xref: /illumos-gate/usr/src/cmd/zoneadmd/zcons.c (revision 8119dad84d6416f13557b0ba8e2aaf9064cbcfd3)
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 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  * Copyright 2012 Joyent, Inc.  All rights reserved.
26  * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
27  */
28 
29 /*
30  * Console support for zones requires a significant infrastructure.  The
31  * core pieces are contained in this file, but other portions of note
32  * are in the zlogin(1) command, the zcons(4D) driver, and in the
33  * devfsadm(8) misc_link generator.
34  *
35  * Care is taken to make the console behave in an "intuitive" fashion for
36  * administrators.  Essentially, we try as much as possible to mimic the
37  * experience of using a system via a tip line and system controller.
38  *
39  * The zone console architecture looks like this:
40  *
41  *                                      Global Zone | Non-Global Zone
42  *                        .--------------.          |
43  *        .-----------.   | zoneadmd -z  |          | .--------. .---------.
44  *        | zlogin -C |   |     myzone   |          | | ttymon | | syslogd |
45  *        `-----------'   `--------------'          | `--------' `---------'
46  *                  |       |       | |             |      |       |
47  *  User            |       |       | |             |      V       V
48  * - - - - - - - - -|- - - -|- - - -|-|- - - - - - -|- - /dev/zconsole - - -
49  *  Kernel          V       V       | |                        |
50  *               [AF_UNIX Socket]   | `--------. .-------------'
51  *                                  |          | |
52  *                                  |          V V
53  *                                  |     +-----------+
54  *                                  |     |  ldterm,  |
55  *                                  |     |   etc.    |
56  *                                  |     +-----------+
57  *                                  |     +-[Anchor]--+
58  *                                  |     |   ptem    |
59  *                                  V     +-----------+
60  *                           +---manager--+-subsidiary+
61  *                           |                        |
62  *                           |      zcons driver      |
63  *                           |    zonename="myzone"   |
64  *                           +------------------------+
65  *
66  * There are basically two major tasks which the console subsystem in
67  * zoneadmd accomplishes:
68  *
69  * - Setup and teardown of zcons driver instances.  One zcons instance
70  *   is maintained per zone; we take advantage of the libdevice APIs
71  *   to online new instances of zcons as needed.  Care is taken to
72  *   prune and manage these appropriately; see init_console_dev() and
73  *   destroy_console_dev().  The end result is the creation of the
74  *   zcons(4D) instance and an open file descriptor to the manager side.
75  *   zcons instances are associated with zones via their zonename device
76  *   property.  This the console instance to persist across reboots,
77  *   and while the zone is halted.
78  *
79  * - Acting as a server for 'zlogin -C' instances.  When zlogin -C is
80  *   run, zlogin connects to zoneadmd via unix domain socket.  zoneadmd
81  *   functions as a two-way proxy for console I/O, relaying user input
82  *   to the manager side of the console, and relaying output from the
83  *   zone to the user.
84  */
85 
86 #include <sys/types.h>
87 #include <sys/socket.h>
88 #include <sys/stat.h>
89 #include <sys/termios.h>
90 #include <sys/zcons.h>
91 #include <sys/mkdev.h>
92 
93 #include <assert.h>
94 #include <ctype.h>
95 #include <errno.h>
96 #include <fcntl.h>
97 #include <stdarg.h>
98 #include <stdio.h>
99 #include <stdlib.h>
100 #include <strings.h>
101 #include <stropts.h>
102 #include <thread.h>
103 #include <ucred.h>
104 #include <unistd.h>
105 #include <zone.h>
106 
107 #include <libdevinfo.h>
108 #include <libdevice.h>
109 #include <libzonecfg.h>
110 
111 #include <syslog.h>
112 #include <sys/modctl.h>
113 
114 #include "zoneadmd.h"
115 
116 #define	ZCONSNEX_DEVTREEPATH	"/pseudo/zconsnex@1"
117 #define	ZCONSNEX_FILEPATH	"/devices/pseudo/zconsnex@1"
118 
119 #define	CONSOLE_SOCKPATH	ZONES_TMPDIR "/%s.console_sock"
120 
121 static int	serverfd = -1;	/* console server unix domain socket fd */
122 char boot_args[BOOTARGS_MAX];
123 char bad_boot_arg[BOOTARGS_MAX];
124 
125 /*
126  * The eventstream is a simple one-directional flow of messages from the
127  * door server to the console subsystem, implemented with a pipe.
128  * It is used to wake up the console poller when it needs to take action,
129  * message the user, die off, etc.
