xref: /freebsd/sys/contrib/openzfs/cmd/zinject/zinject.c (revision 59c8e88e72633afbc47a4ace0d2170d00d51f7dc)
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 https://opensource.org/licenses/CDDL-1.0.
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, 2015 by Delphix. All rights reserved.
24  * Copyright (c) 2017, Intel Corporation.
25  */
26 
27 /*
28  * ZFS Fault Injector
29  *
30  * This userland component takes a set of options and uses libzpool to translate
31  * from a user-visible object type and name to an internal representation.
32  * There are two basic types of faults: device faults and data faults.
33  *
34  *
35  * DEVICE FAULTS
36  *
37  * Errors can be injected into a particular vdev using the '-d' option.  This
38  * option takes a path or vdev GUID to uniquely identify the device within a
39  * pool.  There are four types of errors that can be injected, IO, ENXIO,
40  * ECHILD, and EILSEQ.  These can be controlled through the '-e' option and the
41  * default is ENXIO.  For EIO failures, any attempt to read data from the device
42  * will return EIO, but a subsequent attempt to reopen the device will succeed.
43  * For ENXIO failures, any attempt to read from the device will return EIO, but
44  * any attempt to reopen the device will also return ENXIO.  The EILSEQ failures
45  * only apply to read operations (-T read) and will flip a bit after the device
46  * has read the original data.
47  *
48  * For label faults, the -L option must be specified. This allows faults
49  * to be injected into either the nvlist, uberblock, pad1, or pad2 region
50  * of all the labels for the specified device.
51  *
52  * This form of the command looks like:
53  *
54  * 	zinject -d device [-e errno] [-L <uber | nvlist | pad1 | pad2>] pool
55  *
56  *
57  * DATA FAULTS
58  *
59  * We begin with a tuple of the form:
60  *
61  * 	<type,level,range,object>
62  *
63  * 	type	A string describing the type of data to target.  Each type
64  * 		implicitly describes how to interpret 'object'. Currently,
65  * 		the following values are supported:
66  *
67  * 		data		User data for a file
68  * 		dnode		Dnode for a file or directory
69  *
70  *		The following MOS objects are special.  Instead of injecting
71  *		errors on a particular object or blkid, we inject errors across
72  *		all objects of the given type.
73  *
74  * 		mos		Any data in the MOS
75  * 		mosdir		object directory
76  * 		config		pool configuration
77  * 		bpobj		blkptr list
78  * 		spacemap	spacemap
79  * 		metaslab	metaslab
80  * 		errlog		persistent error log
81  *
82  * 	level	Object level.  Defaults to '0', not applicable to all types.  If
83  * 		a range is given, this corresponds to the indirect block
84  * 		corresponding to the specific range.
85  *
86  *	range	A numerical range [start,end) within the object.  Defaults to
87  *		the full size of the file.
88  *
89  * 	object	A string describing the logical location of the object.  For
90  * 		files and directories (currently the only supported types),
91  * 		this is the path of the object on disk.
92  *
93  * This is translated, via libzpool, into the following internal representation:
94  *
95  * 	<type,objset,object,level,range>
96  *
97  * These types should be self-explanatory.  This tuple is then passed to the
98  * kernel via a special ioctl() to initiate fault injection for the given
99  * object.  Note that 'type' is not strictly necessary for fault injection, but
100  * is used when translating existing faults into a human-readable string.
101  *
102  *
103  * The command itself takes one of the forms:
104  *
105  * 	zinject
106  * 	zinject <-a | -u pool>
107  * 	zinject -c <id|all>
108  * 	zinject [-q] <-t type> [-f freq] [-u] [-a] [-m] [-e errno] [-l level]
109  *	    [-r range] <object>
110  * 	zinject [-f freq] [-a] [-m] [-u] -b objset:object:level:start:end pool
111  *
112  * With no arguments, the command prints all currently registered injection
113  * handlers, with their numeric identifiers.
114  *
115  * The '-c' option will clear the given handler, or all handlers if 'all' is
116  * specified.
117  *
118  * The '-e' option takes a string describing the errno to simulate.  This must
119  * be one of 'io', 'checksum', 'decompress', or 'decrypt'.  In most cases this
120  * will result in the same behavior, but RAID-Z will produce a different set of
121  * ereports for this situation.
122  *
123  * The '-a', '-u', and '-m' flags toggle internal flush behavior.  If '-a' is
124  * specified, then the ARC cache is flushed appropriately.  If '-u' is
125  * specified, then the underlying SPA is unloaded.  Either of these flags can be
126  * specified independently of any other handlers.  The '-m' flag automatically
127  * does an unmount and remount of the underlying dataset to aid in flushing the
128  * cache.
129  *
130  * The '-f' flag controls the frequency of errors injected, expressed as a
131  * real number percentage between 0.0001 and 100.  The default is 100.
132  *
133  * The this form is responsible for actually injecting the handler into the
134  * framework.  It takes the arguments described above, translates them to the
135  * internal tuple using libzpool, and then issues an ioctl() to register the
136  * handler.
137  *
138  * The final form can target a specific bookmark, regardless of whether a
139  * human-readable interface has been designed.  It allows developers to specify
140  * a particular block by number.
141  */
142 
143 #include <errno.h>
144 #include <fcntl.h>
145 #include <stdio.h>
146 #include <stdlib.h>
147 #include <string.h>
148 #include <strings.h>
149 #include <unistd.h>
150 
151 #include <sys/fs/zfs.h>
152 #include <sys/mount.h>
153 
154 #include <libzfs.h>
155 
156 #undef verify	/* both libzfs.h and zfs_context.h want to define this */
157 
158 #include "zinject.h"
159 
160 libzfs_handle_t *g_zfs;
161 int zfs_fd;
162 
163 static const char *const errtable[TYPE_INVAL] = {
164 	"data",
165 	"dnode",
166 	"mos",
167 	"mosdir",
168 	"metaslab",
169 	"config",
170 	"bpobj",
171 	"spacemap",
172 	"errlog",
173 	"uber",
174 	"nvlist",
175 	"pad1",
176 	"pad2"
177 };
178 
179 static err_type_t
180 name_to_type(const char *arg)
181 {
182 	int i;
183 	for (i = 0; i < TYPE_INVAL; i++)
184 		if (strcmp(errtable[i], arg) == 0)
185 			return (i);
186 
187 	return (TYPE_INVAL);
188 }
189 
190 static const char *
191 type_to_name(uint64_t type)
192 {
193 	switch (type) {
194 	case DMU_OT_OBJECT_DIRECTORY:
195 		return ("mosdir");
196 	case DMU_OT_OBJECT_ARRAY:
197 		return ("metaslab");
198 	case DMU_OT_PACKED_NVLIST:
199 		return ("config");
200 	case DMU_OT_BPOBJ:
201 		return ("bpobj");
202 	case DMU_OT_SPACE_MAP:
203 		return ("spacemap");
204 	case DMU_OT_ERROR_LOG:
205 		return ("errlog");
206 	default:
207 		return ("-");
208 	}
209 }
210 
211 
212 /*
213  * Print usage message.
