zinject.c revision 262101
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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012 by Delphix. All rights reserved.
24 */
25
26/*
27 * ZFS Fault Injector
28 *
29 * This userland component takes a set of options and uses libzpool to translate
30 * from a user-visible object type and name to an internal representation.
31 * There are two basic types of faults: device faults and data faults.
32 *
33 *
34 * DEVICE FAULTS
35 *
36 * Errors can be injected into a particular vdev using the '-d' option.  This
37 * option takes a path or vdev GUID to uniquely identify the device within a
38 * pool.  There are two types of errors that can be injected, EIO and ENXIO,
39 * that can be controlled through the '-e' option.  The default is ENXIO.  For
40 * EIO failures, any attempt to read data from the device will return EIO, but
41 * subsequent attempt to reopen the device will succeed.  For ENXIO failures,
42 * any attempt to read from the device will return EIO, but any attempt to
43 * reopen the device will also return ENXIO.
44 * For label faults, the -L option must be specified. This allows faults
45 * to be injected into either the nvlist, uberblock, pad1, or pad2 region
46 * of all the labels for the specified device.
47 *
48 * This form of the command looks like:
49 *
50 * 	zinject -d device [-e errno] [-L <uber | nvlist | pad1 | pad2>] pool
51 *
52 *
53 * DATA FAULTS
54 *
55 * We begin with a tuple of the form:
56 *
57 * 	<type,level,range,object>
58 *
59 * 	type	A string describing the type of data to target.  Each type
60 * 		implicitly describes how to interpret 'object'. Currently,
61 * 		the following values are supported:
62 *
63 * 		data		User data for a file
64 * 		dnode		Dnode for a file or directory
65 *
66 *		The following MOS objects are special.  Instead of injecting
67 *		errors on a particular object or blkid, we inject errors across
68 *		all objects of the given type.
69 *
70 * 		mos		Any data in the MOS
71 * 		mosdir		object directory
72 * 		config		pool configuration
73 * 		bpobj		blkptr list
74 * 		spacemap	spacemap
75 * 		metaslab	metaslab
76 * 		errlog		persistent error log
77 *
78 * 	level	Object level.  Defaults to '0', not applicable to all types.  If
79 * 		a range is given, this corresponds to the indirect block
80 * 		corresponding to the specific range.
81 *
82 *	range	A numerical range [start,end) within the object.  Defaults to
83 *		the full size of the file.
84 *
85 * 	object	A string describing the logical location of the object.  For
86 * 		files and directories (currently the only supported types),
87 * 		this is the path of the object on disk.
88 *
89 * This is translated, via libzpool, into the following internal representation:
90 *
91 * 	<type,objset,object,level,range>
92 *
93 * These types should be self-explanatory.  This tuple is then passed to the
94 * kernel via a special ioctl() to initiate fault injection for the given
95 * object.  Note that 'type' is not strictly necessary for fault injection, but
96 * is used when translating existing faults into a human-readable string.
97 *
98 *
99 * The command itself takes one of the forms:
100 *
101 * 	zinject
102 * 	zinject <-a | -u pool>
103 * 	zinject -c <id|all>
104 * 	zinject [-q] <-t type> [-f freq] [-u] [-a] [-m] [-e errno] [-l level]
105 *	    [-r range] <object>
106 * 	zinject [-f freq] [-a] [-m] [-u] -b objset:object:level:start:end pool
107 *
108 * With no arguments, the command prints all currently registered injection
109 * handlers, with their numeric identifiers.
110 *
111 * The '-c' option will clear the given handler, or all handlers if 'all' is
112 * specified.
113 *
114 * The '-e' option takes a string describing the errno to simulate.  This must
115 * be either 'io' or 'checksum'.  In most cases this will result in the same
116 * behavior, but RAID-Z will produce a different set of ereports for this
117 * situation.
118 *
119 * The '-a', '-u', and '-m' flags toggle internal flush behavior.  If '-a' is
120 * specified, then the ARC cache is flushed appropriately.  If '-u' is
121 * specified, then the underlying SPA is unloaded.  Either of these flags can be
122 * specified independently of any other handlers.  The '-m' flag automatically
123 * does an unmount and remount of the underlying dataset to aid in flushing the
124 * cache.