130  */
131 static int eventstream[2];
132 
133 
134 
135 int
136 eventstream_init()
137 {
138 	if (pipe(eventstream) == -1)
139 		return (-1);
140 	return (0);
141 }
142 
143 void
144 eventstream_write(zone_evt_t evt)
145 {
146 	(void) write(eventstream[0], &evt, sizeof (evt));
147 }
148 
149 static zone_evt_t
150 eventstream_read(void)
151 {
152 	zone_evt_t evt = Z_EVT_NULL;
153 
154 	(void) read(eventstream[1], &evt, sizeof (evt));
155 	return (evt);
156 }
157 
158 /*
159  * count_console_devs() and its helper count_cb() do a walk of the
160  * subtree of the device tree where zone console nodes are represented.
161  * The goal is to count zone console instances already setup for a zone
162  * with the given name.  More than 1 is anomolous, and our caller will
163  * have to deal with that if we find that's the case.
164  *
165  * Note: this algorithm is a linear search of nodes in the zconsnex subtree
166  * of the device tree, and could be a scalability problem, but I don't see
167  * how to avoid it.
168  */
169 
170 /*
171  * cb_data is shared by count_cb and destroy_cb for simplicity.
172  */
173 struct cb_data {
174 	zlog_t *zlogp;
175 	int found;
176 	int killed;
177 };
178 
179 static int
180 count_cb(di_node_t node, void *arg)
181 {
182 	struct cb_data *cb = (struct cb_data *)arg;
183 	char *prop_data;
184 
185 	if (di_prop_lookup_strings(DDI_DEV_T_ANY, node, "zonename",
186 	    &prop_data) != -1) {
187 		assert(prop_data != NULL);
188 		if (strcmp(prop_data, zone_name) == 0) {
189 			cb->found++;
190 			return (DI_WALK_CONTINUE);
191 		}
192 	}
193 	return (DI_WALK_CONTINUE);
194 }
195 
196 static int
197 count_console_devs(zlog_t *zlogp)
198 {
199 	di_node_t root;
200 	struct cb_data cb;
201 
202 	bzero(&cb, sizeof (cb));
203 	cb.zlogp = zlogp;
204 
205 	if ((root = di_init(ZCONSNEX_DEVTREEPATH, DINFOCPYALL)) ==
206 	    DI_NODE_NIL) {
207 		zerror(zlogp, B_TRUE, "%s failed", "di_init");
208 		return (-1);
209 	}
210 
211 	(void) di_walk_node(root, DI_WALK_CLDFIRST, (void *)&cb, count_cb);
212 	di_fini(root);
213 	return (cb.found);
214 }
215 
216 /*
217  * destroy_console_devs() and its helper destroy_cb() tears down any console
218  * instances associated with this zone.  If things went very wrong, we
219  * might have more than one console instance hanging around.  This routine
220  * hunts down and tries to remove all of them.  Of course, if the console
221  * is open, the instance will not detach, which is a potential issue.
222  */
223 static int
224 destroy_cb(di_node_t node, void *arg)
225 {
226 	struct cb_data *cb = (struct cb_data *)arg;
227 	char *prop_data;
228 	char *tmp;
229 	char devpath[MAXPATHLEN];
230 	devctl_hdl_t hdl;
231 
232 	if (di_prop_lookup_strings(DDI_DEV_T_ANY, node, "zonename",
233 	    &prop_data) == -1)
234 		return (DI_WALK_CONTINUE);
235 
236 	assert(prop_data != NULL);
237 	if (strcmp(prop_data, zone_name) != 0) {
238 		/* this is the console for a different zone */
239 		return (DI_WALK_CONTINUE);
240 	}
241 
242 	cb->found++;
243 	tmp = di_devfs_path(node);
244 	(void) snprintf(devpath, sizeof (devpath), "/devices/%s", tmp);
245 	di_devfs_path_free(tmp);
246 
247 	if ((hdl = devctl_device_acquire(devpath, 0)) == NULL) {
248 		zerror(cb->zlogp, B_TRUE, "WARNING: console %s found, "
249 		    "but it could not be controlled.", devpath);
250 		return (DI_WALK_CONTINUE);
251 	}
252 	if (devctl_device_remove(hdl) == 0) {
253 		cb->killed++;
254 	} else {
255 		zerror(cb->zlogp, B_TRUE, "WARNING: console %s found, "
256 		    "but it could not be removed.", devpath);
257 	}
258 	devctl_release(hdl);
259 	return (DI_WALK_CONTINUE);
260 }
261 
262 static int
263 destroy_console_devs(zlog_t *zlogp)
264 {
265 	char conspath[MAXPATHLEN];
266 	di_node_t root;
267 	struct cb_data cb;
268 	int managerfd;
269 	int subfd;
270 
271 	/*
272 	 * Signal the manager side to release its handle on the subsidiary side
273 	 * by issuing a ZC_RELEASESUBSID ioctl.