214  */
215 void
216 usage(void)
217 {
218 	(void) printf(
219 	    "usage:\n"
220 	    "\n"
221 	    "\tzinject\n"
222 	    "\n"
223 	    "\t\tList all active injection records.\n"
224 	    "\n"
225 	    "\tzinject -c <id|all>\n"
226 	    "\n"
227 	    "\t\tClear the particular record (if given a numeric ID), or\n"
228 	    "\t\tall records if 'all' is specified.\n"
229 	    "\n"
230 	    "\tzinject -p <function name> pool\n"
231 	    "\t\tInject a panic fault at the specified function. Only \n"
232 	    "\t\tfunctions which call spa_vdev_config_exit(), or \n"
233 	    "\t\tspa_vdev_exit() will trigger a panic.\n"
234 	    "\n"
235 	    "\tzinject -d device [-e errno] [-L <nvlist|uber|pad1|pad2>] [-F]\n"
236 	    "\t\t[-T <read|write|free|claim|all>] [-f frequency] pool\n\n"
237 	    "\t\tInject a fault into a particular device or the device's\n"
238 	    "\t\tlabel.  Label injection can either be 'nvlist', 'uber',\n "
239 	    "\t\t'pad1', or 'pad2'.\n"
240 	    "\t\t'errno' can be 'nxio' (the default), 'io', 'dtl', or\n"
241 	    "\t\t'corrupt' (bit flip).\n"
242 	    "\t\t'frequency' is a value between 0.0001 and 100.0 that limits\n"
243 	    "\t\tdevice error injection to a percentage of the IOs.\n"
244 	    "\n"
245 	    "\tzinject -d device -A <degrade|fault> -D <delay secs> pool\n"
246 	    "\t\tPerform a specific action on a particular device.\n"
247 	    "\n"
248 	    "\tzinject -d device -D latency:lanes pool\n"
249 	    "\n"
250 	    "\t\tAdd an artificial delay to IO requests on a particular\n"
251 	    "\t\tdevice, such that the requests take a minimum of 'latency'\n"
252 	    "\t\tmilliseconds to complete. Each delay has an associated\n"
253 	    "\t\tnumber of 'lanes' which defines the number of concurrent\n"
254 	    "\t\tIO requests that can be processed.\n"
255 	    "\n"
256 	    "\t\tFor example, with a single lane delay of 10 ms (-D 10:1),\n"
257 	    "\t\tthe device will only be able to service a single IO request\n"
258 	    "\t\tat a time with each request taking 10 ms to complete. So,\n"
259 	    "\t\tif only a single request is submitted every 10 ms, the\n"
260 	    "\t\taverage latency will be 10 ms; but if more than one request\n"
261 	    "\t\tis submitted every 10 ms, the average latency will be more\n"
262 	    "\t\tthan 10 ms.\n"
263 	    "\n"
264 	    "\t\tSimilarly, if a delay of 10 ms is specified to have two\n"
265 	    "\t\tlanes (-D 10:2), then the device will be able to service\n"
266 	    "\t\ttwo requests at a time, each with a minimum latency of\n"
267 	    "\t\t10 ms. So, if two requests are submitted every 10 ms, then\n"
268 	    "\t\tthe average latency will be 10 ms; but if more than two\n"
269 	    "\t\trequests are submitted every 10 ms, the average latency\n"
270 	    "\t\twill be more than 10 ms.\n"
271 	    "\n"
272 	    "\t\tAlso note, these delays are additive. So two invocations\n"
273 	    "\t\tof '-D 10:1', is roughly equivalent to a single invocation\n"
274 	    "\t\tof '-D 10:2'. This also means, one can specify multiple\n"
275 	    "\t\tlanes with differing target latencies. For example, an\n"
276 	    "\t\tinvocation of '-D 10:1' followed by '-D 25:2' will\n"
277 	    "\t\tcreate 3 lanes on the device; one lane with a latency\n"
278 	    "\t\tof 10 ms and two lanes with a 25 ms latency.\n"
279 	    "\n"
280 	    "\tzinject -I [-s <seconds> | -g <txgs>] pool\n"
281 	    "\t\tCause the pool to stop writing blocks yet not\n"
282 	    "\t\treport errors for a duration.  Simulates buggy hardware\n"
283 	    "\t\tthat fails to honor cache flush requests.\n"
284 	    "\t\tDefault duration is 30 seconds.  The machine is panicked\n"
285 	    "\t\tat the end of the duration.\n"
286 	    "\n"
287 	    "\tzinject -b objset:object:level:blkid pool\n"
288 	    "\n"
289 	    "\t\tInject an error into pool 'pool' with the numeric bookmark\n"
290 	    "\t\tspecified by the remaining tuple.  Each number is in\n"
291 	    "\t\thexadecimal, and only one block can be specified.\n"
292 	    "\n"
293 	    "\tzinject [-q] <-t type> [-C dvas] [-e errno] [-l level]\n"
294 	    "\t\t[-r range] [-a] [-m] [-u] [-f freq] <object>\n"
295 	    "\n"
296 	    "\t\tInject an error into the object specified by the '-t' option\n"
297 	    "\t\tand the object descriptor.  The 'object' parameter is\n"
298 	    "\t\tinterpreted depending on the '-t' option.\n"
299 	    "\n"
300 	    "\t\t-q\tQuiet mode.  Only print out the handler number added.\n"
301 	    "\t\t-e\tInject a specific error.  Must be one of 'io',\n"
302 	    "\t\t\t'checksum', 'decompress', or 'decrypt'.  Default is 'io'.\n"
303 	    "\t\t-C\tInject the given error only into specific DVAs. The\n"
304 	    "\t\t\tDVAs should be specified as a list of 0-indexed DVAs\n"
305 	    "\t\t\tseparated by commas (ex. '0,2').\n"
306 	    "\t\t-l\tInject error at a particular block level. Default is "
307 	    "0.\n"
308 	    "\t\t-m\tAutomatically remount underlying filesystem.\n"
309 	    "\t\t-r\tInject error over a particular logical range of an\n"
