xref: /illumos-gate/usr/src/cmd/svc/startd/wait.c (revision 032624d56c174c5c55126582b32e314a6af15522)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * wait.c - asynchronous monitoring of "wait registered" start methods
31  *
32  * Use event ports to poll on the set of fds representing the /proc/[pid]/psinfo
33  * files.  If one of these fds returns an event, then we inform the restarter
34  * that it has stopped.
35  *
36  * The wait_info_list holds the series of processes currently being monitored
37  * for exit.  The wi_fd member, which contains the file descriptor of the psinfo
38  * file being polled upon ("event ported upon"), will be set to -1 if the file
39  * descriptor is inactive (already closed or not yet opened).
40  */
41 
42 #ifdef _FILE_OFFSET_BITS
43 #undef _FILE_OFFSET_BITS
44 #endif /* _FILE_OFFSET_BITS */
45 
46 #include <sys/resource.h>
47 #include <sys/stat.h>
48 #include <sys/types.h>
49 #include <sys/uio.h>
50 #include <sys/wait.h>
51 
52 #include <assert.h>
53 #include <errno.h>
54 #include <fcntl.h>
55 #include <libuutil.h>
56 #include <poll.h>
57 #include <port.h>
58 #include <pthread.h>
59 #include <procfs.h>
60 #include <string.h>
61 #include <stropts.h>
62 #include <unistd.h>
63 
64 #include "startd.h"
65 
66 #define	WAIT_FILES	262144		/* reasonably high maximum */
67 
68 static int port_fd;
69 static scf_handle_t *wait_hndl;
70 static struct rlimit init_fd_rlimit;
71 
72 static uu_list_pool_t *wait_info_pool;
73 static uu_list_t *wait_info_list;
74 
75 static pthread_mutex_t wait_info_lock;
76 
77 /*
78  * void wait_remove(wait_info_t *, int)
79  *   Remove the given wait_info structure from our list, performing various
80  *   cleanup operations along the way.  If the direct flag is false (meaning
81  *   that we are being called with from restarter instance list context), then
82  *   notify the restarter that the associated instance has exited.
83  *
84  *   Since we may no longer be the startd that started this process, we only are
85  *   concerned with a waitpid(3C) failure if the wi_parent field is non-zero.
86  */
87 static void
88 wait_remove(wait_info_t *wi, int direct)
89 {
90 	int status;
91 
92 	if (waitpid(wi->wi_pid, &status, 0) == -1) {
93 		if (wi->wi_parent)
94 			log_framework(LOG_INFO,
95 			    "instance %s waitpid failure: %s\n", wi->wi_fmri,
96 			    strerror(errno));
97 	} else {
98 		if (WEXITSTATUS(status) != 0) {
99 			log_framework(LOG_NOTICE,
100 			    "instance %s exited with status %d\n", wi->wi_fmri,
101 			    WEXITSTATUS(status));
102 		}
103 	}
104 
105 	MUTEX_LOCK(&wait_info_lock);
106 	if (wi->wi_fd != -1) {
107 		startd_close(wi->wi_fd);
108 		wi->wi_fd = -1;
109 	}
110 	uu_list_remove(wait_info_list, wi);
111 	MUTEX_UNLOCK(&wait_info_lock);
112 
113 	/*
114 	 * Make an attempt to clear out any utmpx record associated with this
115 	 * PID.
116 	 */
117 	utmpx_mark_dead(wi->wi_pid, status, B_FALSE);
118 
119 	if (!direct) {
120 		/*
121 		 * Bind wait_hndl lazily.