274 	 */
275 	(void) snprintf(conspath, sizeof (conspath), "/dev/zcons/%s/%s",
276 	    zone_name, ZCONS_MANAGER_NAME);
277 	if ((managerfd = open(conspath, O_RDWR | O_NOCTTY)) != -1) {
278 		(void) snprintf(conspath, sizeof (conspath), "/dev/zcons/%s/%s",
279 		    zone_name, ZCONS_SUBSIDIARY_NAME);
280 		if ((subfd = open(conspath, O_RDWR | O_NOCTTY)) != -1) {
281 			if (ioctl(managerfd, ZC_RELEASESUBSID,
282 			    (caddr_t)(intptr_t)subfd) != 0)
283 				zerror(zlogp, B_TRUE, "WARNING: error while "
284 				    "releasing subsidiary handle of zone "
285 				    "console for %s", zone_name);
286 			(void) close(subfd);
287 		} else {
288 			zerror(zlogp, B_TRUE, "WARNING: could not open "
289 			    "subsidiary side of zone console for %s to "
290 			    "release subsidiary handle", zone_name);
291 		}
292 		(void) close(managerfd);
293 	} else {
294 		zerror(zlogp, B_TRUE, "WARNING: could not open manager side of "
295 		    "zone console for %s to release subsidiary handle",
296 		    zone_name);
297 	}
298 
299 	bzero(&cb, sizeof (cb));
300 	cb.zlogp = zlogp;
301 
302 	if ((root = di_init(ZCONSNEX_DEVTREEPATH, DINFOCPYALL)) ==
303 	    DI_NODE_NIL) {
304 		zerror(zlogp, B_TRUE, "%s failed", "di_init");
305 		return (-1);
306 	}
307 
308 	(void) di_walk_node(root, DI_WALK_CLDFIRST, (void *)&cb, destroy_cb);
309 	if (cb.found > 1) {
310 		zerror(zlogp, B_FALSE, "WARNING: multiple zone console "
311 		    "instances detected for zone '%s'; %d of %d "
312 		    "successfully removed.",
313 		    zone_name, cb.killed, cb.found);
314 	}
315 
316 	di_fini(root);
317 	return (0);
318 }
319 
320 /*
321  * init_console_dev() drives the device-tree configuration of the zone
322  * console device.  The general strategy is to use the libdevice (devctl)
323  * interfaces to instantiate a new zone console node.  We do a lot of
324  * sanity checking, and are careful to reuse a console if one exists.
325  *
326  * Once the device is in the device tree, we kick devfsadm via di_init_devs()
327  * to ensure that the appropriate symlinks (to the manager and subsidiary
328  * console devices) are placed in /dev in the global zone.
329  */
330 static int
331 init_console_dev(zlog_t *zlogp)
332 {
333 	char conspath[MAXPATHLEN];
334 	devctl_hdl_t bus_hdl = NULL;
335 	devctl_hdl_t dev_hdl = NULL;
336 	devctl_ddef_t ddef_hdl = NULL;
337 	di_devlink_handle_t dl = NULL;
338 	int rv = -1;
339 	int ndevs;
340 	int managerfd;
341 	int subfd;
342 	int i;
343 
344 	/*
345 	 * Don't re-setup console if it is working and ready already; just
346 	 * skip ahead to making devlinks, which we do for sanity's sake.
347 	 */
348 	ndevs = count_console_devs(zlogp);
349 	if (ndevs == 1) {
350 		goto devlinks;
351 	} else if (ndevs > 1 || ndevs == -1) {
352 		/*
353 		 * For now, this seems like a reasonable but harsh punishment.
354 		 * If needed, we could try to get clever and delete all but
355 		 * the console which is pointed at by the current symlink.
356 		 */
357 		if (destroy_console_devs(zlogp) == -1) {
358 			goto error;
359 		}
360 	}
361 
362 	/*
363 	 * Time to make the consoles!
364 	 */
365 	if ((bus_hdl = devctl_bus_acquire(ZCONSNEX_FILEPATH, 0)) == NULL) {
366 		zerror(zlogp, B_TRUE, "%s failed", "devctl_bus_acquire");
367 		goto error;
368 	}
369 	if ((ddef_hdl = devctl_ddef_alloc("zcons", 0)) == NULL) {
370 		zerror(zlogp, B_TRUE, "failed to allocate ddef handle");
371 		goto error;
372 	}
373 	/*
374 	 * Set three properties on this node; the first is the name of the
375 	 * zone; the second is a flag which lets pseudo know that it is
376 	 * OK to automatically allocate an instance # for this device;
377 	 * the third tells the device framework not to auto-detach this
378 	 * node-- we need the node to still be there when we ask devfsadmd
379 	 * to make links, and when we need to open it.