125 *
126 * The '-f' flag controls the frequency of errors injected, expressed as a
127 * integer percentage between 1 and 100.  The default is 100.
128 *
129 * The this form is responsible for actually injecting the handler into the
130 * framework.  It takes the arguments described above, translates them to the
131 * internal tuple using libzpool, and then issues an ioctl() to register the
132 * handler.
133 *
134 * The final form can target a specific bookmark, regardless of whether a
135 * human-readable interface has been designed.  It allows developers to specify
136 * a particular block by number.
137 */
138
139#include <errno.h>
140#include <fcntl.h>
141#include <stdio.h>
142#include <stdlib.h>
143#include <strings.h>
144#include <unistd.h>
145
146#include <sys/fs/zfs.h>
147#include <sys/param.h>
148#include <sys/mount.h>
149
150#include <libzfs.h>
151#include <libzfs_compat.h>
152
153#undef verify	/* both libzfs.h and zfs_context.h want to define this */
154
155#include "zinject.h"
156
157libzfs_handle_t *g_zfs;
158int zfs_fd;
159
160#ifndef ECKSUM
161#define	ECKSUM	EBADE
162#endif
163
164static const char *errtable[TYPE_INVAL] = {
165	"data",
166	"dnode",
167	"mos",
168	"mosdir",
169	"metaslab",
170	"config",
171	"bpobj",
172	"spacemap",
173	"errlog",
174	"uber",
175	"nvlist",
176	"pad1",
177	"pad2"
178};
179
180static err_type_t
181name_to_type(const char *arg)
182{
183	int i;
184	for (i = 0; i < TYPE_INVAL; i++)
185		if (strcmp(errtable[i], arg) == 0)
186			return (i);
187
188	return (TYPE_INVAL);
189}
190
191static const char *
192type_to_name(uint64_t type)
193{
194	switch (type) {
195	case DMU_OT_OBJECT_DIRECTORY:
196		return ("mosdir");
197	case DMU_OT_OBJECT_ARRAY:
198		return ("metaslab");
199	case DMU_OT_PACKED_NVLIST:
200		return ("config");
201	case DMU_OT_BPOBJ:
202		return ("bpobj");
203	case DMU_OT_SPACE_MAP:
204		return ("spacemap");
205	case DMU_OT_ERROR_LOG:
206		return ("errlog");
207	default:
208		return ("-");
209	}
210}
211
212
213/*
214 * Print usage message.
215 */
216void
217usage(void)
218{
219	(void) printf(
220	    "usage:\n"
221	    "\n"
222	    "\tzinject\n"
223	    "\n"
224	    "\t\tList all active injection records.\n"
225	    "\n"
226	    "\tzinject -c <id|all>\n"
227	    "\n"
228	    "\t\tClear the particular record (if given a numeric ID), or\n"
229	    "\t\tall records if 'all' is specificed.\n"
230	    "\n"
231	    "\tzinject -p <function name> pool\n"
232	    "\t\tInject a panic fault at the specified function. Only \n"
233	    "\t\tfunctions which call spa_vdev_config_exit(), or \n"
234	    "\t\tspa_vdev_exit() will trigger a panic.\n"
235	    "\n"
236	    "\tzinject -d device [-e errno] [-L <nvlist|uber|pad1|pad2>] [-F]\n"
237	    "\t    [-T <read|write|free|claim|all> pool\n"
238	    "\t\tInject a fault into a particular device or the device's\n"
239	    "\t\tlabel.  Label injection can either be 'nvlist', 'uber',\n "
240	    "\t\t'pad1', or 'pad2'.\n"
241	    "\t\t'errno' can be 'nxio' (the default), 'io', or 'dtl'.\n"
242	    "\n"
243	    "\tzinject -d device -A <degrade|fault> pool\n"
244	    "\t\tPerform a specific action on a particular device\n"
245	    "\n"
246	    "\tzinject -I [-s <seconds> | -g <txgs>] pool\n"
247	    "\t\tCause the pool to stop writing blocks yet not\n"
248	    "\t\treport errors for a duration.  Simulates buggy hardware\n"