122 		 */
123 		if (wait_hndl == NULL) {
124 			for (wait_hndl =
125 			    libscf_handle_create_bound(SCF_VERSION);
126 			    wait_hndl == NULL;
127 			    wait_hndl =
128 			    libscf_handle_create_bound(SCF_VERSION)) {
129 				log_error(LOG_INFO, "[wait_remove] Unable to "
130 				    "bind a new repository handle: %s\n",
131 				    scf_strerror(scf_error()));
132 				(void) sleep(2);
133 			}
134 		}
135 
136 		log_framework(LOG_DEBUG,
137 		    "wait_remove requesting stop of %s\n", wi->wi_fmri);
138 		(void) stop_instance_fmri(wait_hndl, wi->wi_fmri, RSTOP_EXIT);
139 	}
140 
141 	uu_list_node_fini(wi, &wi->wi_link, wait_info_pool);
142 	startd_free(wi, sizeof (wait_info_t));
143 }
144 
145 /*
146  * int wait_register(pid_t, char *, int, int)
147  *   wait_register is called after we have called fork(2), and know which pid we
148  *   wish to monitor.  However, since the child may have already exited by the
149  *   time we are called, we must handle the error cases from open(2)
150  *   appropriately.  The am_parent flag is recorded to handle waitpid(2)
151  *   behaviour on removal; similarly, the direct flag is passed through to a
152  *   potential call to wait_remove() to govern its behaviour in different
153  *   contexts.
154  *
155  *   Returns 0 if registration successful, 1 if child pid did not exist, and -1
156  *   if a different error occurred.
157  */
158 int
159 wait_register(pid_t pid, const char *inst_fmri, int am_parent, int direct)
160 {
161 	char *fname = uu_msprintf("/proc/%ld/psinfo", pid);
162 	int fd;
163 	wait_info_t *wi;
164 
165 	assert(pid != 0);
166 
167 	if (fname == NULL)
168 		return (-1);
169 
170 	wi = startd_alloc(sizeof (wait_info_t));
171 
172 	uu_list_node_init(wi, &wi->wi_link, wait_info_pool);
173 
174 	wi->wi_fd = -1;
175 	wi->wi_pid = pid;
176 	wi->wi_fmri = inst_fmri;
177 	wi->wi_parent = am_parent;
178 
179 	MUTEX_LOCK(&wait_info_lock);
180 	(void) uu_list_insert_before(wait_info_list, NULL, wi);
181 	MUTEX_UNLOCK(&wait_info_lock);
182 
183 	if ((fd = open(fname, O_RDONLY)) == -1) {
184 		if (errno == ENOENT) {
185 			/*
186 			 * Child has already exited.
187 			 */
188 			wait_remove(wi, direct);
189 			uu_free(fname);
190 			return (1);
191 		} else {
192 			log_error(LOG_WARNING,
193 			    "open %s failed; not monitoring %s: %s\n", fname,
194 			    inst_fmri, strerror(errno));
195 			uu_free(fname);
196 			return (-1);
197 		}
198 	}
199 
200 	uu_free(fname);
201 
202 	wi->wi_fd = fd;
203 
204 	if (port_associate(port_fd, PORT_SOURCE_FD, fd, 0, wi)) {
205 		log_error(LOG_WARNING,
206 		    "initial port_association of %d / %s failed: %s\n", fd,
207 		    inst_fmri, strerror(errno));
208 		return (-1);
209 	}
210 
211 	log_framework(LOG_DEBUG, "monitoring PID %ld on fd %d (%s)\n", pid, fd,
212 	    inst_fmri);
213 
214 	return (0);
215 }
216 
217 /*ARGSUSED*/
218 void *
219 wait_thread(void *args)
220 {
221 	for (;;) {
222 		port_event_t pe;
223 		int fd;
224 		wait_info_t *wi;
225 
226 		if (port_get(port_fd, &pe, NULL) != 0) {
227 			if (errno == EINTR)
228 				continue;
229 			else {
230 				log_error(LOG_WARNING,
231 				    "port_get() failed with %s\n",
232 				    strerror(errno));
233 				bad_error("port_get", errno);
234 			}
235 		}
236 
237 		fd = pe.portev_object;
238 		wi = pe.portev_user;
239 		assert(wi != NULL);
240 		assert(fd == wi->wi_fd);
241 
242 		if ((pe.portev_events & POLLHUP) == POLLHUP) {
243 			psinfo_t psi;
244 
245 			if (lseek(fd, 0, SEEK_SET) != 0 ||
246 			    read(fd, &psi, sizeof (psinfo_t)) !=
247 			    sizeof (psinfo_t)) {
248 				log_framework(LOG_WARNING,
249 				    "couldn't get psinfo data for %s (%s); "
250 				    "assuming failed\n", wi->wi_fmri,
251 				    strerror(errno));
252 				    goto err_remove;
253 			}
254 
255 			if (psi.pr_nlwp != 0 ||
256 			    psi.pr_nzomb != 0 ||
257 			    psi.pr_lwp.pr_lwpid != 0) {
258 				/*
259 				 * We have determined, in accordance with the
260 				 * definition in proc(4), this process is not a
261 				 * zombie.  Reassociate.