380 	 */
381 	if (devctl_ddef_string(ddef_hdl, "zonename", zone_name) == -1) {
382 		zerror(zlogp, B_TRUE, "failed to create zonename property");
383 		goto error;
384 	}
385 	if (devctl_ddef_int(ddef_hdl, "auto-assign-instance", 1) == -1) {
386 		zerror(zlogp, B_TRUE, "failed to create auto-assign-instance "
387 		    "property");
388 		goto error;
389 	}
390 	if (devctl_ddef_int(ddef_hdl, "ddi-no-autodetach", 1) == -1) {
391 		zerror(zlogp, B_TRUE, "failed to create ddi-no-auto-detach "
392 		    "property");
393 		goto error;
394 	}
395 	if (devctl_bus_dev_create(bus_hdl, ddef_hdl, 0, &dev_hdl) == -1) {
396 		zerror(zlogp, B_TRUE, "failed to create console node");
397 		goto error;
398 	}
399 
400 devlinks:
401 	if ((dl = di_devlink_init("zcons", DI_MAKE_LINK)) != NULL) {
402 		(void) di_devlink_fini(&dl);
403 	} else {
404 		zerror(zlogp, B_TRUE, "failed to create devlinks");
405 		goto error;
406 	}
407 
408 	/*
409 	 * Open the manager side of the console and issue the ZC_HOLDSUBSID
410 	 * ioctl, which will cause the manager to retain a reference to the
411 	 * subsidiary.  This prevents ttymon from blowing through the
412 	 * subsidiary's STREAMS anchor.
413 	 */
414 	(void) snprintf(conspath, sizeof (conspath), "/dev/zcons/%s/%s",
415 	    zone_name, ZCONS_MANAGER_NAME);
416 	if ((managerfd = open(conspath, O_RDWR | O_NOCTTY)) == -1) {
417 		zerror(zlogp, B_TRUE, "ERROR: could not open manager side of "
418 		    "zone console for %s to acquire subsidiary handle",
419 		    zone_name);
420 		goto error;
421 	}
422 	(void) snprintf(conspath, sizeof (conspath), "/dev/zcons/%s/%s",
423 	    zone_name, ZCONS_SUBSIDIARY_NAME);
424 	if ((subfd = open(conspath, O_RDWR | O_NOCTTY)) == -1) {
425 		zerror(zlogp, B_TRUE, "ERROR: could not open subsidiary side "
426 		    "of zone console for %s to acquire subsidiary handle",
427 		    zone_name);
428 		(void) close(managerfd);
429 		goto error;
430 	}
431 	/*
432 	 * This ioctl can occasionally return ENXIO if devfs doesn't have
433 	 * everything plumbed up yet due to heavy zone startup load. Wait for
434 	 * 1 sec. and retry a few times before we fail to boot the zone.
435 	 */
436 	for (i = 0; i < 5; i++) {
437 		if (ioctl(managerfd, ZC_HOLDSUBSID, (caddr_t)(intptr_t)subfd)
438 		    == 0) {
439 			rv = 0;
440 			break;
441 		} else if (errno != ENXIO) {
442 			break;
443 		}
444 		(void) sleep(1);
445 	}
446 	if (rv != 0)
447 		zerror(zlogp, B_TRUE, "ERROR: error while acquiring "
448 		    "subsidiary handle of zone console for %s", zone_name);
449 
450 	(void) close(subfd);
451 	(void) close(managerfd);
452 
453 error:
454 	if (ddef_hdl)
455 		devctl_ddef_free(ddef_hdl);
456 	if (bus_hdl)
457 		devctl_release(bus_hdl);
458 	if (dev_hdl)
459 		devctl_release(dev_hdl);
460 	return (rv);
461 }
462 
463 static int
464 init_console_sock(zlog_t *zlogp)
465 {
466 	int servfd;
467 	struct sockaddr_un servaddr;
468 
469 	bzero(&servaddr, sizeof (servaddr));
470 	servaddr.sun_family = AF_UNIX;
471 	(void) snprintf(servaddr.sun_path, sizeof (servaddr.sun_path),
472 	    CONSOLE_SOCKPATH, zone_name);
473 
474 	if ((servfd = socket(AF_UNIX, SOCK_STREAM, 0)) == -1) {
475 		zerror(zlogp, B_TRUE, "console setup: could not create socket");
476 		return (-1);
477 	}
478 	(void) unlink(servaddr.sun_path);
479 
480 	if (bind(servfd, (struct sockaddr *)&servaddr,
481 	    sizeof (servaddr)) == -1) {
482 		zerror(zlogp, B_TRUE,
483 		    "console setup: could not bind to socket");
484 		goto out;
485 	}
486 
487 	if (listen(servfd, 4) == -1) {
488 		zerror(zlogp, B_TRUE,
489 		    "console setup: could not listen on socket");
490 		goto out;
491 	}
492 	return (servfd);
493 
494 out:
495 	(void) unlink(servaddr.sun_path);
496 	(void) close(servfd);
497 	return (-1);
498 }
499 
500 static void
501 destroy_console_sock(int servfd)
502 {
503 	char path[MAXPATHLEN];
504 
505 	(void) snprintf(path, sizeof (path), CONSOLE_SOCKPATH, zone_name);
506 	(void) unlink(path);
507 	(void) shutdown(servfd, SHUT_RDWR);
508 	(void) close(servfd);
509 }
510 
511 /*
512  * Read the "ident" string from the client's descriptor; this routine also
513  * tolerates being called with pid=NULL, for times when you want to "eat"
514  * the ident string from a client without saving it.