249	    "\t\tthat fails to honor cache flush requests.\n"
250	    "\t\tDefault duration is 30 seconds.  The machine is panicked\n"
251	    "\t\tat the end of the duration.\n"
252	    "\n"
253	    "\tzinject -b objset:object:level:blkid pool\n"
254	    "\n"
255	    "\t\tInject an error into pool 'pool' with the numeric bookmark\n"
256	    "\t\tspecified by the remaining tuple.  Each number is in\n"
257	    "\t\thexidecimal, and only one block can be specified.\n"
258	    "\n"
259	    "\tzinject [-q] <-t type> [-e errno] [-l level] [-r range]\n"
260	    "\t    [-a] [-m] [-u] [-f freq] <object>\n"
261	    "\n"
262	    "\t\tInject an error into the object specified by the '-t' option\n"
263	    "\t\tand the object descriptor.  The 'object' parameter is\n"
264	    "\t\tinterperted depending on the '-t' option.\n"
265	    "\n"
266	    "\t\t-q\tQuiet mode.  Only print out the handler number added.\n"
267	    "\t\t-e\tInject a specific error.  Must be either 'io' or\n"
268	    "\t\t\t'checksum'.  Default is 'io'.\n"
269	    "\t\t-l\tInject error at a particular block level. Default is "
270	    "0.\n"
271	    "\t\t-m\tAutomatically remount underlying filesystem.\n"
272	    "\t\t-r\tInject error over a particular logical range of an\n"
273	    "\t\t\tobject.  Will be translated to the appropriate blkid\n"
274	    "\t\t\trange according to the object's properties.\n"
275	    "\t\t-a\tFlush the ARC cache.  Can be specified without any\n"
276	    "\t\t\tassociated object.\n"
277	    "\t\t-u\tUnload the associated pool.  Can be specified with only\n"
278	    "\t\t\ta pool object.\n"
279	    "\t\t-f\tOnly inject errors a fraction of the time.  Expressed as\n"
280	    "\t\t\ta percentage between 1 and 100.\n"
281	    "\n"
282	    "\t-t data\t\tInject an error into the plain file contents of a\n"
283	    "\t\t\tfile.  The object must be specified as a complete path\n"
284	    "\t\t\tto a file on a ZFS filesystem.\n"
285	    "\n"
286	    "\t-t dnode\tInject an error into the metadnode in the block\n"
287	    "\t\t\tcorresponding to the dnode for a file or directory.  The\n"
288	    "\t\t\t'-r' option is incompatible with this mode.  The object\n"
289	    "\t\t\tis specified as a complete path to a file or directory\n"
290	    "\t\t\ton a ZFS filesystem.\n"
291	    "\n"
292	    "\t-t <mos>\tInject errors into the MOS for objects of the given\n"
293	    "\t\t\ttype.  Valid types are: mos, mosdir, config, bpobj,\n"
294	    "\t\t\tspacemap, metaslab, errlog.  The only valid <object> is\n"
295	    "\t\t\tthe poolname.\n");
296}
297
298static int
299iter_handlers(int (*func)(int, const char *, zinject_record_t *, void *),
300    void *data)
301{
302	zfs_cmd_t zc = { 0 };
303	int ret;
304
305	while (ioctl(zfs_fd, ZFS_IOC_INJECT_LIST_NEXT, &zc) == 0)
306		if ((ret = func((int)zc.zc_guid, zc.zc_name,
307		    &zc.zc_inject_record, data)) != 0)
308			return (ret);
309
310	if (errno != ENOENT) {
311		(void) fprintf(stderr, "Unable to list handlers: %s\n",
312		    strerror(errno));
313		return (-1);
314	}
315
316	return (0);
317}
318
319static int
320print_data_handler(int id, const char *pool, zinject_record_t *record,
321    void *data)
322{
323	int *count = data;
324
325	if (record->zi_guid != 0 || record->zi_func[0] != '\0')
326		return (0);
327
328	if (*count == 0) {
329		(void) printf("%3s  %-15s  %-6s  %-6s  %-8s  %3s  %-15s\n",
