xref: /titanic_44/usr/src/cmd/svc/startd/wait.c (revision 8eea8e29cc4374d1ee24c25a07f45af132db3499)
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 2004 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 	uu_list_remove(wait_info_list, wi);
107 	MUTEX_UNLOCK(&wait_info_lock);
108 
109 	/*
110 	 * Make an attempt to clear out any utmpx record associated with this
111 	 * PID.
112 	 */
113 	utmpx_mark_dead(wi->wi_pid, status, B_FALSE);
114 
115 	if (!direct) {
116 		/*
117 		 * Bind wait_hndl lazily.
118 		 */
119 		if (wait_hndl == NULL) {
120 			for (wait_hndl =
121 			    libscf_handle_create_bound(SCF_VERSION);
122 			    wait_hndl == NULL;
123 			    wait_hndl =
124 			    libscf_handle_create_bound(SCF_VERSION)) {
125 				log_error(LOG_INFO, "[wait_remove] Unable to "
126 				    "bind a new repository handle: %s\n",
127 				    scf_strerror(scf_error()));
128 				(void) sleep(2);
129 			}
130 		}
131 
132 		log_framework(LOG_DEBUG,
133 		    "wait_remove requesting stop of %s\n", wi->wi_fmri);
134 		(void) stop_instance_fmri(wait_hndl, wi->wi_fmri, RSTOP_EXIT);
135 	}
136 
137 	uu_list_node_fini(wi, &wi->wi_link, wait_info_pool);
138 	startd_free(wi, sizeof (wait_info_t));
139 }
140 
141 /*
142  * int wait_register(pid_t, char *, int, int)
143  *   wait_register is called after we have called fork(2), and know which pid we
144  *   wish to monitor.  However, since the child may have already exited by the
145  *   time we are called, we must handle the error cases from open(2)
146  *   appropriately.  The am_parent flag is recorded to handle waitpid(2)
147  *   behaviour on removal; similarly, the direct flag is passed through to a
148  *   potential call to wait_remove() to govern its behaviour in different
149  *   contexts.
150  *
151  *   Returns 0 if registration successful, 1 if child pid did not exist, and -1
152  *   if a different error occurred.
153  */
154 int
155 wait_register(pid_t pid, const char *inst_fmri, int am_parent, int direct)
156 {
157 	char *fname = uu_msprintf("/proc/%ld/psinfo", pid);
158 	int fd;
159 	wait_info_t *wi;
160 
161 	assert(pid != 0);
162 
163 	if (fname == NULL)
164 		return (-1);
165 
166 	wi = startd_alloc(sizeof (wait_info_t));
167 
168 	uu_list_node_init(wi, &wi->wi_link, wait_info_pool);
169 
170 	wi->wi_fd = -1;
171 	wi->wi_pid = pid;
172 	wi->wi_fmri = inst_fmri;
173 	wi->wi_parent = am_parent;
174 
175 	MUTEX_LOCK(&wait_info_lock);
176 	(void) uu_list_insert_before(wait_info_list, NULL, wi);
177 	MUTEX_UNLOCK(&wait_info_lock);
178 
179 	if ((fd = open(fname, O_RDONLY)) == -1) {
180 		if (errno == ENOENT) {
181 			/*
182 			 * Child has already exited.