515  */
516 static int
517 get_client_ident(int clifd, pid_t *pid, char *locale, size_t locale_len,
518     int *disconnect)
519 {
520 	char buf[BUFSIZ], *bufp;
521 	size_t buflen = sizeof (buf);
522 	char c = '\0';
523 	int i = 0, r;
524 
525 	/* "eat up the ident string" case, for simplicity */
526 	if (pid == NULL) {
527 		assert(locale == NULL && locale_len == 0);
528 		while (read(clifd, &c, 1) == 1) {
529 			if (c == '\n')
530 				return (0);
531 		}
532 	}
533 
534 	bzero(buf, sizeof (buf));
535 	while ((buflen > 1) && (r = read(clifd, &c, 1)) == 1) {
536 		buflen--;
537 		if (c == '\n')
538 			break;
539 
540 		buf[i] = c;
541 		i++;
542 	}
543 	if (r == -1)
544 		return (-1);
545 
546 	/*
547 	 * We've filled the buffer, but still haven't seen \n.  Keep eating
548 	 * until we find it; we don't expect this to happen, but this is
549 	 * defensive.
550 	 */
551 	if (c != '\n') {
552 		while ((r = read(clifd, &c, sizeof (c))) > 0)
553 			if (c == '\n')
554 				break;
555 	}
556 
557 	/*
558 	 * Parse buffer for message of the form:
559 	 * IDENT <pid> <locale> <disconnect flag>
560 	 */
561 	bufp = buf;
562 	if (strncmp(bufp, "IDENT ", 6) != 0)
563 		return (-1);
564 	bufp += 6;
565 	errno = 0;
566 	*pid = strtoll(bufp, &bufp, 10);
567 	if (errno != 0)
568 		return (-1);
569 
570 	while (*bufp != '\0' && isspace(*bufp))
571 		bufp++;
572 	buflen = strlen(bufp) - 1;
573 	*disconnect = atoi(&bufp[buflen]);
574 	bufp[buflen - 1] = '\0';
575 	(void) strlcpy(locale, bufp, locale_len);
576 
577 	return (0);
578 }
579 
580 static int
581 accept_client(int servfd, pid_t *pid, char *locale, size_t locale_len,
582     int *disconnect)
583 {
584 	int connfd;
585 	struct sockaddr_un cliaddr;
586 	socklen_t clilen;
587 
588 	clilen = sizeof (cliaddr);
589 	connfd = accept(servfd, (struct sockaddr *)&cliaddr, &clilen);
590 	if (connfd == -1)
591 		return (-1);
592 	if (get_client_ident(connfd, pid, locale, locale_len,
593 	    disconnect) == -1) {
594 		(void) shutdown(connfd, SHUT_RDWR);
595 		(void) close(connfd);
596 		return (-1);
597 	}
598 	(void) write(connfd, "OK\n", 3);
599 	return (connfd);
600 }
601 
602 static void
603 reject_client(int servfd, pid_t clientpid)
604 {
605 	int connfd;
606 	struct sockaddr_un cliaddr;
607 	socklen_t clilen;
608 	char nak[MAXPATHLEN];
609 
610 	clilen = sizeof (cliaddr);
611 	connfd = accept(servfd, (struct sockaddr *)&cliaddr, &clilen);
612 
613 	/*
614 	 * After hear its ident string, tell client to get lost.