330		    "ID", "POOL", "OBJSET", "OBJECT", "TYPE", "LVL",  "RANGE");
331		(void) printf("---  ---------------  ------  "
332		    "------  --------  ---  ---------------\n");
333	}
334
335	*count += 1;
336
337	(void) printf("%3d  %-15s  %-6llu  %-6llu  %-8s  %3d  ", id, pool,
338	    (u_longlong_t)record->zi_objset, (u_longlong_t)record->zi_object,
339	    type_to_name(record->zi_type), record->zi_level);
340
341	if (record->zi_start == 0 &&
342	    record->zi_end == -1ULL)
343		(void) printf("all\n");
344	else
345		(void) printf("[%llu, %llu]\n", (u_longlong_t)record->zi_start,
346		    (u_longlong_t)record->zi_end);
347
348	return (0);
349}
350
351static int
352print_device_handler(int id, const char *pool, zinject_record_t *record,
353    void *data)
354{
355	int *count = data;
356
357	if (record->zi_guid == 0 || record->zi_func[0] != '\0')
358		return (0);
359
360	if (*count == 0) {
361		(void) printf("%3s  %-15s  %s\n", "ID", "POOL", "GUID");
362		(void) printf("---  ---------------  ----------------\n");
363	}
364
365	*count += 1;
366
367	(void) printf("%3d  %-15s  %llx\n", id, pool,
368	    (u_longlong_t)record->zi_guid);
369
370	return (0);
371}
372
373static int
374print_panic_handler(int id, const char *pool, zinject_record_t *record,
375    void *data)
376{
377	int *count = data;
378
379	if (record->zi_func[0] == '\0')
380		return (0);
381
382	if (*count == 0) {
383		(void) printf("%3s  %-15s  %s\n", "ID", "POOL", "FUNCTION");
384		(void) printf("---  ---------------  ----------------\n");
385	}
386
387	*count += 1;
388
389	(void) printf("%3d  %-15s  %s\n", id, pool, record->zi_func);
390
391	return (0);
392}
393
394/*
395 * Print all registered error handlers.  Returns the number of handlers
396 * registered.
397 */
398static int
399print_all_handlers(void)
400{
401	int count = 0, total = 0;
402
403	(void) iter_handlers(print_device_handler, &count);
404	if (count > 0) {
405		total += count;
406		(void) printf("\n");
407		count = 0;
408	}
409
410	(void) iter_handlers(print_data_handler, &count);
411	if (count > 0) {
412		total += count;
413		(void) printf("\n");
414		count = 0;
415	}
416
417	(void) iter_handlers(print_panic_handler, &count);
418
419	return (count + total);
420}
421
422/* ARGSUSED */
423static int
424cancel_one_handler(int id, const char *pool, zinject_record_t *record,
425    void *data)
426{
427	zfs_cmd_t zc = { 0 };
428
429	zc.zc_guid = (uint64_t)id;
430
431	if (ioctl(zfs_fd, ZFS_IOC_CLEAR_FAULT, &zc) != 0) {
432		(void) fprintf(stderr, "failed to remove handler %d: %s\n",
433		    id, strerror(errno));
434		return (1);
435	}
436
437	return (0);
438}
439
440/*
441 * Remove all fault injection handlers.
442 */
443static int
444cancel_all_handlers(void)
445{
446	int ret = iter_handlers(cancel_one_handler, NULL);
447
448	if (ret == 0)
449		(void) printf("removed all registered handlers\n");
450
451	return (ret);
452}
453
454/*
455 * Remove a specific fault injection handler.
456 */
457static int
458cancel_handler(int id)
459{
460	zfs_cmd_t zc = { 0 };
461
462	zc.zc_guid = (uint64_t)id;
463
464	if (ioctl(zfs_fd, ZFS_IOC_CLEAR_FAULT, &zc) != 0) {
465		(void) fprintf(stderr, "failed to remove handler %d: %s\n",
466		    id, strerror(errno));
467		return (1);
468	}
469
470	(void) printf("removed handler %d\n", id);
471
472	return (0);
473}
474
475/*
476 * Register a new fault injection handler.
