
356 	pd->pd_proc = NULL;
357 	p->p_procdesc = NULL;
358 	procdesc_free(pd);
359 }
360 
361 /*
362  * procdesc_close() - last close on a process descriptor.  If the process is
363  * still running, terminate with SIGKILL (unless PDF_DAEMON is set) and let
364  * its reaper clean up the mess; if not, we have to clean up the zombie
365  * ourselves.
366  */
367 static int
368 procdesc_close(struct file *fp, struct thread *td)
369 {
370 	struct procdesc *pd;
371 	struct proc *p;
372 
373 	KASSERT(fp->f_type == DTYPE_PROCDESC, ("procdesc_close: !procdesc"));
374 
375 	pd = fp->f_data;
376 	fp->f_ops = &badfileops;
377 	fp->f_data = NULL;
378 
379 	sx_xlock(&proctree_lock);
380 	PROCDESC_LOCK(pd);
381 	pd->pd_flags |= PDF_CLOSED;
382 	PROCDESC_UNLOCK(pd);
383 	p = pd->pd_proc;
384 	if (p == NULL) {
385 		/*
386 		 * This is the case where process' exit status was already
387 		 * collected and procdesc_reap() was already called.
388 		 */
389 		sx_xunlock(&proctree_lock);
390 	} else {
391 		PROC_LOCK(p);
392 		AUDIT_ARG_PROCESS(p);
393 		if (p->p_state == PRS_ZOMBIE) {
394 			/*
395 			 * If the process is already dead and just awaiting
396 			 * reaping, do that now.  This will release the
397 			 * process's reference to the process descriptor when it
398 			 * calls back into procdesc_reap().
399 			 */
400 			proc_reap(curthread, p, NULL, 0);
401 		} else {
402 			/*
403 			 * If the process is not yet dead, we need to kill it,
404 			 * but we can't wait around synchronously for it to go
405 			 * away, as that path leads to madness (and deadlocks).
406 			 * First, detach the process from its descriptor so that
407 			 * its exit status will be reported normally.
408 			 */
409 			pd->pd_proc = NULL;
410 			p->p_procdesc = NULL;
411 			procdesc_free(pd);
412 
413 			/*
414 			 * Next, reparent it to its reaper (usually init(8)) so
415 			 * that there's someone to pick up the pieces; finally,
416 			 * terminate with prejudice.
417 			 */
418 			p->p_sigparent = SIGCHLD;
419 			proc_reparent(p, p->p_reaper, true);
420 			if ((pd->pd_flags & PDF_DAEMON) == 0)
421 				kern_psignal(p, SIGKILL);
422 			PROC_UNLOCK(p);
423 			sx_xunlock(&proctree_lock);
424 		}
425 	}
426 
427 	/*
428 	 * Release the file descriptor's reference on the process descriptor.
429 	 */
430 	procdesc_free(pd);
431 	return (0);
432 }
433 
434 static int
435 procdesc_poll(struct file *fp, int events, struct ucred *active_cred,
436     struct thread *td)
437 {
438 	struct procdesc *pd;
439 	int revents;
440 
441 	revents = 0;
442 	pd = fp->f_data;
443 	PROCDESC_LOCK(pd);
444 	if (pd->pd_flags & PDF_EXITED)
445 		revents |= POLLHUP;
446 	if (revents == 0) {
447 		selrecord(td, &pd->pd_selinfo);
448 		pd->pd_flags |= PDF_SELECTED;
449 	}
450 	PROCDESC_UNLOCK(pd);
451 	return (revents);
452 }
453 
454 static void
455 procdesc_kqops_detach(struct knote *kn)
456 {
457 	struct procdesc *pd;
458 
459 	pd = kn->kn_fp->f_data;
460 	knlist_remove(&pd->pd_selinfo.si_note, kn, 0);
461 }
462 
463 static int
464 procdesc_kqops_event(struct knote *kn, long hint)
465 {
466 	struct procdesc *pd;
467 	u_int event;
468 
469 	pd = kn->kn_fp->f_data;
470 	if (hint == 0) {
471 		/*
472 		 * Initial test after registration. Generate a NOTE_EXIT in
473 		 * case the process already terminated before registration.
