xref: /illumos-gate/usr/src/cmd/svc/startd/wait.c (revision bbf21555)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  *
25  * Copyright 2018 Joyent, Inc.
26  */
27 
28 /*
29  * wait.c - asynchronous monitoring of "wait registered" start methods
30  *
31  * Use event ports to poll on the set of fds representing the /proc/[pid]/psinfo
32  * files.  If one of these fds returns an event, then we inform the restarter
33  * that it has stopped.
34  *
35  * The wait_info_list holds the series of processes currently being monitored
36  * for exit.  The wi_fd member, which contains the file descriptor of the psinfo
37  * file being polled upon ("event ported upon"), will be set to -1 if the file
38  * descriptor is inactive (already closed or not yet opened).
39  */
40 
41 #ifdef _FILE_OFFSET_BITS
42 #undef _FILE_OFFSET_BITS
43 #endif /* _FILE_OFFSET_BITS */
44 
45 #include <sys/resource.h>
46 #include <sys/stat.h>
47 #include <sys/types.h>
48 #include <sys/uio.h>
49 #include <sys/wait.h>
50 
51 #include <assert.h>
52 #include <errno.h>
53 #include <fcntl.h>
54 #include <libuutil.h>
55 #include <poll.h>
56 #include <port.h>
57 #include <pthread.h>
58 #include <procfs.h>
59 #include <string.h>
60 #include <stropts.h>
61 #include <unistd.h>
62 
63 #include "startd.h"
64 
65 #define	WAIT_FILES	262144		/* reasonably high maximum */
66 
67 static int port_fd;
68 static scf_handle_t *wait_hndl;
69 static struct rlimit init_fd_rlimit;
70 
71 static uu_list_pool_t *wait_info_pool;
72 static uu_list_t *wait_info_list;
73 
74 static pthread_mutex_t wait_info_lock;
75 
76 /*
77  * void wait_remove(wait_info_t *, int)
78  *   Remove the given wait_info structure from our list, performing various
79  *   cleanup operations along the way.  If the direct flag is false (meaning
80  *   that we are being called with from restarter instance list context) and
81  *   the instance should not be ignored, then notify the restarter that the
82  *   associated instance has exited. If the wi_ignore flag is true then it
83  *   means that the stop was initiated from within svc.startd, rather than
84  *   from outside it.
85  *
86  *   Since we may no longer be the startd that started this process, we only are
87  *   concerned with a waitpid(3C) failure if the wi_parent field is non-zero.
88  */
89 static void
wait_remove(wait_info_t * wi,int direct)90 wait_remove(wait_info_t *wi, int direct)
91 {
92 	int status;
93 	stop_cause_t cause = RSTOP_EXIT;
94 
95 	if (waitpid(wi->wi_pid, &status, 0) == -1) {
96 		if (wi->wi_parent)
97 			log_framework(LOG_INFO,
98 			    "instance %s waitpid failure: %s\n", wi->wi_fmri,
99 			    strerror(errno));
100 	} else {
101 		if (WEXITSTATUS(status) != 0) {
102 			log_framework(LOG_NOTICE,
103 			    "instance %s exited with status %d\n", wi->wi_fmri,
104 			    WEXITSTATUS(status));
105 			if (WEXITSTATUS(status) == SMF_EXIT_ERR_CONFIG)
106 				cause = RSTOP_ERR_CFG;
107 			else
108 				cause = RSTOP_ERR_EXIT;
109 		}
110 	}
111 
112 	MUTEX_LOCK(&wait_info_lock);
113 	if (wi->wi_fd != -1) {
114 		startd_close(wi->wi_fd);
115 		wi->wi_fd = -1;
116 	}
117 	uu_list_remove(wait_info_list, wi);
118 	MUTEX_UNLOCK(&wait_info_lock);
119 
120 	/*
121 	 * Make an attempt to clear out any utmpx record associated with this
122 	 * PID.
123 	 */
124 	utmpx_mark_dead(wi->wi_pid, status, B_FALSE);
125 
126 	if (!direct && !wi->wi_ignore) {
127 		/*
128 		 * Bind wait_hndl lazily.