477 */
478static int
479register_handler(const char *pool, int flags, zinject_record_t *record,
480    int quiet)
481{
482	zfs_cmd_t zc = { 0 };
483
484	(void) strcpy(zc.zc_name, pool);
485	zc.zc_inject_record = *record;
486	zc.zc_guid = flags;
487
488	if (ioctl(zfs_fd, ZFS_IOC_INJECT_FAULT, &zc) != 0) {
489		(void) fprintf(stderr, "failed to add handler: %s\n",
490		    strerror(errno));
491		return (1);
492	}
493
494	if (flags & ZINJECT_NULL)
495		return (0);
496
497	if (quiet) {
498		(void) printf("%llu\n", (u_longlong_t)zc.zc_guid);
499	} else {
500		(void) printf("Added handler %llu with the following "
501		    "properties:\n", (u_longlong_t)zc.zc_guid);
502		(void) printf("  pool: %s\n", pool);
503		if (record->zi_guid) {
504			(void) printf("  vdev: %llx\n",
505			    (u_longlong_t)record->zi_guid);
506		} else if (record->zi_func[0] != '\0') {
507			(void) printf("  panic function: %s\n",
508			    record->zi_func);
509		} else if (record->zi_duration > 0) {
510			(void) printf(" time: %lld seconds\n",
511			    (u_longlong_t)record->zi_duration);
512		} else if (record->zi_duration < 0) {
513			(void) printf(" txgs: %lld \n",
514			    (u_longlong_t)-record->zi_duration);
515		} else {
516			(void) printf("objset: %llu\n",
517			    (u_longlong_t)record->zi_objset);
518			(void) printf("object: %llu\n",
519			    (u_longlong_t)record->zi_object);
520			(void) printf("  type: %llu\n",
521			    (u_longlong_t)record->zi_type);
522			(void) printf(" level: %d\n", record->zi_level);
523			if (record->zi_start == 0 &&
524			    record->zi_end == -1ULL)
525				(void) printf(" range: all\n");
526			else
527				(void) printf(" range: [%llu, %llu)\n",
528				    (u_longlong_t)record->zi_start,
529				    (u_longlong_t)record->zi_end);
530		}
531	}
532
533	return (0);
534}
535
536int
537perform_action(const char *pool, zinject_record_t *record, int cmd)
538{
539	zfs_cmd_t zc = { 0 };
540
541	ASSERT(cmd == VDEV_STATE_DEGRADED || cmd == VDEV_STATE_FAULTED);
542	(void) strlcpy(zc.zc_name, pool, sizeof (zc.zc_name));
543	zc.zc_guid = record->zi_guid;
544	zc.zc_cookie = cmd;
545
546	if (ioctl(zfs_fd, ZFS_IOC_VDEV_SET_STATE, &zc) == 0)
547		return (0);
548
549	return (1);
550}
551
552int
553main(int argc, char **argv)
554{
555	int c;
556	char *range = NULL;
557	char *cancel = NULL;
558	char *end;
559	char *raw = NULL;
560	char *device = NULL;
561	int level = 0;
562	int quiet = 0;
563	int error = 0;
564	int domount = 0;
565	int io_type = ZIO_TYPES;
566	int action = VDEV_STATE_UNKNOWN;
567	err_type_t type = TYPE_INVAL;
568	err_type_t label = TYPE_INVAL;
569	zinject_record_t record = { 0 };
570	char pool[MAXNAMELEN];
571	char dataset[MAXNAMELEN];
572	zfs_handle_t *zhp;
573	int nowrites = 0;
574	int dur_txg = 0;
575	int dur_secs = 0;
576	int ret;
577	int flags = 0;
578
579	if ((g_zfs = libzfs_init()) == NULL) {
580		(void) fprintf(stderr, "internal error: failed to "
581		    "initialize ZFS library\n");
582		return (1);
583	}
584
585	libzfs_print_on_error(g_zfs, B_TRUE);
586
587	if ((zfs_fd = open(ZFS_DEV, O_RDWR)) < 0) {
588		(void) fprintf(stderr, "failed to open ZFS device\n");
589		return (1);
590	}
591
592	if (argc == 1) {
593		/*
594		 * No arguments.  Print the available handlers.  If there are no
595		 * available handlers, direct the user to '-h' for help
596		 * information.