262 				 */
263 				if (port_associate(port_fd, PORT_SOURCE_FD, fd,
264 					0, wi))
265 					log_error(LOG_WARNING,
266 					    "port_association of %d / %s "
267 					    "failed\n", fd, wi->wi_fmri);
268 				continue;
269 			}
270 		} else if (
271 		    (pe.portev_events & POLLERR) == 0) {
272 			if (port_associate(port_fd, PORT_SOURCE_FD, fd, 0, wi))
273 				log_error(LOG_WARNING,
274 				    "port_association of %d / %s "
275 				    "failed\n", fd, wi->wi_fmri);
276 			continue;
277 		}
278 
279 err_remove:
280 		wait_remove(wi, 0);
281 	}
282 
283 	/*LINTED E_FUNC_HAS_NO_RETURN_STMT*/
284 }
285 
286 void
287 wait_prefork()
288 {
289 	MUTEX_LOCK(&wait_info_lock);
290 }
291 
292 void
293 wait_postfork(pid_t pid)
294 {
295 	wait_info_t *wi;
296 
297 	MUTEX_UNLOCK(&wait_info_lock);
298 
299 	if (pid != 0)
300 		return;
301 
302 	/*
303 	 * Close all of the child's wait-related fds.  The wait_thread() is
304 	 * gone, so no need to worry about returning events.  We always exec(2)
305 	 * after a fork request, so we needn't free the list elements
306 	 * themselves.
307 	 */
308 
309 	for (wi = uu_list_first(wait_info_list);
310 	    wi != NULL;
311 	    wi = uu_list_next(wait_info_list, wi)) {
312 		if (wi->wi_fd != -1)
313 			startd_close(wi->wi_fd);
314 	}
315 
316 	startd_close(port_fd);
317 
318 	(void) setrlimit(RLIMIT_NOFILE, &init_fd_rlimit);
319 }
320 
321 void
322 wait_init()
323 {
324 	struct rlimit fd_new;
325 
326 	(void) getrlimit(RLIMIT_NOFILE, &init_fd_rlimit);
327 	(void) getrlimit(RLIMIT_NOFILE, &fd_new);
328 
329 	fd_new.rlim_max = fd_new.rlim_cur = WAIT_FILES;
330 
331 	(void) setrlimit(RLIMIT_NOFILE, &fd_new);
332 
333 	if ((port_fd = port_create()) == -1)
334 		uu_die("wait_init couldn't port_create");
335 
336 	wait_info_pool = uu_list_pool_create("wait_info", sizeof (wait_info_t),
337 	    offsetof(wait_info_t, wi_link), NULL, UU_LIST_POOL_DEBUG);
338 	if (wait_info_pool == NULL)
339 		uu_die("wait_init couldn't create wait_info_pool");
340 
341 	wait_info_list = uu_list_create(wait_info_pool, wait_info_list, 0);
342 	if (wait_info_list == NULL)
343 		uu_die("wait_init couldn't create wait_info_list");
344 
345 	(void) pthread_mutex_init(&wait_info_lock, &mutex_attrs);
346 }
347