615 	 */
616 	if (get_client_ident(connfd, NULL, NULL, 0, NULL) == 0) {
617 		(void) snprintf(nak, sizeof (nak), "%lu\n",
618 		    clientpid);
619 		(void) write(connfd, nak, strlen(nak));
620 	}
621 	(void) shutdown(connfd, SHUT_RDWR);
622 	(void) close(connfd);
623 }
624 
625 static void
626 event_message(int clifd, char *clilocale, zone_evt_t evt, int dflag)
627 {
628 	char *str, *lstr = NULL;
629 	char lmsg[BUFSIZ];
630 	char outbuf[BUFSIZ];
631 
632 	if (clifd == -1)
633 		return;
634 
635 	switch (evt) {
636 	case Z_EVT_ZONE_BOOTING:
637 		if (*boot_args == '\0') {
638 			str = "NOTICE: Zone booting up";
639 			break;
640 		}
641 		/*LINTED*/
642 		(void) snprintf(lmsg, sizeof (lmsg), localize_msg(clilocale,
643 		    "NOTICE: Zone booting up with arguments: %s"), boot_args);
644 		lstr = lmsg;
645 		break;
646 	case Z_EVT_ZONE_READIED:
647 		str = "NOTICE: Zone readied";
648 		break;
649 	case Z_EVT_ZONE_HALTED:
650 		if (dflag)
651 			str = "NOTICE: Zone halted.  Disconnecting...";
652 		else
653 			str = "NOTICE: Zone halted";
654 		break;
655 	case Z_EVT_ZONE_REBOOTING:
656 		if (*boot_args == '\0') {
657 			str = "NOTICE: Zone rebooting";
658 			break;
659 		}
660 		/*LINTED*/
661 		(void) snprintf(lmsg, sizeof (lmsg), localize_msg(clilocale,
662 		    "NOTICE: Zone rebooting with arguments: %s"), boot_args);
663 		lstr = lmsg;
664 		break;
665 	case Z_EVT_ZONE_UNINSTALLING:
666 		str = "NOTICE: Zone is being uninstalled.  Disconnecting...";
667 		break;
668 	case Z_EVT_ZONE_BOOTFAILED:
669 		if (dflag)
670 			str = "NOTICE: Zone boot failed.  Disconnecting...";
671 		else
672 			str = "NOTICE: Zone boot failed";
673 		break;
674 	case Z_EVT_ZONE_BADARGS:
675 		/*LINTED*/
676 		(void) snprintf(lmsg, sizeof (lmsg),
677 		    localize_msg(clilocale,
678 		    "WARNING: Ignoring invalid boot arguments: %s"),
679 		    bad_boot_arg);
680 		lstr = lmsg;
681 		break;
682 	default:
683 		return;
684 	}
685 
686 	if (lstr == NULL)
687 		lstr = localize_msg(clilocale, str);
688 	(void) snprintf(outbuf, sizeof (outbuf), "\r\n[%s]\r\n", lstr);
689 	(void) write(clifd, outbuf, strlen(outbuf));
690 }
691 
692 /*
693  * Check to see if the client at the other end of the socket is still
694  * alive; we know it is not if it throws EPIPE at us when we try to write
695  * an otherwise harmless 0-length message to it.
696  */
697 static int
698 test_client(int clifd)
699 {
700 	if ((write(clifd, "", 0) == -1) && errno == EPIPE)
701 		return (-1);
702 	return (0);
703 }
704 
705 /*
706  * This routine drives the console I/O loop.  It polls for input from the
707  * manager side of the console (output to the console), and from the client
708  * (input from the console user).  Additionally, it polls on the server fd,
709  * and disconnects any clients that might try to hook up with the zone while
710  * the console is in use.
711  *
712  * When the client first calls us up, it is expected to send a line giving
713  * its "identity"; this consists of the string 'IDENT <pid> <locale>'.
714  * This is so that we can report that the console is busy along with
715  * some diagnostics about who has it busy; the locale is used so that
716  * asynchronous messages about zone state (like the NOTICE: zone halted
717  * messages) can be output in the user's locale.