597		 */
598		if (print_all_handlers() == 0) {
599			(void) printf("No handlers registered.\n");
600			(void) printf("Run 'zinject -h' for usage "
601			    "information.\n");
602		}
603
604		return (0);
605	}
606
607	while ((c = getopt(argc, argv,
608	    ":aA:b:d:D:f:Fg:qhIc:t:T:l:mr:s:e:uL:p:")) != -1) {
609		switch (c) {
610		case 'a':
611			flags |= ZINJECT_FLUSH_ARC;
612			break;
613		case 'A':
614			if (strcasecmp(optarg, "degrade") == 0) {
615				action = VDEV_STATE_DEGRADED;
616			} else if (strcasecmp(optarg, "fault") == 0) {
617				action = VDEV_STATE_FAULTED;
618			} else {
619				(void) fprintf(stderr, "invalid action '%s': "
620				    "must be 'degrade' or 'fault'\n", optarg);
621				usage();
622				return (1);
623			}
624			break;
625		case 'b':
626			raw = optarg;
627			break;
628		case 'c':
629			cancel = optarg;
630			break;
631		case 'd':
632			device = optarg;
633			break;
634		case 'D':
635			record.zi_timer = strtoull(optarg, &end, 10);
636			if (errno != 0 || *end != '\0') {
637				(void) fprintf(stderr, "invalid i/o delay "
638				    "value: '%s'\n", optarg);
639				usage();
640				return (1);
641			}
642			break;
643		case 'e':
644			if (strcasecmp(optarg, "io") == 0) {
645				error = EIO;
646			} else if (strcasecmp(optarg, "checksum") == 0) {
647				error = ECKSUM;
648			} else if (strcasecmp(optarg, "nxio") == 0) {
649				error = ENXIO;
650			} else if (strcasecmp(optarg, "dtl") == 0) {
651				error = ECHILD;
652			} else {
653				(void) fprintf(stderr, "invalid error type "
654				    "'%s': must be 'io', 'checksum' or "
655				    "'nxio'\n", optarg);
656				usage();
657				return (1);
658			}
659			break;
660		case 'f':
661			record.zi_freq = atoi(optarg);
662			if (record.zi_freq < 1 || record.zi_freq > 100) {
663				(void) fprintf(stderr, "frequency range must "
664				    "be in the range (0, 100]\n");
665				return (1);
666			}
667			break;
668		case 'F':
669			record.zi_failfast = B_TRUE;
670			break;
671		case 'g':
672			dur_txg = 1;
673			record.zi_duration = (int)strtol(optarg, &end, 10);
674			if (record.zi_duration <= 0 || *end != '\0') {
675				(void) fprintf(stderr, "invalid duration '%s': "
676				    "must be a positive integer\n", optarg);
677				usage();
678				return (1);
679			}
680			/* store duration of txgs as its negative */
681			record.zi_duration *= -1;
682			break;
683		case 'h':
684			usage();
685			return (0);
686		case 'I':
687			/* default duration, if one hasn't yet been defined */
688			nowrites = 1;
689			if (dur_secs == 0 && dur_txg == 0)
690				record.zi_duration = 30;
691			break;
692		case 'l':
693			level = (int)strtol(optarg, &end, 10);
694			if (*end != '\0') {
695				(void) fprintf(stderr, "invalid level '%s': "
696				    "must be an integer\n", optarg);
697				usage();
698				return (1);
699			}
700			break;
701		case 'm':
702			domount = 1;
703			break;
704		case 'p':
705			(void) strlcpy(record.zi_func, optarg,
706			    sizeof (record.zi_func));
707			record.zi_cmd = ZINJECT_PANIC;
708			break;
709		case 'q':
710			quiet = 1;
711			break;
712		case 'r':
713			range = optarg;
714			break;
715		case 's':
716			dur_secs = 1;
717			record.zi_duration = (int)strtol(optarg, &end, 10);
718			if (record.zi_duration <= 0 || *end != '\0') {
719				(void) fprintf(stderr, "invalid duration '%s': "
720				    "must be a positive integer\n", optarg);
721				usage();
722				return (1);
723			}
724			break;
725		case 'T':
726			if (strcasecmp(optarg, "read") == 0) {
727				io_type = ZIO_TYPE_READ;