474 		 */
475 		event = pd->pd_flags & PDF_EXITED ? NOTE_EXIT : 0;
476 	} else {
477 		/* Mask off extra data. */
478 		event = (u_int)hint & NOTE_PCTRLMASK;
479 	}
480 
481 	/* If the user is interested in this event, record it. */
482 	if (kn->kn_sfflags & event)
483 		kn->kn_fflags |= event;
484 
485 	/* Process is gone, so flag the event as finished. */
486 	if (event == NOTE_EXIT) {
487 		kn->kn_flags |= EV_EOF | EV_ONESHOT;
488 		if (kn->kn_fflags & NOTE_EXIT)
489 			kn->kn_data = pd->pd_xstat;
490 		if (kn->kn_fflags == 0)
491 			kn->kn_flags |= EV_DROP;
492 		return (1);
493 	}
494 
495 	return (kn->kn_fflags != 0);
496 }
497 
498 static struct filterops procdesc_kqops = {
499 	.f_isfd = 1,
500 	.f_detach = procdesc_kqops_detach,
501 	.f_event = procdesc_kqops_event,
502 };
503 
504 static int
505 procdesc_kqfilter(struct file *fp, struct knote *kn)
506 {
507 	struct procdesc *pd;
508 
509 	pd = fp->f_data;
510 	switch (kn->kn_filter) {
511 	case EVFILT_PROCDESC:
512 		kn->kn_fop = &procdesc_kqops;
513 		kn->kn_flags |= EV_CLEAR;
514 		knlist_add(&pd->pd_selinfo.si_note, kn, 0);
515 		return (0);
516 	default:
517 		return (EINVAL);
518 	}
519 }
520 
521 static int
522 procdesc_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
523     struct thread *td)
524 {
525 	struct procdesc *pd;
526 	struct timeval pstart, boottime;
527 
528 	/*
529 	 * XXXRW: Perhaps we should cache some more information from the
530 	 * process so that we can return it reliably here even after it has
531 	 * died.  For example, caching its credential data.
532 	 */
533 	bzero(sb, sizeof(*sb));
534 	pd = fp->f_data;
535 	sx_slock(&proctree_lock);
536 	if (pd->pd_proc != NULL) {
537 		PROC_LOCK(pd->pd_proc);
538 		AUDIT_ARG_PROCESS(pd->pd_proc);
539 
540 		/* Set birth and [acm] times to process start time. */
541 		pstart = pd->pd_proc->p_stats->p_start;
542 		getboottime(&boottime);
543 		timevaladd(&pstart, &boottime);
544 		TIMEVAL_TO_TIMESPEC(&pstart, &sb->st_birthtim);
545 		sb->st_atim = sb->st_birthtim;
546 		sb->st_ctim = sb->st_birthtim;
547 		sb->st_mtim = sb->st_birthtim;
548 		if (pd->pd_proc->p_state != PRS_ZOMBIE)
549 			sb->st_mode = S_IFREG | S_IRWXU;
550 		else
551 			sb->st_mode = S_IFREG;
552 		sb->st_uid = pd->pd_proc->p_ucred->cr_ruid;
553 		sb->st_gid = pd->pd_proc->p_ucred->cr_rgid;
554 		PROC_UNLOCK(pd->pd_proc);
555 	} else
556 		sb->st_mode = S_IFREG;
557 	sx_sunlock(&proctree_lock);
558 	return (0);
559 }
560 
561 static int
562 procdesc_fill_kinfo(struct file *fp, struct kinfo_file *kif,
563     struct filedesc *fdp)
564 {
565 	struct procdesc *pdp;
566 
567 	kif->kf_type = KF_TYPE_PROCDESC;
568 	pdp = fp->f_data;
569 	kif->kf_un.kf_proc.kf_pid = pdp->pd_pid;
570 	return (0);
571 }
572