zdb: show dedup table and log attributes
[zfs.git] / cmd / zinject / zinject.c
blobed60cce3dd164ea3c080954ac864f9b127e70df4
1 /*
2 * CDDL HEADER START
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.
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.
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]
19 * CDDL HEADER END
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 * Copyright (c) 2023-2024, Klara Inc.
29 * ZFS Fault Injector
31 * This userland component takes a set of options and uses libzpool to translate
32 * from a user-visible object type and name to an internal representation.
33 * There are two basic types of faults: device faults and data faults.
36 * DEVICE FAULTS
38 * Errors can be injected into a particular vdev using the '-d' option. This
39 * option takes a path or vdev GUID to uniquely identify the device within a
40 * pool. There are four types of errors that can be injected, IO, ENXIO,
41 * ECHILD, and EILSEQ. These can be controlled through the '-e' option and the
42 * default is ENXIO. For EIO failures, any attempt to read data from the device
43 * will return EIO, but a subsequent attempt to reopen the device will succeed.
44 * For ENXIO failures, any attempt to read from the device will return EIO, but
45 * any attempt to reopen the device will also return ENXIO. The EILSEQ failures
46 * only apply to read operations (-T read) and will flip a bit after the device
47 * has read the original data.
49 * For label faults, the -L option must be specified. This allows faults
50 * to be injected into either the nvlist, uberblock, pad1, or pad2 region
51 * of all the labels for the specified device.
53 * This form of the command looks like:
55 * zinject -d device [-e errno] [-L <uber | nvlist | pad1 | pad2>] pool
58 * DATA FAULTS
60 * We begin with a tuple of the form:
62 * <type,level,range,object>
64 * type A string describing the type of data to target. Each type
65 * implicitly describes how to interpret 'object'. Currently,
66 * the following values are supported:
68 * data User data for a file
69 * dnode Dnode for a file or directory
71 * The following MOS objects are special. Instead of injecting
72 * errors on a particular object or blkid, we inject errors across
73 * all objects of the given type.
75 * mos Any data in the MOS
76 * mosdir object directory
77 * config pool configuration
78 * bpobj blkptr list
79 * spacemap spacemap
80 * metaslab metaslab
81 * errlog persistent error log
83 * level Object level. Defaults to '0', not applicable to all types. If
84 * a range is given, this corresponds to the indirect block
85 * corresponding to the specific range.
87 * range A numerical range [start,end) within the object. Defaults to
88 * the full size of the file.
90 * object A string describing the logical location of the object. For
91 * files and directories (currently the only supported types),
92 * this is the path of the object on disk.
94 * This is translated, via libzpool, into the following internal representation:
96 * <type,objset,object,level,range>
98 * These types should be self-explanatory. This tuple is then passed to the
99 * kernel via a special ioctl() to initiate fault injection for the given
100 * object. Note that 'type' is not strictly necessary for fault injection, but
101 * is used when translating existing faults into a human-readable string.
104 * The command itself takes one of the forms:
106 * zinject
107 * zinject <-a | -u pool>
108 * zinject -c <id|all>
109 * zinject [-q] <-t type> [-f freq] [-u] [-a] [-m] [-e errno] [-l level]
110 * [-r range] <object>
111 * zinject [-f freq] [-a] [-m] [-u] -b objset:object:level:start:end pool
113 * With no arguments, the command prints all currently registered injection
114 * handlers, with their numeric identifiers.
116 * The '-c' option will clear the given handler, or all handlers if 'all' is
117 * specified.
119 * The '-e' option takes a string describing the errno to simulate. This must
120 * be one of 'io', 'checksum', 'decompress', or 'decrypt'. In most cases this
121 * will result in the same behavior, but RAID-Z will produce a different set of
122 * ereports for this situation.
124 * The '-a', '-u', and '-m' flags toggle internal flush behavior. If '-a' is
125 * specified, then the ARC cache is flushed appropriately. If '-u' is
126 * specified, then the underlying SPA is unloaded. Either of these flags can be
127 * specified independently of any other handlers. The '-m' flag automatically
128 * does an unmount and remount of the underlying dataset to aid in flushing the
129 * cache.
131 * The '-f' flag controls the frequency of errors injected, expressed as a
132 * real number percentage between 0.0001 and 100. The default is 100.
134 * The this form is responsible for actually injecting the handler into the
135 * framework. It takes the arguments described above, translates them to the
136 * internal tuple using libzpool, and then issues an ioctl() to register the
137 * handler.