310 	    "\t\t\tobject.  Will be translated to the appropriate blkid\n"
311 	    "\t\t\trange according to the object's properties.\n"
312 	    "\t\t-a\tFlush the ARC cache.  Can be specified without any\n"
313 	    "\t\t\tassociated object.\n"
314 	    "\t\t-u\tUnload the associated pool.  Can be specified with only\n"
315 	    "\t\t\ta pool object.\n"
316 	    "\t\t-f\tOnly inject errors a fraction of the time.  Expressed as\n"
317 	    "\t\t\ta percentage between 0.0001 and 100.\n"
318 	    "\n"
319 	    "\t-t data\t\tInject an error into the plain file contents of a\n"
320 	    "\t\t\tfile.  The object must be specified as a complete path\n"
321 	    "\t\t\tto a file on a ZFS filesystem.\n"
322 	    "\n"
323 	    "\t-t dnode\tInject an error into the metadnode in the block\n"
324 	    "\t\t\tcorresponding to the dnode for a file or directory.  The\n"
325 	    "\t\t\t'-r' option is incompatible with this mode.  The object\n"
326 	    "\t\t\tis specified as a complete path to a file or directory\n"
327 	    "\t\t\ton a ZFS filesystem.\n"
328 	    "\n"
329 	    "\t-t <mos>\tInject errors into the MOS for objects of the given\n"
330 	    "\t\t\ttype.  Valid types are: mos, mosdir, config, bpobj,\n"
331 	    "\t\t\tspacemap, metaslab, errlog.  The only valid <object> is\n"
332 	    "\t\t\tthe poolname.\n");
333 }
334 
335 static int
336 iter_handlers(int (*func)(int, const char *, zinject_record_t *, void *),
337     void *data)
338 {
339 	zfs_cmd_t zc = {"\0"};
340 	int ret;
341 
342 	while (zfs_ioctl(g_zfs, ZFS_IOC_INJECT_LIST_NEXT, &zc) == 0)
343 		if ((ret = func((int)zc.zc_guid, zc.zc_name,
344 		    &zc.zc_inject_record, data)) != 0)
345 			return (ret);
346 
347 	if (errno != ENOENT) {
348 		(void) fprintf(stderr, "Unable to list handlers: %s\n",
349 		    strerror(errno));
350 		return (-1);
351 	}
352 
353 	return (0);
354 }
355 
356 static int
357 print_data_handler(int id, const char *pool, zinject_record_t *record,
358     void *data)
359 {
360 	int *count = data;
361 
362 	if (record->zi_guid != 0 || record->zi_func[0] != '\0')
363 		return (0);
364 
365 	if (*count == 0) {
366 		(void) printf("%3s  %-15s  %-6s  %-6s  %-8s  %3s  %-4s  "
367 		    "%-15s\n", "ID", "POOL", "OBJSET", "OBJECT", "TYPE",
368 		    "LVL", "DVAs", "RANGE");
369 		(void) printf("---  ---------------  ------  "
370 		    "------  --------  ---  ----  ---------------\n");
371 	}
372 
373 	*count += 1;
374 
375 	(void) printf("%3d  %-15s  %-6llu  %-6llu  %-8s  %-3d  0x%02x  ",
376 	    id, pool, (u_longlong_t)record->zi_objset,
377 	    (u_longlong_t)record->zi_object, type_to_name(record->zi_type),
378 	    record->zi_level, record->zi_dvas);
379 
380 
381 	if (record->zi_start == 0 &&
382 	    record->zi_end == -1ULL)
383 		(void) printf("all\n");
384 	else
385 		(void) printf("[%llu, %llu]\n", (u_longlong_t)record->zi_start,
386 		    (u_longlong_t)record->zi_end);
387 
388 	return (0);
389 }
390 
391 static int
392 print_device_handler(int id, const char *pool, zinject_record_t *record,
393     void *data)
394 {
395 	int *count = data;
396 
397 	if (record->zi_guid == 0 || record->zi_func[0] != '\0')
398 		return (0);
399 
400 	if (record->zi_cmd == ZINJECT_DELAY_IO)
401 		return (0);
402 
403 	if (*count == 0) {
404 		(void) printf("%3s  %-15s  %s\n", "ID", "POOL", "GUID");
405 		(void) printf("---  ---------------  ----------------\n");
406 	}
407 
408 	*count += 1;
409 
410 	(void) printf("%3d  %-15s  %llx\n", id, pool,
411 	    (u_longlong_t)record->zi_guid);
412 
413 	return (0);
414 }
415 
416 static int
417 print_delay_handler(int id, const char *pool, zinject_record_t *record,
418     void *data)
419 {
420 	int *count = data;
421 
422 	if (record->zi_guid == 0 || record->zi_func[0] != '\0')
423 		return (0);
424 
425 	if (record->zi_cmd != ZINJECT_DELAY_IO)
426 		return (0);
427 
428 	if (*count == 0) {
429 		(void) printf("%3s  %-15s  %-15s  %-15s  %s\n",
430 		    "ID", "POOL", "DELAY (ms)", "LANES", "GUID");
431 		(void) printf("---  ---------------  ---------------  "
432 		    "---------------  ----------------\n");
433 	}
434 
435 	*count += 1;
436 
437 	(void) printf("%3d  %-15s  %-15llu  %-15llu  %llx\n", id, pool,
438 	    (u_longlong_t)NSEC2MSEC(record->zi_timer),
439 	    (u_longlong_t)record->zi_nlanes,
440 	    (u_longlong_t)record->zi_guid);
441 
442 	return (0);
443 }
444 
445 static int
446 print_panic_handler(int id, const char *pool, zinject_record_t *record,
447     void *data)
448 {
449 	int *count = data;
450 
451 	if (record->zi_func[0] == '\0')
452 		return (0);
453 
454 	if (*count == 0) {
455 		(void) printf("%3s  %-15s  %s\n", "ID", "POOL", "FUNCTION");
456 		(void) printf("---  ---------------  ----------------\n");
457 	}
458 
459 	*count += 1;
460 
461 	(void) printf("%3d  %-15s  %s\n", id, pool, record->zi_func);
462 
463 	return (0);
464 }
465 
466 /*
467  * Print all registered error handlers.  Returns the number of handlers
468  * registered.