129 		 */
130 		if (wait_hndl == NULL) {
131 			for (wait_hndl =
132 			    libscf_handle_create_bound(SCF_VERSION);
133 			    wait_hndl == NULL;
134 			    wait_hndl =
135 			    libscf_handle_create_bound(SCF_VERSION)) {
136 				log_error(LOG_INFO, "[wait_remove] Unable to "
137 				    "bind a new repository handle: %s\n",
138 				    scf_strerror(scf_error()));
139 				(void) sleep(2);
140 			}
141 		}
142 
143 		log_framework(LOG_DEBUG,
144 		    "wait_remove requesting stop of %s\n", wi->wi_fmri);
145 		(void) stop_instance_fmri(wait_hndl, wi->wi_fmri, cause);
146 	}
147 
148 	uu_list_node_fini(wi, &wi->wi_link, wait_info_pool);
149 	startd_free(wi, sizeof (wait_info_t));
150 }
151 
152 /*
153  * void wait_ignore_by_fmri(const char *)
154  *   wait_ignore_by_fmri is called when svc.startd is going to stop the
155  *   instance. Since we need to wait on the process and close the utmpx record,
156  *   we're going to set the wi_ignore flag, so that when the process exits we
157  *   clean up, but don't tell the restarter to stop it.
158  */
159 void
wait_ignore_by_fmri(const char * fmri)160 wait_ignore_by_fmri(const char *fmri)
161 {
162 	wait_info_t *wi;
163 
164 	MUTEX_LOCK(&wait_info_lock);
165 
166 	for (wi = uu_list_first(wait_info_list); wi != NULL;
167 	    wi = uu_list_next(wait_info_list, wi)) {
168 		if (strcmp(wi->wi_fmri, fmri) == 0)
169 			break;
170 	}
171 
172 	if (wi != NULL) {
173 		wi->wi_ignore = 1;
174 	}
175 
176 	MUTEX_UNLOCK(&wait_info_lock);
177 }
178 
179 /*
180  * int wait_register(pid_t, char *, int, int)
181  *   wait_register is called after we have called fork(2), and know which pid we
182  *   wish to monitor.  However, since the child may have already exited by the
183  *   time we are called, we must handle the error cases from open(2)
184  *   appropriately.  The am_parent flag is recorded to handle waitpid(2)
185  *   behaviour on removal; similarly, the direct flag is passed through to a
186  *   potential call to wait_remove() to govern its behaviour in different
187  *   contexts.
188  *
189  *   Returns 0 if registration successful, 1 if child pid did not exist, and -1
190  *   if a different error occurred.
191  */
192 int
wait_register(pid_t pid,const char * inst_fmri,int am_parent,int direct)193 wait_register(pid_t pid, const char *inst_fmri, int am_parent, int direct)
194 {
195 	char *fname = uu_msprintf("/proc/%ld/psinfo", pid);
196 	int fd;
197 	wait_info_t *wi;
198 
199 	assert(pid != 0);
200 
201 	if (fname == NULL)
202 		return (-1);
203 
204 	wi = startd_alloc(sizeof (wait_info_t));
205 
206 	uu_list_node_init(wi, &wi->wi_link, wait_info_pool);
207 
208 	wi->wi_fd = -1;
209 	wi->wi_pid = pid;
210 	wi->wi_fmri = inst_fmri;
211 	wi->wi_parent = am_parent;
212 	wi->wi_ignore = 0;
213 
214 	MUTEX_LOCK(&wait_info_lock);
215 	(void) uu_list_insert_before(wait_info_list, NULL, wi);
216 	MUTEX_UNLOCK(&wait_info_lock);
217 
218 	if ((fd = open(fname, O_RDONLY)) == -1) {
219 		if (errno == ENOENT) {
220 			/*
221 			 * Child has already exited.