139 * The final form can target a specific bookmark, regardless of whether a
140 * human-readable interface has been designed. It allows developers to specify
141 * a particular block by number.
144 #include <errno.h>
145 #include <fcntl.h>
146 #include <stdio.h>
147 #include <stdlib.h>
148 #include <string.h>
149 #include <strings.h>
150 #include <unistd.h>
152 #include <sys/fs/zfs.h>
153 #include <sys/mount.h>
155 #include <libzfs.h>
157 #undef verify /* both libzfs.h and zfs_context.h want to define this */
159 #include "zinject.h"
161 libzfs_handle_t *g_zfs;
162 int zfs_fd;
164 static const char *const 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"
180 static err_type_t
181 name_to_type(const char *arg)
183 int i;
184 for (i = 0; i < TYPE_INVAL; i++)
185 if (strcmp(errtable[i], arg) == 0)
186 return (i);
188 return (TYPE_INVAL);
191 static const char *
192 type_to_name(uint64_t type)
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 ("-");
212 struct errstr {
213 int err;
214 const char *str;
216 static const struct errstr errstrtable[] = {
217 { EIO, "io" },
218 { ECKSUM, "checksum" },
219 { EINVAL, "decompress" },
220 { EACCES, "decrypt" },
221 { ENXIO, "nxio" },
222 { ECHILD, "dtl" },
223 { EILSEQ, "corrupt" },
224 { ENOSYS, "noop" },
225 { 0, NULL },
228 static int
229 str_to_err(const char *str)
231 for (int i = 0; errstrtable[i].str != NULL; i++)
232 if (strcasecmp(errstrtable[i].str, str) == 0)
233 return (errstrtable[i].err);
234 return (-1);
236 static const char *
237 err_to_str(int err)
239 for (int i = 0; errstrtable[i].str != NULL; i++)
240 if (errstrtable[i].err == err)
241 return (errstrtable[i].str);
242 return ("[unknown]");
246 * Print usage message.
248 void
249 usage(void)
251 (void) printf(
252 "usage:\n"
253 "\n"
254 "\tzinject\n"
255 "\n"
256 "\t\tList all active injection records.\n"
257 "\n"
258 "\tzinject -c <id|all>\n"
259 "\n"
260 "\t\tClear the particular record (if given a numeric ID), or\n"
261 "\t\tall records if 'all' is specified.\n"
262 "\n"
263 "\tzinject -p <function name> pool\n"
264 "\t\tInject a panic fault at the specified function. Only \n"
265 "\t\tfunctions which call spa_vdev_config_exit(), or \n"
266 "\t\tspa_vdev_exit() will trigger a panic.\n"
267 "\n"
268 "\tzinject -d device [-e errno] [-L <nvlist|uber|pad1|pad2>] [-F]\n"
269 "\t\t[-T <read|write|free|claim|flush|all>] [-f frequency] pool\n\n"
270 "\t\tInject a fault into a particular device or the device's\n"
271 "\t\tlabel. Label injection can either be 'nvlist', 'uber',\n "
272 "\t\t'pad1', or 'pad2'.\n"
273 "\t\t'errno' can be 'nxio' (the default), 'io', 'dtl',\n"
274 "\t\t'corrupt' (bit flip), or 'noop' (successfully do nothing).\n"
275 "\t\t'frequency' is a value between 0.0001 and 100.0 that limits\n"
276 "\t\tdevice error injection to a percentage of the IOs.\n"
277 "\n"
278 "\tzinject -d device -A <degrade|fault> -D <delay secs> pool\n"
279 "\t\tPerform a specific action on a particular device.\n"
280 "\n"
281 "\tzinject -d device -D latency:lanes pool\n"
282 "\n"
283 "\t\tAdd an artificial delay to IO requests on a particular\n"
284 "\t\tdevice, such that the requests take a minimum of 'latency'\n"
285 "\t\tmilliseconds to complete. Each delay has an associated\n"
286 "\t\tnumber of 'lanes' which defines the number of concurrent\n"
287 "\t\tIO requests that can be processed.\n"
288 "\n"
289 "\t\tFor example, with a single lane delay of 10 ms (-D 10:1),\n"
290 "\t\tthe device will only be able to service a single IO request\n"
291 "\t\tat a time with each request taking 10 ms to complete. So,\n"