718  */
719 static void
720 do_console_io(zlog_t *zlogp, int consfd, int servfd)
721 {
722 	struct pollfd pollfds[4];
723 	char ibuf[BUFSIZ];
724 	int cc, ret;
725 	int clifd = -1;
726 	int pollerr = 0;
727 	char clilocale[MAXPATHLEN];
728 	pid_t clipid = 0;
729 	int disconnect = 0;
730 
731 	/* console side, watch for read events */
732 	pollfds[0].fd = consfd;
733 	pollfds[0].events = POLLIN | POLLRDNORM | POLLRDBAND |
734 	    POLLPRI | POLLERR | POLLHUP | POLLNVAL;
735 
736 	/* client side, watch for read events */
737 	pollfds[1].fd = clifd;
738 	pollfds[1].events = pollfds[0].events;
739 
740 	/* the server socket; watch for events (new connections) */
741 	pollfds[2].fd = servfd;
742 	pollfds[2].events = pollfds[0].events;
743 
744 	/* the eventstram; watch for events (e.g.: zone halted) */
745 	pollfds[3].fd = eventstream[1];
746 	pollfds[3].events = pollfds[0].events;
747 
748 	for (;;) {
749 		pollfds[0].revents = pollfds[1].revents = 0;
750 		pollfds[2].revents = pollfds[3].revents = 0;
751 
752 		ret = poll(pollfds,
753 		    sizeof (pollfds) / sizeof (struct pollfd), -1);
754 		if (ret == -1 && errno != EINTR) {
755 			zerror(zlogp, B_TRUE, "poll failed");
756 			/* we are hosed, close connection */
757 			break;
758 		}
759 
760 		/* event from console side */
761 		if (pollfds[0].revents) {
762 			if (pollfds[0].revents &
763 			    (POLLIN | POLLRDNORM | POLLRDBAND | POLLPRI)) {
764 				errno = 0;
765 				cc = read(consfd, ibuf, BUFSIZ);
766 				if (cc <= 0 && (errno != EINTR) &&
767 				    (errno != EAGAIN))
768 					break;
769 				/*
770 				 * Lose I/O if no one is listening
771 				 */
772 				if (clifd != -1 && cc > 0)
773 					(void) write(clifd, ibuf, cc);
774 			} else {
775 				pollerr = pollfds[0].revents;
776 				zerror(zlogp, B_FALSE,
777 				    "closing connection with (console) "
778 				    "pollerr %d\n", pollerr);
779 				break;
780 			}
781 		}
782 
783 		/* event from client side */
784 		if (pollfds[1].revents) {
785 			if (pollfds[1].revents &
786 			    (POLLIN | POLLRDNORM | POLLRDBAND | POLLPRI)) {
787 				errno = 0;
788 				cc = read(clifd, ibuf, BUFSIZ);
789 				if (cc <= 0 && (errno != EINTR) &&
790 				    (errno != EAGAIN))
791 					break;
792 				(void) write(consfd, ibuf, cc);
793 			} else {
794 				pollerr = pollfds[1].revents;
795 				zerror(zlogp, B_FALSE,
796 				    "closing connection with (client) "
797 				    "pollerr %d\n", pollerr);
798 				break;
799 			}
800 		}
801 
802 		/* event from server socket */
803 		if (pollfds[2].revents &&
804 		    (pollfds[2].revents & (POLLIN | POLLRDNORM))) {
805 			if (clifd != -1) {
806 				/*
807 				 * Test the client to see if it is really
808 				 * still alive.  If it has died but we
809 				 * haven't yet detected that, we might
810 				 * deny a legitimate connect attempt.  If it
811 				 * is dead, we break out; once we tear down
812 				 * the old connection, the new connection
813 				 * will happen.
814 				 */
815 				if (test_client(clifd) == -1) {
816 					break;
817 				}
818 				/* we're already handling a client */
819 				reject_client(servfd, clipid);
820 
821 
822 			} else if ((clifd = accept_client(servfd, &clipid,
823 			    clilocale, sizeof (clilocale),
824 			    &disconnect)) != -1) {
825 				pollfds[1].fd = clifd;
826 
827 			} else {
828 				break;
829 			}
830 		}
831 
832 		/*
833 		 * Watch for events on the eventstream.  This is how we get
834 		 * notified of the zone halting, etc.  It provides us a
835 		 * "wakeup" from poll when important things happen, which
836 		 * is good.
837 		 */
838 		if (pollfds[3].revents) {
839 			int evt = eventstream_read();
840 			/*
841 			 * After we drain out the event, if we aren't servicing
842 			 * a console client, we hop back out to our caller,
843 			 * which will check to see if it is time to shutdown
844 			 * the daemon, or if we should take another console
845 			 * service lap.
846 			 */
847 			if (clifd == -1) {
848 				break;
849 			}
850 			event_message(clifd, clilocale, evt, disconnect);
851 			/*
852 			 * Special handling for the message that the zone is
853 			 * uninstalling; we boot the client, then break out
854 			 * of this function.  When we return to the
855 			 * serve_console loop, we will see that the zone is
856 			 * in a state < READY, and so zoneadmd will shutdown.