469  */
470 static int
471 print_all_handlers(void)
472 {
473 	int count = 0, total = 0;
474 
475 	(void) iter_handlers(print_device_handler, &count);
476 	if (count > 0) {
477 		total += count;
478 		(void) printf("\n");
479 		count = 0;
480 	}
481 
482 	(void) iter_handlers(print_delay_handler, &count);
483 	if (count > 0) {
484 		total += count;
485 		(void) printf("\n");
486 		count = 0;
487 	}
488 
489 	(void) iter_handlers(print_data_handler, &count);
490 	if (count > 0) {
491 		total += count;
492 		(void) printf("\n");
493 		count = 0;
494 	}
495 
496 	(void) iter_handlers(print_panic_handler, &count);
497 
498 	return (count + total);
499 }
500 
501 static int
502 cancel_one_handler(int id, const char *pool, zinject_record_t *record,
503     void *data)
504 {
505 	(void) pool, (void) record, (void) data;
506 	zfs_cmd_t zc = {"\0"};
507 
508 	zc.zc_guid = (uint64_t)id;
509 
510 	if (zfs_ioctl(g_zfs, ZFS_IOC_CLEAR_FAULT, &zc) != 0) {
511 		(void) fprintf(stderr, "failed to remove handler %d: %s\n",
512 		    id, strerror(errno));
513 		return (1);
514 	}
515 
516 	return (0);
517 }
518 
519 /*
520  * Remove all fault injection handlers.
521  */
522 static int
523 cancel_all_handlers(void)
524 {
525 	int ret = iter_handlers(cancel_one_handler, NULL);
526 
527 	if (ret == 0)
528 		(void) printf("removed all registered handlers\n");
529 
530 	return (ret);
531 }
532 
533 /*
534  * Remove a specific fault injection handler.
535  */
536 static int
537 cancel_handler(int id)
538 {
539 	zfs_cmd_t zc = {"\0"};
540 
541 	zc.zc_guid = (uint64_t)id;
542 
543 	if (zfs_ioctl(g_zfs, ZFS_IOC_CLEAR_FAULT, &zc) != 0) {
544 		(void) fprintf(stderr, "failed to remove handler %d: %s\n",
545 		    id, strerror(errno));
546 		return (1);
547 	}
548 
549 	(void) printf("removed handler %d\n", id);
550 
551 	return (0);
552 }
553 
554 /*
555  * Register a new fault injection handler.
556  */
557 static int
558 register_handler(const char *pool, int flags, zinject_record_t *record,
559     int quiet)
560 {
561 	zfs_cmd_t zc = {"\0"};
562 
563 	(void) strlcpy(zc.zc_name, pool, sizeof (zc.zc_name));
564 	zc.zc_inject_record = *record;
565 	zc.zc_guid = flags;
566 
567 	if (zfs_ioctl(g_zfs, ZFS_IOC_INJECT_FAULT, &zc) != 0) {
568 		(void) fprintf(stderr, "failed to add handler: %s\n",
569 		    errno == EDOM ? "block level exceeds max level of object" :
570 		    strerror(errno));
571 		return (1);
572 	}
573 
574 	if (flags & ZINJECT_NULL)
575 		return (0);
576 
577 	if (quiet) {
578 		(void) printf("%llu\n", (u_longlong_t)zc.zc_guid);
579 	} else {
580 		(void) printf("Added handler %llu with the following "
581 		    "properties:\n", (u_longlong_t)zc.zc_guid);
582 		(void) printf("  pool: %s\n", pool);
583 		if (record->zi_guid) {
584 			(void) printf("  vdev: %llx\n",
585 			    (u_longlong_t)record->zi_guid);
586 		} else if (record->zi_func[0] != '\0') {
587 			(void) printf("  panic function: %s\n",
588 			    record->zi_func);
589 		} else if (record->zi_duration > 0) {
590 			(void) printf(" time: %lld seconds\n",
591 			    (u_longlong_t)record->zi_duration);
592 		} else if (record->zi_duration < 0) {
593 			(void) printf(" txgs: %lld \n",
594 			    (u_longlong_t)-record->zi_duration);
595 		} else {
596 			(void) printf("objset: %llu\n",
597 			    (u_longlong_t)record->zi_objset);
598 			(void) printf("object: %llu\n",
599 			    (u_longlong_t)record->zi_object);
600 			(void) printf("  type: %llu\n",
601 			    (u_longlong_t)record->zi_type);
602 			(void) printf(" level: %d\n", record->zi_level);
603 			if (record->zi_start == 0 &&
604 			    record->zi_end == -1ULL)
605 				(void) printf(" range: all\n");
606 			else
607 				(void) printf(" range: [%llu, %llu)\n",
608 				    (u_longlong_t)record->zi_start,
609 				    (u_longlong_t)record->zi_end);
610 			(void) printf("  dvas: 0x%x\n", record->zi_dvas);
611 		}
612 	}
613 
614 	return (0);
615 }
616 
617 static int
618 perform_action(const char *pool, zinject_record_t *record, int cmd)
619 {
620 	zfs_cmd_t zc = {"\0"};
621 
622 	ASSERT(cmd == VDEV_STATE_DEGRADED || cmd == VDEV_STATE_FAULTED);
623 	(void) strlcpy(zc.zc_name, pool, sizeof (zc.zc_name));
624 	zc.zc_guid = record->zi_guid;
625 	zc.zc_cookie = cmd;
626 
627 	if (zfs_ioctl(g_zfs, ZFS_IOC_VDEV_SET_STATE, &zc) == 0)
628 		return (0);
629 
630 	return (1);
631 }
632 
633 static int
634 parse_delay(char *str, uint64_t *delay, uint64_t *nlanes)
635 {
636 	unsigned long scan_delay;
637 	unsigned long scan_nlanes;
638 
639 	if (sscanf(str, "%lu:%lu", &scan_delay, &scan_nlanes) != 2)
640 		return (1);
641 
642 	/*
643 	 * We explicitly disallow a delay of zero here, because we key
644 	 * off this value being non-zero in translate_device(), to
645 	 * determine if the fault is a ZINJECT_DELAY_IO fault or not.