183 			 */
184 			wait_remove(wi, direct);
185 			uu_free(fname);
186 			return (1);
187 		} else {
188 			log_error(LOG_WARNING,
189 			    "open %s failed; not monitoring %s: %s\n", fname,
190 			    inst_fmri, strerror(errno));
191 			uu_free(fname);
192 			return (-1);
193 		}
194 	}
195 
196 	uu_free(fname);
197 
198 	wi->wi_fd = fd;
199 
200 	if (port_associate(port_fd, PORT_SOURCE_FD, fd, 0, wi)) {
201 		log_error(LOG_WARNING,
202 		    "initial port_association of %d / %s failed: %s\n", fd,
203 		    inst_fmri, strerror(errno));
204 		return (-1);
205 	}
206 
207 	log_framework(LOG_DEBUG, "monitoring PID %ld on fd %d (%s)\n", pid, fd,
208 	    inst_fmri);
209 
210 	return (0);
211 }
212 
213 /*ARGSUSED*/
214 void *
215 wait_thread(void *args)
216 {
217 	for (;;) {
218 		port_event_t pe;
219 		int fd;
220 		wait_info_t *wi;
221 		struct timespec ts;
222 
223 		ts.tv_sec = 1;
224 		ts.tv_nsec = 0;
225 
226 		if (port_get(port_fd, &pe, &ts) == -1)
227 			if (errno == EINTR || errno == ETIME)
228 				continue;
229 			else
230 				log_error(LOG_WARNING,
231 				    "port_get returned %s\n", strerror(errno));
232 
233 		fd = pe.portev_object;
234 		wi = pe.portev_user;
235 
236 		if ((pe.portev_events & POLLHUP) == POLLHUP) {
237 			psinfo_t psi;
238 
239 			if (lseek(fd, 0, SEEK_SET) != 0 ||
240 			    read(fd, &psi, sizeof (psinfo_t)) !=
241 			    sizeof (psinfo_t)) {
242 				log_framework(LOG_WARNING,
243 				    "couldn't get psinfo data for %s (%s); "
244 				    "assuming failed\n", wi->wi_fmri,
245 				    strerror(errno));
246 				    goto err_remove;
247 			}
248 
249 			if (psi.pr_nlwp != 0 ||
250 			    psi.pr_nzomb != 0 ||
251 			    psi.pr_lwp.pr_lwpid != 0) {
252 				/*
253 				 * We have determined, in accordance with the
254 				 * definition in proc(4), this process is not a
255 				 * zombie.  Reassociate.
256 				 */
257 				if (port_associate(port_fd, PORT_SOURCE_FD, fd,
258 					0, wi))
259 					log_error(LOG_WARNING,
260 					    "port_association of %d / %s "
261 					    "failed\n", fd, wi->wi_fmri);
262 				continue;
263 			}
264 		} else if (
265 		    (pe.portev_events & POLLERR) == 0) {
266 			if (port_associate(port_fd, PORT_SOURCE_FD, fd, 0, wi))
267 				log_error(LOG_WARNING,
268 				    "port_association of %d / %s "
269 				    "failed\n", fd, wi->wi_fmri);
270 			continue;
271 		}
272 
273 err_remove:
274 		startd_close(fd);
275 		wi->wi_fd = -1;
276 
277 		wait_remove(wi, 0);
278 	}
279 
280 	/*LINTED E_FUNC_HAS_NO_RETURN_STMT*/
281 }
282 
283 void
284 wait_prefork()
285 {
286 	MUTEX_LOCK(&wait_info_lock);
287 }
288 
289 void
290 wait_postfork(pid_t pid)
291 {
292 	wait_info_t *wi;
293 
294 	MUTEX_UNLOCK(&wait_info_lock);
295 
296 	if (pid != 0)
297 		return;
298 
299 	/*
300 	 * Close all of the child's wait-related fds.  The wait_thread() is
301 	 * gone, so no need to worry about returning events.  We always exec(2)
302 	 * after a fork request, so we needn't free the list elements
303 	 * themselves.
304 	 */
305 
306 	for (wi = uu_list_first(wait_info_list);
307 	    wi != NULL;
308 	    wi = uu_list_next(wait_info_list, wi)) {
309 		if (wi->wi_fd != -1)
310 			startd_close(wi->wi_fd);
311 	}
312 
313 	startd_close(port_fd);
314 
315 	(void) setrlimit(RLIMIT_NOFILE, &init_fd_rlimit);
316 }
317 
318 void
319 wait_init()
320 {
321 	struct rlimit fd_new;
322 
323 	(void) getrlimit(RLIMIT_NOFILE, &init_fd_rlimit);
324 	(void) getrlimit(RLIMIT_NOFILE, &fd_new);
325 
326 	fd_new.rlim_max = fd_new.rlim_cur = WAIT_FILES;
327 
328 	(void) setrlimit(RLIMIT_NOFILE, &fd_new);
329 
330 	if ((port_fd = port_create()) == -1)
331 		uu_die("wait_init couldn't port_create");
332 
333 	wait_info_pool = uu_list_pool_create("wait_info", sizeof (wait_info_t),
334 	    offsetof(wait_info_t, wi_link), NULL, UU_LIST_POOL_DEBUG);
335 	if (wait_info_pool == NULL)
336 		uu_die("wait_init couldn't create wait_info_pool");
337 
338 	wait_info_list = uu_list_create(wait_info_pool, wait_info_list, 0);
339 	if (wait_info_list == NULL)
340 		uu_die("wait_init couldn't create wait_info_list");
341 
342 	(void) pthread_mutex_init(&wait_info_lock, &mutex_attrs);
343 }
344