292 "\t\tif only a single request is submitted every 10 ms, the\n"
293 "\t\taverage latency will be 10 ms; but if more than one request\n"
294 "\t\tis submitted every 10 ms, the average latency will be more\n"
295 "\t\tthan 10 ms.\n"
296 "\n"
297 "\t\tSimilarly, if a delay of 10 ms is specified to have two\n"
298 "\t\tlanes (-D 10:2), then the device will be able to service\n"
299 "\t\ttwo requests at a time, each with a minimum latency of\n"
300 "\t\t10 ms. So, if two requests are submitted every 10 ms, then\n"
301 "\t\tthe average latency will be 10 ms; but if more than two\n"
302 "\t\trequests are submitted every 10 ms, the average latency\n"
303 "\t\twill be more than 10 ms.\n"
304 "\n"
305 "\t\tAlso note, these delays are additive. So two invocations\n"
306 "\t\tof '-D 10:1', is roughly equivalent to a single invocation\n"
307 "\t\tof '-D 10:2'. This also means, one can specify multiple\n"
308 "\t\tlanes with differing target latencies. For example, an\n"
309 "\t\tinvocation of '-D 10:1' followed by '-D 25:2' will\n"
310 "\t\tcreate 3 lanes on the device; one lane with a latency\n"
311 "\t\tof 10 ms and two lanes with a 25 ms latency.\n"
312 "\n"
313 "\tzinject -P import|export -s <seconds> pool\n"
314 "\t\tAdd an artificial delay to a future pool import or export,\n"
315 "\t\tsuch that the operation takes a minimum of supplied seconds\n"
316 "\t\tto complete.\n"
317 "\n"
318 "\tzinject -I [-s <seconds> | -g <txgs>] pool\n"
319 "\t\tCause the pool to stop writing blocks yet not\n"
320 "\t\treport errors for a duration. Simulates buggy hardware\n"
321 "\t\tthat fails to honor cache flush requests.\n"
322 "\t\tDefault duration is 30 seconds. The machine is panicked\n"
323 "\t\tat the end of the duration.\n"
324 "\n"
325 "\tzinject -b objset:object:level:blkid pool\n"
326 "\n"
327 "\t\tInject an error into pool 'pool' with the numeric bookmark\n"
328 "\t\tspecified by the remaining tuple. Each number is in\n"
329 "\t\thexadecimal, and only one block can be specified.\n"
330 "\n"
331 "\tzinject [-q] <-t type> [-C dvas] [-e errno] [-l level]\n"
332 "\t\t[-r range] [-a] [-m] [-u] [-f freq] <object>\n"
333 "\n"
334 "\t\tInject an error into the object specified by the '-t' option\n"
335 "\t\tand the object descriptor. The 'object' parameter is\n"
336 "\t\tinterpreted depending on the '-t' option.\n"
337 "\n"
338 "\t\t-q\tQuiet mode. Only print out the handler number added.\n"
339 "\t\t-e\tInject a specific error. Must be one of 'io',\n"
340 "\t\t\t'checksum', 'decompress', or 'decrypt'. Default is 'io'.\n"
341 "\t\t-C\tInject the given error only into specific DVAs. The\n"
342 "\t\t\tDVAs should be specified as a list of 0-indexed DVAs\n"
343 "\t\t\tseparated by commas (ex. '0,2').\n"
344 "\t\t-l\tInject error at a particular block level. Default is "
345 "0.\n"
346 "\t\t-m\tAutomatically remount underlying filesystem.\n"
347 "\t\t-r\tInject error over a particular logical range of an\n"
348 "\t\t\tobject. Will be translated to the appropriate blkid\n"
349 "\t\t\trange according to the object's properties.\n"
350 "\t\t-a\tFlush the ARC cache. Can be specified without any\n"
351 "\t\t\tassociated object.\n"
352 "\t\t-u\tUnload the associated pool. Can be specified with only\n"
353 "\t\t\ta pool object.\n"
354 "\t\t-f\tOnly inject errors a fraction of the time. Expressed as\n"
355 "\t\t\ta percentage between 0.0001 and 100.\n"
356 "\n"
357 "\t-t data\t\tInject an error into the plain file contents of a\n"
358 "\t\t\tfile. The object must be specified as a complete path\n"
359 "\t\t\tto a file on a ZFS filesystem.\n"
360 "\n"