728			} else if (strcasecmp(optarg, "write") == 0) {
729				io_type = ZIO_TYPE_WRITE;
730			} else if (strcasecmp(optarg, "free") == 0) {
731				io_type = ZIO_TYPE_FREE;
732			} else if (strcasecmp(optarg, "claim") == 0) {
733				io_type = ZIO_TYPE_CLAIM;
734			} else if (strcasecmp(optarg, "all") == 0) {
735				io_type = ZIO_TYPES;
736			} else {
737				(void) fprintf(stderr, "invalid I/O type "
738				    "'%s': must be 'read', 'write', 'free', "
739				    "'claim' or 'all'\n", optarg);
740				usage();
741				return (1);
742			}
743			break;
744		case 't':
745			if ((type = name_to_type(optarg)) == TYPE_INVAL &&
746			    !MOS_TYPE(type)) {
747				(void) fprintf(stderr, "invalid type '%s'\n",
748				    optarg);
749				usage();
750				return (1);
751			}
752			break;
753		case 'u':
754			flags |= ZINJECT_UNLOAD_SPA;
755			break;
756		case 'L':
757			if ((label = name_to_type(optarg)) == TYPE_INVAL &&
758			    !LABEL_TYPE(type)) {
759				(void) fprintf(stderr, "invalid label type "
760				    "'%s'\n", optarg);
761				usage();
762				return (1);
763			}
764			break;
765		case ':':
766			(void) fprintf(stderr, "option -%c requires an "
767			    "operand\n", optopt);
768			usage();
769			return (1);
770		case '?':
771			(void) fprintf(stderr, "invalid option '%c'\n",
772			    optopt);
773			usage();
774			return (2);
775		}
776	}
777
778	argc -= optind;
779	argv += optind;
780
781	if (record.zi_duration != 0)
782		record.zi_cmd = ZINJECT_IGNORED_WRITES;
783
784	if (cancel != NULL) {
785		/*
786		 * '-c' is invalid with any other options.
787		 */
788		if (raw != NULL || range != NULL || type != TYPE_INVAL ||
789		    level != 0 || record.zi_cmd != ZINJECT_UNINITIALIZED) {
790			(void) fprintf(stderr, "cancel (-c) incompatible with "
791			    "any other options\n");
792			usage();
793			return (2);
794		}
795		if (argc != 0) {
796			(void) fprintf(stderr, "extraneous argument to '-c'\n");
797			usage();
798			return (2);
799		}
800
801		if (strcmp(cancel, "all") == 0) {
802			return (cancel_all_handlers());
803		} else {
804			int id = (int)strtol(cancel, &end, 10);
805			if (*end != '\0') {
806				(void) fprintf(stderr, "invalid handle id '%s':"
807				    " must be an integer or 'all'\n", cancel);
808				usage();
809				return (1);
810			}
811			return (cancel_handler(id));
812		}
813	}
814
815	if (device != NULL) {
816		/*
817		 * Device (-d) injection uses a completely different mechanism
818		 * for doing injection, so handle it separately here.
819		 */
820		if (raw != NULL || range != NULL || type != TYPE_INVAL ||
821		    level != 0 || record.zi_cmd != ZINJECT_UNINITIALIZED) {
822			(void) fprintf(stderr, "device (-d) incompatible with "
823			    "data error injection\n");
824			usage();
825			return (2);
826		}
827
828		if (argc != 1) {
829			(void) fprintf(stderr, "device (-d) injection requires "
830			    "a single pool name\n");
831			usage();
832			return (2);
833		}
834
835		(void) strcpy(pool, argv[0]);
836		dataset[0] = '\0';
837
838		if (error == ECKSUM) {
839			(void) fprintf(stderr, "device error type must be "
840			    "'io' or 'nxio'\n");
841			return (1);
842		}
843
844		record.zi_iotype = io_type;
845		if (translate_device(pool, device, label, &record) != 0)
846			return (1);
847		if (!error)
848			error = ENXIO;
849
850		if (action != VDEV_STATE_UNKNOWN)
851			return (perform_action(pool, &record, action));
852
853	} else if (raw != NULL) {
854		if (range != NULL || type != TYPE_INVAL || level != 0 ||
855		    record.zi_cmd != ZINJECT_UNINITIALIZED) {