857 			 */
858 			if (evt == Z_EVT_ZONE_UNINSTALLING) {
859 				break;
860 			}
861 			/*
862 			 * Diconnect if -C and -d options were specified and
863 			 * zone was halted or failed to boot.
864 			 */
865 			if ((evt == Z_EVT_ZONE_HALTED ||
866 			    evt == Z_EVT_ZONE_BOOTFAILED) && disconnect) {
867 				break;
868 			}
869 		}
870 
871 	}
872 
873 	if (clifd != -1) {
874 		(void) shutdown(clifd, SHUT_RDWR);
875 		(void) close(clifd);
876 	}
877 }
878 
879 int
880 init_console(zlog_t *zlogp)
881 {
882 	if (init_console_dev(zlogp) == -1) {
883 		zerror(zlogp, B_FALSE,
884 		    "console setup: device initialization failed");
885 		return (-1);
886 	}
887 
888 	if ((serverfd = init_console_sock(zlogp)) == -1) {
889 		zerror(zlogp, B_FALSE,
890 		    "console setup: socket initialization failed");
891 		return (-1);
892 	}
893 	return (0);
894 }
895 
896 /*
897  * serve_console() is the master loop for driving console I/O.  It is also the
898  * routine which is ultimately responsible for "pulling the plug" on zoneadmd
899  * when it realizes that the daemon should shut down.
900  *
901  * The rules for shutdown are: there must be no console client, and the zone
902  * state must be < ready.  However, we need to give things a chance to actually
903  * get going when the daemon starts up-- otherwise the daemon would immediately
904  * exit on startup if the zone was in the installed state, so we first drop
905  * into the do_console_io() loop in order to give *something* a chance to
906  * happen.
907  */
908 void
909 serve_console(zlog_t *zlogp)
910 {
911 	int managerfd;
912 	zone_state_t zstate;
913 	char conspath[MAXPATHLEN];
914 
915 	(void) snprintf(conspath, sizeof (conspath),
916 	    "/dev/zcons/%s/%s", zone_name, ZCONS_MANAGER_NAME);
917 
918 	for (;;) {
919 		managerfd = open(conspath, O_RDWR|O_NONBLOCK|O_NOCTTY);
920 		if (managerfd == -1) {
921 			zerror(zlogp, B_TRUE, "failed to open console manager");
922 			(void) mutex_lock(&lock);
923 			goto death;
924 		}
925 
926 		/*
927 		 * Setting RPROTDIS on the stream means that the control
928 		 * portion of messages received (which we don't care about)
929 		 * will be discarded by the stream head.  If we allowed such
930 		 * messages, we wouldn't be able to use read(2), as it fails
931 		 * (EBADMSG) when a message with a control element is received.
932 		 */
933 		if (ioctl(managerfd, I_SRDOPT, RNORM|RPROTDIS) == -1) {
934 			zerror(zlogp, B_TRUE, "failed to set options on "
935 			    "console manager");
936 			(void) mutex_lock(&lock);
937 			goto death;
938 		}
939 
940 		do_console_io(zlogp, managerfd, serverfd);
941 
942 		/*
943 		 * We would prefer not to do this, but hostile zone processes
944 		 * can cause the stream to become tainted, and reads will
945 		 * fail.  So, in case something has gone seriously ill,
946 		 * we dismantle the stream and reopen the console when we
947 		 * take another lap.
948 		 */
949 		(void) close(managerfd);
950 
951 		(void) mutex_lock(&lock);
952 		/*
953 		 * We need to set death_throes (see below) atomically with
954 		 * respect to noticing that (a) we have no console client and
955 		 * (b) the zone is not installed.  Otherwise we could get a
956 		 * request to boot during this time.  Once we set death_throes,
957 		 * any incoming door stuff will be turned away.
958 		 */
959 		if (zone_get_state(zone_name, &zstate) == Z_OK) {
960 			if (zstate < ZONE_STATE_READY)
961 				goto death;
962 		} else {
963 			zerror(zlogp, B_FALSE,
964 			    "unable to determine state of zone");
965 			goto death;
966 		}
967 		/*
968 		 * Even if zone_get_state() fails, stay conservative, and
969 		 * take another lap.
970 		 */
971 		(void) mutex_unlock(&lock);
972 	}
973 
974 death:
975 	assert(MUTEX_HELD(&lock));
976 	in_death_throes = B_TRUE;
977 	(void) mutex_unlock(&lock);
978 
979 	destroy_console_sock(serverfd);
980 	(void) destroy_console_devs(zlogp);
981 }
982