646 	 */
647 	if (scan_delay == 0)
648 		return (1);
649 
650 	/*
651 	 * The units for the CLI delay parameter is milliseconds, but
652 	 * the data passed to the kernel is interpreted as nanoseconds.
653 	 * Thus we scale the milliseconds to nanoseconds here, and this
654 	 * nanosecond value is used to pass the delay to the kernel.
655 	 */
656 	*delay = MSEC2NSEC(scan_delay);
657 	*nlanes = scan_nlanes;
658 
659 	return (0);
660 }
661 
662 static int
663 parse_frequency(const char *str, uint32_t *percent)
664 {
665 	double val;
666 	char *post;
667 
668 	val = strtod(str, &post);
669 	if (post == NULL || *post != '\0')
670 		return (EINVAL);
671 
672 	/* valid range is [0.0001, 100.0] */
673 	val /= 100.0f;
674 	if (val < 0.000001f || val > 1.0f)
675 		return (ERANGE);
676 
677 	/* convert to an integer for use by kernel */
678 	*percent = ((uint32_t)(val * ZI_PERCENTAGE_MAX));
679 
680 	return (0);
681 }
682 
683 /*
684  * This function converts a string specifier for DVAs into a bit mask.
685  * The dva's provided by the user should be 0 indexed and separated by
686  * a comma. For example:
687  *	"1"	-> 0b0010  (0x2)
688  *	"0,1"	-> 0b0011  (0x3)
689  *	"0,1,2"	-> 0b0111  (0x7)
690  */
691 static int
692 parse_dvas(const char *str, uint32_t *dvas_out)
693 {
694 	const char *c = str;
695 	uint32_t mask = 0;
696 	boolean_t need_delim = B_FALSE;
697 
698 	/* max string length is 5 ("0,1,2") */
699 	if (strlen(str) > 5 || strlen(str) == 0)
700 		return (EINVAL);
701 
702 	while (*c != '\0') {
703 		switch (*c) {
704 		case '0':
705 		case '1':
706 		case '2':
707 			/* check for pipe between DVAs */
708 			if (need_delim)
709 				return (EINVAL);
710 
711 			/* check if this DVA has been set already */
712 			if (mask & (1 << ((*c) - '0')))
713 				return (EINVAL);
714 
715 			mask |= (1 << ((*c) - '0'));
716 			need_delim = B_TRUE;
717 			break;
718 		case ',':
719 			need_delim = B_FALSE;
720 			break;
721 		default:
722 			/* check for invalid character */
723 			return (EINVAL);
724 		}
725 		c++;
726 	}
727 
728 	/* check for dangling delimiter */
729 	if (!need_delim)
730 		return (EINVAL);
731 
732 	*dvas_out = mask;
733 	return (0);
734 }
735 
736 int
737 main(int argc, char **argv)
738 {
739 	int c;
740 	char *range = NULL;
741 	char *cancel = NULL;
742 	char *end;
743 	char *raw = NULL;
744 	char *device = NULL;
745 	int level = 0;
746 	int quiet = 0;
747 	int error = 0;
748 	int domount = 0;
749 	int io_type = ZIO_TYPES;
750 	int action = VDEV_STATE_UNKNOWN;
751 	err_type_t type = TYPE_INVAL;
752 	err_type_t label = TYPE_INVAL;
753 	zinject_record_t record = { 0 };
754 	char pool[MAXNAMELEN] = "";
755 	char dataset[MAXNAMELEN] = "";
756 	zfs_handle_t *zhp = NULL;
757 	int nowrites = 0;
758 	int dur_txg = 0;
759 	int dur_secs = 0;
760 	int ret;
761 	int flags = 0;
762 	uint32_t dvas = 0;
763 
764 	if ((g_zfs = libzfs_init()) == NULL) {
765 		(void) fprintf(stderr, "%s\n", libzfs_error_init(errno));
766 		return (1);
767 	}
768 
769 	libzfs_print_on_error(g_zfs, B_TRUE);
770 
771 	if ((zfs_fd = open(ZFS_DEV, O_RDWR)) < 0) {
772 		(void) fprintf(stderr, "failed to open ZFS device\n");
773 		libzfs_fini(g_zfs);
774 		return (1);
775 	}
776 
777 	if (argc == 1) {
778 		/*
779 		 * No arguments.  Print the available handlers.  If there are no
780 		 * available handlers, direct the user to '-h' for help
781 		 * information.