222 			 */
223 			wait_remove(wi, direct);
224 			uu_free(fname);
225 			return (1);
226 		} else {
227 			log_error(LOG_WARNING,
228 			    "open %s failed; not monitoring %s: %s\n", fname,
229 			    inst_fmri, strerror(errno));
230 			uu_free(fname);
231 			return (-1);
232 		}
233 	}
234 
235 	uu_free(fname);
236 
237 	wi->wi_fd = fd;
238 
239 	if (port_associate(port_fd, PORT_SOURCE_FD, fd, 0, wi)) {
240 		log_error(LOG_WARNING,
241 		    "initial port_association of %d / %s failed: %s\n", fd,
242 		    inst_fmri, strerror(errno));
243 		return (-1);
244 	}
245 
246 	log_framework(LOG_DEBUG, "monitoring PID %ld on fd %d (%s)\n", pid, fd,
247 	    inst_fmri);
248 
249 	return (0);
250 }
251 
252 /*ARGSUSED*/
253 void *
wait_thread(void * args)254 wait_thread(void *args)
255 {
256 	(void) pthread_setname_np(pthread_self(), "wait");
257 
258 	for (;;) {
259 		port_event_t pe;
260 		int fd;
261 		wait_info_t *wi;
262 
263 		if (port_get(port_fd, &pe, NULL) != 0) {
264 			if (errno == EINTR)
265 				continue;
266 			else {
267 				log_error(LOG_WARNING,
268 				    "port_get() failed with %s\n",
269 				    strerror(errno));
270 				bad_error("port_get", errno);
271 			}
272 		}
273 
274 		fd = pe.portev_object;
275 		wi = pe.portev_user;
276 		assert(wi != NULL);
277 		assert(fd == wi->wi_fd);
278 
279 		if ((pe.portev_events & POLLHUP) == POLLHUP) {
280 			psinfo_t psi;
281 
282 			if (lseek(fd, 0, SEEK_SET) != 0 ||
283 			    read(fd, &psi, sizeof (psinfo_t)) !=
284 			    sizeof (psinfo_t)) {
285 				log_framework(LOG_WARNING,
286 				    "couldn't get psinfo data for %s (%s); "
287 				    "assuming failed\n", wi->wi_fmri,
288 				    strerror(errno));
289 				goto err_remove;
290 			}
291 
292 			if (psi.pr_nlwp != 0 ||
293 			    psi.pr_nzomb != 0 ||
294 			    psi.pr_lwp.pr_lwpid != 0) {
295 				/*
296 				 * We have determined, in accordance with the
297 				 * definition in proc(5), this process is not a
298 				 * zombie.  Reassociate.
299 				 */
300 				if (port_associate(port_fd, PORT_SOURCE_FD, fd,
301 				    0, wi))
302 					log_error(LOG_WARNING,
303 					    "port_association of %d / %s "
304 					    "failed\n", fd, wi->wi_fmri);
305 				continue;
306 			}
307 		} else if (
308 		    (pe.portev_events & POLLERR) == 0) {
309 			if (port_associate(port_fd, PORT_SOURCE_FD, fd, 0, wi))
310 				log_error(LOG_WARNING,
311 				    "port_association of %d / %s "
312 				    "failed\n", fd, wi->wi_fmri);
313 			continue;
314 		}
315 
316 err_remove:
317 		wait_remove(wi, 0);
318 	}
319 
320 	/*LINTED E_FUNC_HAS_NO_RETURN_STMT*/
321 }
322 
323 void
wait_prefork()324 wait_prefork()
325 {
326 	MUTEX_LOCK(&wait_info_lock);
327 }
328 
329 void
wait_postfork(pid_t pid)330 wait_postfork(pid_t pid)
331 {
332 	wait_info_t *wi;
333 
334 	MUTEX_UNLOCK(&wait_info_lock);
335 
336 	if (pid != 0)
337 		return;
338 
339 	/*
340 	 * Close all of the child's wait-related fds.  The wait_thread() is
341 	 * gone, so no need to worry about returning events.  We always exec(2)
342 	 * after a fork request, so we needn't free the list elements
343 	 * themselves.
344 	 */
345 
346 	for (wi = uu_list_first(wait_info_list);
347 	    wi != NULL;
348 	    wi = uu_list_next(wait_info_list, wi)) {
349 		if (wi->wi_fd != -1)
350 			startd_close(wi->wi_fd);
351 	}
352 
353 	startd_close(port_fd);
354 
355 	(void) setrlimit(RLIMIT_NOFILE, &init_fd_rlimit);
356 }
357 
358 void
wait_init()359 wait_init()
360 {
361 	struct rlimit fd_new;
362 
363 	(void) getrlimit(RLIMIT_NOFILE, &init_fd_rlimit);
364 	(void) getrlimit(RLIMIT_NOFILE, &fd_new);
365 
366 	fd_new.rlim_max = fd_new.rlim_cur = WAIT_FILES;
367 
368 	(void) setrlimit(RLIMIT_NOFILE, &fd_new);
369 
370 	if ((port_fd = port_create()) == -1)
371 		uu_die("wait_init couldn't port_create");
372 
373 	wait_info_pool = uu_list_pool_create("wait_info", sizeof (wait_info_t),
374 	    offsetof(wait_info_t, wi_link), NULL, UU_LIST_POOL_DEBUG);
375 	if (wait_info_pool == NULL)
376 		uu_die("wait_init couldn't create wait_info_pool");
377 
378 	wait_info_list = uu_list_create(wait_info_pool, wait_info_list, 0);
379 	if (wait_info_list == NULL)
380 		uu_die("wait_init couldn't create wait_info_list");
381 
382 	(void) pthread_mutex_init(&wait_info_lock, &mutex_attrs);
383 }
384