361 "\t-t dnode\tInject an error into the metadnode in the block\n"
362 "\t\t\tcorresponding to the dnode for a file or directory. The\n"
363 "\t\t\t'-r' option is incompatible with this mode. The object\n"
364 "\t\t\tis specified as a complete path to a file or directory\n"
365 "\t\t\ton a ZFS filesystem.\n"
366 "\n"
367 "\t-t <mos>\tInject errors into the MOS for objects of the given\n"
368 "\t\t\ttype. Valid types are: mos, mosdir, config, bpobj,\n"
369 "\t\t\tspacemap, metaslab, errlog. The only valid <object> is\n"
370 "\t\t\tthe poolname.\n");
373 static int
374 iter_handlers(int (*func)(int, const char *, zinject_record_t *, void *),
375 void *data)
377 zfs_cmd_t zc = {"\0"};
378 int ret;
380 while (zfs_ioctl(g_zfs, ZFS_IOC_INJECT_LIST_NEXT, &zc) == 0)
381 if ((ret = func((int)zc.zc_guid, zc.zc_name,
382 &zc.zc_inject_record, data)) != 0)
383 return (ret);
385 if (errno != ENOENT) {
386 (void) fprintf(stderr, "Unable to list handlers: %s\n",
387 strerror(errno));
388 return (-1);
391 return (0);
394 static int
395 print_data_handler(int id, const char *pool, zinject_record_t *record,
396 void *data)
398 int *count = data;
400 if (record->zi_guid != 0 || record->zi_func[0] != '\0' ||
401 record->zi_duration != 0) {
402 return (0);
405 if (*count == 0) {
406 (void) printf("%3s %-15s %-6s %-6s %-8s %3s %-4s "
407 "%-15s\n", "ID", "POOL", "OBJSET", "OBJECT", "TYPE",
408 "LVL", "DVAs", "RANGE");
409 (void) printf("--- --------------- ------ "
410 "------ -------- --- ---- ---------------\n");
413 *count += 1;
415 (void) printf("%3d %-15s %-6llu %-6llu %-8s %-3d 0x%02x ",
416 id, pool, (u_longlong_t)record->zi_objset,
417 (u_longlong_t)record->zi_object, type_to_name(record->zi_type),
418 record->zi_level, record->zi_dvas);
421 if (record->zi_start == 0 &&
422 record->zi_end == -1ULL)
423 (void) printf("all\n");
424 else
425 (void) printf("[%llu, %llu]\n", (u_longlong_t)record->zi_start,
426 (u_longlong_t)record->zi_end);
428 return (0);
431 static int
432 print_device_handler(int id, const char *pool, zinject_record_t *record,
433 void *data)
435 static const char *iotypestr[] = {
436 "null", "read", "write", "free", "claim", "flush", "trim", "all",
439 int *count = data;
441 if (record->zi_guid == 0 || record->zi_func[0] != '\0')
442 return (0);
444 if (record->zi_cmd == ZINJECT_DELAY_IO)
445 return (0);
447 if (*count == 0) {
448 (void) printf("%3s %-15s %-16s %-5s %-10s %-9s\n",
449 "ID", "POOL", "GUID", "TYPE", "ERROR", "FREQ");
450 (void) printf(
451 "--- --------------- ---------------- "
452 "----- ---------- ---------\n");
455 *count += 1;
457 double freq = record->zi_freq == 0 ? 100.0f :
458 (((double)record->zi_freq) / ZI_PERCENTAGE_MAX) * 100.0f;
460 (void) printf("%3d %-15s %llx %-5s %-10s %8.4f%%\n", id, pool,
461 (u_longlong_t)record->zi_guid, iotypestr[record->zi_iotype],
462 err_to_str(record->zi_error), freq);
464 return (0);
467 static int
468 print_delay_handler(int id, const char *pool, zinject_record_t *record,
469 void *data)
471 int *count = data;
473 if (record->zi_guid == 0 || record->zi_func[0] != '\0')
474 return (0);
476 if (record->zi_cmd != ZINJECT_DELAY_IO)
477 return (0);
479 if (*count == 0) {
480 (void) printf("%3s %-15s %-15s %-15s %s\n",
481 "ID", "POOL", "DELAY (ms)", "LANES", "GUID");
482 (void) printf("--- --------------- --------------- "
483 "--------------- ----------------\n");
486 *count += 1;
488 (void) printf("%3d %-15s %-15llu %-15llu %llx\n", id, pool,
489 (u_longlong_t)NSEC2MSEC(record->zi_timer),
490 (u_longlong_t)record->zi_nlanes,
491 (u_longlong_t)record->zi_guid);
493 return (0);
496 static int