782 		 */
783 		if (print_all_handlers() == 0) {
784 			(void) printf("No handlers registered.\n");
785 			(void) printf("Run 'zinject -h' for usage "
786 			    "information.\n");
787 		}
788 		libzfs_fini(g_zfs);
789 		return (0);
790 	}
791 
792 	while ((c = getopt(argc, argv,
793 	    ":aA:b:C:d:D:f:Fg:qhIc:t:T:l:mr:s:e:uL:p:")) != -1) {
794 		switch (c) {
795 		case 'a':
796 			flags |= ZINJECT_FLUSH_ARC;
797 			break;
798 		case 'A':
799 			if (strcasecmp(optarg, "degrade") == 0) {
800 				action = VDEV_STATE_DEGRADED;
801 			} else if (strcasecmp(optarg, "fault") == 0) {
802 				action = VDEV_STATE_FAULTED;
803 			} else {
804 				(void) fprintf(stderr, "invalid action '%s': "
805 				    "must be 'degrade' or 'fault'\n", optarg);
806 				usage();
807 				libzfs_fini(g_zfs);
808 				return (1);
809 			}
810 			break;
811 		case 'b':
812 			raw = optarg;
813 			break;
814 		case 'c':
815 			cancel = optarg;
816 			break;
817 		case 'C':
818 			ret = parse_dvas(optarg, &dvas);
819 			if (ret != 0) {
820 				(void) fprintf(stderr, "invalid DVA list '%s': "
821 				    "DVAs should be 0 indexed and separated by "
822 				    "commas.\n", optarg);
823 				usage();
824 				libzfs_fini(g_zfs);
825 				return (1);
826 			}
827 			break;
828 		case 'd':
829 			device = optarg;
830 			break;
831 		case 'D':
832 			errno = 0;
833 			ret = parse_delay(optarg, &record.zi_timer,
834 			    &record.zi_nlanes);
835 			if (ret != 0) {
836 
837 				(void) fprintf(stderr, "invalid i/o delay "
838 				    "value: '%s'\n", optarg);
839 				usage();
840 				libzfs_fini(g_zfs);
841 				return (1);
842 			}
843 			break;
844 		case 'e':
845 			if (strcasecmp(optarg, "io") == 0) {
846 				error = EIO;
847 			} else if (strcasecmp(optarg, "checksum") == 0) {
848 				error = ECKSUM;
849 			} else if (strcasecmp(optarg, "decompress") == 0) {
850 				error = EINVAL;
851 			} else if (strcasecmp(optarg, "decrypt") == 0) {
852 				error = EACCES;
853 			} else if (strcasecmp(optarg, "nxio") == 0) {
854 				error = ENXIO;
855 			} else if (strcasecmp(optarg, "dtl") == 0) {
856 				error = ECHILD;
857 			} else if (strcasecmp(optarg, "corrupt") == 0) {
858 				error = EILSEQ;
859 			} else {
860 				(void) fprintf(stderr, "invalid error type "
861 				    "'%s': must be 'io', 'checksum' or "
862 				    "'nxio'\n", optarg);
863 				usage();
864 				libzfs_fini(g_zfs);
865 				return (1);
866 			}
867 			break;
868 		case 'f':
869 			ret = parse_frequency(optarg, &record.zi_freq);
870 			if (ret != 0) {
871 				(void) fprintf(stderr, "%sfrequency value must "
872 				    "be in the range [0.0001, 100.0]\n",
873 				    ret == EINVAL ? "invalid value: " :
874 				    ret == ERANGE ? "out of range: " : "");
875 				libzfs_fini(g_zfs);
876 				return (1);
877 			}
878 			break;
879 		case 'F':
880 			record.zi_failfast = B_TRUE;
881 			break;
882 		case 'g':
883 			dur_txg = 1;
884 			record.zi_duration = (int)strtol(optarg, &end, 10);
885 			if (record.zi_duration <= 0 || *end != '\0') {
886 				(void) fprintf(stderr, "invalid duration '%s': "
887 				    "must be a positive integer\n", optarg);
888 				usage();
889 				libzfs_fini(g_zfs);
890 				return (1);
891 			}
892 			/* store duration of txgs as its negative */
893 			record.zi_duration *= -1;
894 			break;
895 		case 'h':
896 			usage();
897 			libzfs_fini(g_zfs);
898 			return (0);
899 		case 'I':
900 			/* default duration, if one hasn't yet been defined */
901 			nowrites = 1;
902 			if (dur_secs == 0 && dur_txg == 0)
903 				record.zi_duration = 30;
904 			break;
905 		case 'l':
906 			level = (int)strtol(optarg, &end, 10);
907 			if (*end != '\0') {
908 				(void) fprintf(stderr, "invalid level '%s': "
909 				    "must be an integer\n", optarg);
910 				usage();
911 				libzfs_fini(g_zfs);
912 				return (1);
913 			}
914 			break;
915 		case 'm':
916 			domount = 1;
917 			break;
918 		case 'p':
919 			(void) strlcpy(record.zi_func, optarg,
920 			    sizeof (record.zi_func));
921 			record.zi_cmd = ZINJECT_PANIC;
922 			break;
923 		case 'q':
924 			quiet = 1;
925 			break;
926 		case 'r':
927 			range = optarg;
928 			flags |= ZINJECT_CALC_RANGE;
929 			break;
930 		case 's':
931 			dur_secs = 1;
932 			record.zi_duration = (int)strtol(optarg, &end, 10);
933 			if (record.zi_duration <= 0 || *end != '\0') {
934 				(void) fprintf(stderr, "invalid duration '%s': "
935 				    "must be a positive integer\n", optarg);
936 				usage();
937 				libzfs_fini(g_zfs);
938 				return (1);
939 			}
940 			break;
941 		case 'T':
942 			if (strcasecmp(optarg, "read") == 0) {
943 				io_type = ZIO_TYPE_READ;
944 			} else if (strcasecmp(optarg, "write") == 0) {
945 				io_type = ZIO_TYPE_WRITE;
946 			} else if (strcasecmp(optarg, "free") == 0) {
947 				io_type = ZIO_TYPE_FREE;
948 			} else if (strcasecmp(optarg, "claim") == 0) {
949 				io_type = ZIO_TYPE_CLAIM;
950 			} else if (strcasecmp(optarg, "all") == 0) {
951 				io_type = ZIO_TYPES;
952 			} else {
953 				(void) fprintf(stderr, "invalid I/O type "
954 				    "'%s': must be 'read', 'write', 'free', "
955 				    "'claim' or 'all'\n", optarg);
956 				usage();
957 				libzfs_fini(g_zfs);
958 				return (1);
959 			}
960 			break;
961 		case 't':
962 			if ((type = name_to_type(optarg)) == TYPE_INVAL &&
963 			    !MOS_TYPE(type)) {
964 				(void) fprintf(stderr, "invalid type '%s'\n",
965 				    optarg);
966 				usage();
967 				libzfs_fini(g_zfs);
968 				return (1);
969 			}
970 			break;
971 		case 'u':
972 			flags |= ZINJECT_UNLOAD_SPA;
973 			break;
974 		case 'L':
975 			if ((label = name_to_type(optarg)) == TYPE_INVAL &&
976 			    !LABEL_TYPE(type)) {
977 				(void) fprintf(stderr, "invalid label type "
978 				    "'%s'\n", optarg);
979 				usage();
980 				libzfs_fini(g_zfs);
981 				return (1);
982 			}
983 			break;
984 		case ':':
985 			(void) fprintf(stderr, "option -%c requires an "
986 			    "operand\n", optopt);
987 			usage();
988 			libzfs_fini(g_zfs);
989 			return (1);
990 		case '?':
991 			(void) fprintf(stderr, "invalid option '%c'\n",
992 			    optopt);
993 			usage();
994 			libzfs_fini(g_zfs);
995 			return (2);
996 		}
997 	}
998 
999 	argc -= optind;
1000 	argv += optind;
1001 
1002 	if (record.zi_duration != 0)
1003 		record.zi_cmd = ZINJECT_IGNORED_WRITES;
1004 
1005 	if (cancel != NULL) {
1006 		/*
1007 		 * '-c' is invalid with any other options.