856			(void) fprintf(stderr, "raw (-b) format with "
857			    "any other options\n");
858			usage();
859			return (2);
860		}
861
862		if (argc != 1) {
863			(void) fprintf(stderr, "raw (-b) format expects a "
864			    "single pool name\n");
865			usage();
866			return (2);
867		}
868
869		(void) strcpy(pool, argv[0]);
870		dataset[0] = '\0';
871
872		if (error == ENXIO) {
873			(void) fprintf(stderr, "data error type must be "
874			    "'checksum' or 'io'\n");
875			return (1);
876		}
877
878		record.zi_cmd = ZINJECT_DATA_FAULT;
879		if (translate_raw(raw, &record) != 0)
880			return (1);
881		if (!error)
882			error = EIO;
883	} else if (record.zi_cmd == ZINJECT_PANIC) {
884		if (raw != NULL || range != NULL || type != TYPE_INVAL ||
885		    level != 0 || device != NULL) {
886			(void) fprintf(stderr, "panic (-p) incompatible with "
887			    "other options\n");
888			usage();
889			return (2);
890		}
891
892		if (argc < 1 || argc > 2) {
893			(void) fprintf(stderr, "panic (-p) injection requires "
894			    "a single pool name and an optional id\n");
895			usage();
896			return (2);
897		}
898
899		(void) strcpy(pool, argv[0]);
900		if (argv[1] != NULL)
901			record.zi_type = atoi(argv[1]);
902		dataset[0] = '\0';
903	} else if (record.zi_cmd == ZINJECT_IGNORED_WRITES) {
904		if (nowrites == 0) {
905			(void) fprintf(stderr, "-s or -g meaningless "
906			    "without -I (ignore writes)\n");
907			usage();
908			return (2);
909		} else if (dur_secs && dur_txg) {
910			(void) fprintf(stderr, "choose a duration either "
911			    "in seconds (-s) or a number of txgs (-g) "
912			    "but not both\n");
913			usage();
914			return (2);
915		} else if (argc != 1) {
916			(void) fprintf(stderr, "ignore writes (-I) "
917			    "injection requires a single pool name\n");
918			usage();
919			return (2);
920		}
921
922		(void) strcpy(pool, argv[0]);
923		dataset[0] = '\0';
924	} else if (type == TYPE_INVAL) {
925		if (flags == 0) {
926			(void) fprintf(stderr, "at least one of '-b', '-d', "
927			    "'-t', '-a', '-p', '-I' or '-u' "
928			    "must be specified\n");
929			usage();
930			return (2);
931		}
932
933		if (argc == 1 && (flags & ZINJECT_UNLOAD_SPA)) {
934			(void) strcpy(pool, argv[0]);
935			dataset[0] = '\0';
936		} else if (argc != 0) {
937			(void) fprintf(stderr, "extraneous argument for "
938			    "'-f'\n");
939			usage();
940			return (2);
941		}
942
943		flags |= ZINJECT_NULL;
944	} else {
945		if (argc != 1) {
946			(void) fprintf(stderr, "missing object\n");
947			usage();
948			return (2);
949		}
950
951		if (error == ENXIO) {
952			(void) fprintf(stderr, "data error type must be "
953			    "'checksum' or 'io'\n");
954			return (1);
955		}
956
957		record.zi_cmd = ZINJECT_DATA_FAULT;
958		if (translate_record(type, argv[0], range, level, &record, pool,
959		    dataset) != 0)
960			return (1);
961		if (!error)
962			error = EIO;
963	}
964
965	/*
966	 * If this is pool-wide metadata, unmount everything.  The ioctl() will
967	 * unload the pool, so that we trigger spa-wide reopen of metadata next
968	 * time we access the pool.
969	 */
970	if (dataset[0] != '\0' && domount) {
971		if ((zhp = zfs_open(g_zfs, dataset, ZFS_TYPE_DATASET)) == NULL)
972			return (1);
973
974		if (zfs_unmount(zhp, NULL, 0) != 0)
975			return (1);
976	}
977
978	record.zi_error = error;
979
980	ret = register_handler(pool, flags, &record, quiet);
981
982	if (dataset[0] != '\0' && domount)
983		ret = (zfs_mount(zhp, NULL, 0) != 0);
984
985	libzfs_fini(g_zfs);
986
987	return (ret);
988}
989