1008 		 */
1009 		if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1010 		    level != 0 || record.zi_cmd != ZINJECT_UNINITIALIZED ||
1011 		    record.zi_freq > 0 || dvas != 0) {
1012 			(void) fprintf(stderr, "cancel (-c) incompatible with "
1013 			    "any other options\n");
1014 			usage();
1015 			libzfs_fini(g_zfs);
1016 			return (2);
1017 		}
1018 		if (argc != 0) {
1019 			(void) fprintf(stderr, "extraneous argument to '-c'\n");
1020 			usage();
1021 			libzfs_fini(g_zfs);
1022 			return (2);
1023 		}
1024 
1025 		if (strcmp(cancel, "all") == 0) {
1026 			return (cancel_all_handlers());
1027 		} else {
1028 			int id = (int)strtol(cancel, &end, 10);
1029 			if (*end != '\0') {
1030 				(void) fprintf(stderr, "invalid handle id '%s':"
1031 				    " must be an integer or 'all'\n", cancel);
1032 				usage();
1033 				libzfs_fini(g_zfs);
1034 				return (1);
1035 			}
1036 			return (cancel_handler(id));
1037 		}
1038 	}
1039 
1040 	if (device != NULL) {
1041 		/*
1042 		 * Device (-d) injection uses a completely different mechanism
1043 		 * for doing injection, so handle it separately here.
1044 		 */
1045 		if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1046 		    level != 0 || record.zi_cmd != ZINJECT_UNINITIALIZED ||
1047 		    dvas != 0) {
1048 			(void) fprintf(stderr, "device (-d) incompatible with "
1049 			    "data error injection\n");
1050 			usage();
1051 			libzfs_fini(g_zfs);
1052 			return (2);
1053 		}
1054 
1055 		if (argc != 1) {
1056 			(void) fprintf(stderr, "device (-d) injection requires "
1057 			    "a single pool name\n");
1058 			usage();
1059 			libzfs_fini(g_zfs);
1060 			return (2);
1061 		}
1062 
1063 		(void) strlcpy(pool, argv[0], sizeof (pool));
1064 		dataset[0] = '\0';
1065 
1066 		if (error == ECKSUM) {
1067 			(void) fprintf(stderr, "device error type must be "
1068 			    "'io', 'nxio' or 'corrupt'\n");
1069 			libzfs_fini(g_zfs);
1070 			return (1);
1071 		}
1072 
1073 		if (error == EILSEQ &&
1074 		    (record.zi_freq == 0 || io_type != ZIO_TYPE_READ)) {
1075 			(void) fprintf(stderr, "device corrupt errors require "
1076 			    "io type read and a frequency value\n");
1077 			libzfs_fini(g_zfs);
1078 			return (1);
1079 		}
1080 
1081 		record.zi_iotype = io_type;
1082 		if (translate_device(pool, device, label, &record) != 0) {
1083 			libzfs_fini(g_zfs);
1084 			return (1);
1085 		}
1086 
1087 		if (record.zi_nlanes) {
1088 			switch (io_type) {
1089 			case ZIO_TYPE_READ:
1090 			case ZIO_TYPE_WRITE:
1091 			case ZIO_TYPES:
1092 				break;
1093 			default:
1094 				(void) fprintf(stderr, "I/O type for a delay "
1095 				    "must be 'read' or 'write'\n");
1096 				usage();
1097 				libzfs_fini(g_zfs);
1098 				return (1);
1099 			}
1100 		}
1101 
1102 		if (!error)
1103 			error = ENXIO;
1104 
1105 		if (action != VDEV_STATE_UNKNOWN)
1106 			return (perform_action(pool, &record, action));
1107 
1108 	} else if (raw != NULL) {
1109 		if (range != NULL || type != TYPE_INVAL || level != 0 ||
1110 		    record.zi_cmd != ZINJECT_UNINITIALIZED ||
1111 		    record.zi_freq > 0 || dvas != 0) {
1112 			(void) fprintf(stderr, "raw (-b) format with "
1113 			    "any other options\n");
1114 			usage();
1115 			libzfs_fini(g_zfs);
1116 			return (2);
1117 		}
1118 
1119 		if (argc != 1) {
1120 			(void) fprintf(stderr, "raw (-b) format expects a "
1121 			    "single pool name\n");
1122 			usage();
1123 			libzfs_fini(g_zfs);
1124 			return (2);
1125 		}
1126 
1127 		(void) strlcpy(pool, argv[0], sizeof (pool));
1128 		dataset[0] = '\0';
1129 
1130 		if (error == ENXIO) {
1131 			(void) fprintf(stderr, "data error type must be "
1132 			    "'checksum' or 'io'\n");
1133 			libzfs_fini(g_zfs);
1134 			return (1);
1135 		}
1136 
1137 		record.zi_cmd = ZINJECT_DATA_FAULT;
1138 		if (translate_raw(raw, &record) != 0) {
1139 			libzfs_fini(g_zfs);
1140 			return (1);
1141 		}
1142 		if (!error)
1143 			error = EIO;
1144 	} else if (record.zi_cmd == ZINJECT_PANIC) {
1145 		if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1146 		    level != 0 || device != NULL || record.zi_freq > 0 ||
1147 		    dvas != 0) {
1148 			(void) fprintf(stderr, "panic (-p) incompatible with "
1149 			    "other options\n");
1150 			usage();
1151 			libzfs_fini(g_zfs);
1152 			return (2);
1153 		}
1154 
1155 		if (argc < 1 || argc > 2) {
1156 			(void) fprintf(stderr, "panic (-p) injection requires "
1157 			    "a single pool name and an optional id\n");
1158 			usage();
1159 			libzfs_fini(g_zfs);
1160 			return (2);
1161 		}
1162 
1163 		(void) strlcpy(pool, argv[0], sizeof (pool));
1164 		if (argv[1] != NULL)
1165 			record.zi_type = atoi(argv[1]);
1166 		dataset[0] = '\0';
1167 	} else if (record.zi_cmd == ZINJECT_IGNORED_WRITES) {
1168 		if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1169 		    level != 0 || record.zi_freq > 0 || dvas != 0) {
1170 			(void) fprintf(stderr, "hardware failure (-I) "
1171 			    "incompatible with other options\n");
1172 			usage();
1173 			libzfs_fini(g_zfs);
1174 			return (2);
1175 		}
1176 
1177 		if (nowrites == 0) {
1178 			(void) fprintf(stderr, "-s or -g meaningless "
1179 			    "without -I (ignore writes)\n");
1180 			usage();
1181 			libzfs_fini(g_zfs);
1182 			return (2);
1183 		} else if (dur_secs && dur_txg) {
1184 			(void) fprintf(stderr, "choose a duration either "
1185 			    "in seconds (-s) or a number of txgs (-g) "
1186 			    "but not both\n");
1187 			usage();
1188 			libzfs_fini(g_zfs);
1189 			return (2);
1190 		} else if (argc != 1) {
1191 			(void) fprintf(stderr, "ignore writes (-I) "
1192 			    "injection requires a single pool name\n");
1193 			usage();
1194 			libzfs_fini(g_zfs);
1195 			return (2);
1196 		}
1197 
1198 		(void) strlcpy(pool, argv[0], sizeof (pool));
1199 		dataset[0] = '\0';
1200 	} else if (type == TYPE_INVAL) {
1201 		if (flags == 0) {
1202 			(void) fprintf(stderr, "at least one of '-b', '-d', "
1203 			    "'-t', '-a', '-p', '-I' or '-u' "
1204 			    "must be specified\n");
1205 			usage();
1206 			libzfs_fini(g_zfs);
1207 			return (2);
1208 		}
1209 
1210 		if (argc == 1 && (flags & ZINJECT_UNLOAD_SPA)) {
1211 			(void) strlcpy(pool, argv[0], sizeof (pool));
1212 			dataset[0] = '\0';
1213 		} else if (argc != 0) {
1214 			(void) fprintf(stderr, "extraneous argument for "
1215 			    "'-f'\n");
1216 			usage();
1217 			libzfs_fini(g_zfs);
1218 			return (2);
1219 		}
1220 
1221 		flags |= ZINJECT_NULL;
1222 	} else {
1223 		if (argc != 1) {
1224 			(void) fprintf(stderr, "missing object\n");
1225 			usage();
1226 			libzfs_fini(g_zfs);
1227 			return (2);
1228 		}
1229 
1230 		if (error == ENXIO || error == EILSEQ) {
1231 			(void) fprintf(stderr, "data error type must be "
1232 			    "'checksum' or 'io'\n");
1233 			libzfs_fini(g_zfs);
1234 			return (1);
1235 		}
1236 
1237 		if (dvas != 0) {
1238 			if (error == EACCES || error == EINVAL) {
1239 				(void) fprintf(stderr, "the '-C' option may "
1240 				    "not be used with logical data errors "
1241 				    "'decrypt' and 'decompress'\n");
1242 				libzfs_fini(g_zfs);
1243 				return (1);
1244 			}
1245 
1246 			record.zi_dvas = dvas;
1247 		}
1248 
1249 		if (error == EACCES) {
1250 			if (type != TYPE_DATA) {
1251 				(void) fprintf(stderr, "decryption errors "
1252 				    "may only be injected for 'data' types\n");
1253 				libzfs_fini(g_zfs);
1254 				return (1);
1255 			}
1256 
1257 			record.zi_cmd = ZINJECT_DECRYPT_FAULT;
1258 			/*
1259 			 * Internally, ZFS actually uses ECKSUM for decryption
1260 			 * errors since EACCES is used to indicate the key was
1261 			 * not found.
1262 			 */
1263 			error = ECKSUM;
1264 		} else {
1265 			record.zi_cmd = ZINJECT_DATA_FAULT;
1266 		}
1267 
1268 		if (translate_record(type, argv[0], range, level, &record, pool,
1269 		    dataset) != 0) {
1270 			libzfs_fini(g_zfs);
1271 			return (1);
1272 		}
1273 		if (!error)
1274 			error = EIO;
1275 	}
1276 
1277 	/*
1278 	 * If this is pool-wide metadata, unmount everything.  The ioctl() will
1279 	 * unload the pool, so that we trigger spa-wide reopen of metadata next
1280 	 * time we access the pool.
1281 	 */
1282 	if (dataset[0] != '\0' && domount) {
1283 		if ((zhp = zfs_open(g_zfs, dataset,
1284 		    ZFS_TYPE_DATASET)) == NULL) {
1285 			libzfs_fini(g_zfs);
1286 			return (1);
1287 		}
1288 		if (zfs_unmount(zhp, NULL, 0) != 0) {
1289 			libzfs_fini(g_zfs);
1290 			return (1);
1291 		}
1292 	}
1293 
1294 	record.zi_error = error;
1295 
1296 	ret = register_handler(pool, flags, &record, quiet);
1297 
1298 	if (dataset[0] != '\0' && domount)
1299 		ret = (zfs_mount(zhp, NULL, 0) != 0);
1300 
1301 	libzfs_fini(g_zfs);
1302 
1303 	return (ret);
1304 }
1305