497 print_panic_handler(int id, const char *pool, zinject_record_t *record,
498 void *data)
500 int *count = data;
502 if (record->zi_func[0] == '\0')
503 return (0);
505 if (*count == 0) {
506 (void) printf("%3s %-15s %s\n", "ID", "POOL", "FUNCTION");
507 (void) printf("--- --------------- ----------------\n");
510 *count += 1;
512 (void) printf("%3d %-15s %s\n", id, pool, record->zi_func);
514 return (0);
517 static int
518 print_pool_delay_handler(int id, const char *pool, zinject_record_t *record,
519 void *data)
521 int *count = data;
523 if (record->zi_cmd != ZINJECT_DELAY_IMPORT &&
524 record->zi_cmd != ZINJECT_DELAY_EXPORT) {
525 return (0);
528 if (*count == 0) {
529 (void) printf("%3s %-19s %-11s %s\n",
530 "ID", "POOL", "DELAY (sec)", "COMMAND");
531 (void) printf("--- ------------------- -----------"
532 " -------\n");
535 *count += 1;
537 (void) printf("%3d %-19s %-11llu %s\n",
538 id, pool, (u_longlong_t)record->zi_duration,
539 record->zi_cmd == ZINJECT_DELAY_IMPORT ? "import": "export");
541 return (0);
545 * Print all registered error handlers. Returns the number of handlers
546 * registered.
548 static int
549 print_all_handlers(void)
551 int count = 0, total = 0;
553 (void) iter_handlers(print_device_handler, &count);
554 if (count > 0) {
555 total += count;
556 (void) printf("\n");
557 count = 0;
560 (void) iter_handlers(print_delay_handler, &count);
561 if (count > 0) {
562 total += count;
563 (void) printf("\n");
564 count = 0;
567 (void) iter_handlers(print_data_handler, &count);
568 if (count > 0) {
569 total += count;
570 (void) printf("\n");
571 count = 0;
574 (void) iter_handlers(print_pool_delay_handler, &count);
575 if (count > 0) {
576 total += count;
577 (void) printf("\n");
578 count = 0;
581 (void) iter_handlers(print_panic_handler, &count);
583 return (count + total);
586 static int
587 cancel_one_handler(int id, const char *pool, zinject_record_t *record,
588 void *data)
590 (void) pool, (void) record, (void) data;
591 zfs_cmd_t zc = {"\0"};
593 zc.zc_guid = (uint64_t)id;
595 if (zfs_ioctl(g_zfs, ZFS_IOC_CLEAR_FAULT, &zc) != 0) {
596 (void) fprintf(stderr, "failed to remove handler %d: %s\n",
597 id, strerror(errno));
598 return (1);
601 return (0);
605 * Remove all fault injection handlers.
607 static int
608 cancel_all_handlers(void)
610 int ret = iter_handlers(cancel_one_handler, NULL);
612 if (ret == 0)
613 (void) printf("removed all registered handlers\n");
615 return (ret);
619 * Remove a specific fault injection handler.
621 static int
622 cancel_handler(int id)
624 zfs_cmd_t zc = {"\0"};
626 zc.zc_guid = (uint64_t)id;
628 if (zfs_ioctl(g_zfs, ZFS_IOC_CLEAR_FAULT, &zc) != 0) {
629 (void) fprintf(stderr, "failed to remove handler %d: %s\n",
630 id, strerror(errno));
631 return (1);
634 (void) printf("removed handler %d\n", id);
636 return (0);
640 * Register a new fault injection handler.
642 static int
643 register_handler(const char *pool, int flags, zinject_record_t *record,
644 int quiet)
646 zfs_cmd_t zc = {"\0"};
648 (void) strlcpy(zc.zc_name, pool, sizeof (zc.zc_name));
649 zc.zc_inject_record = *record;
650 zc.zc_guid = flags;
652 if (zfs_ioctl(g_zfs, ZFS_IOC_INJECT_FAULT, &zc) != 0) {
653 const char *errmsg = strerror(errno);
655 switch (errno) {
656 case EDOM:
657 errmsg = "block level exceeds max level of object";
658 break;
659 case EEXIST:
660 if (record->zi_cmd == ZINJECT_DELAY_IMPORT)
661 errmsg = "pool already imported";
662 if (record->zi_cmd == ZINJECT_DELAY_EXPORT)
663 errmsg = "a handler already exists";
664 break;
665 case ENOENT:
666 /* import delay injector running on older zfs module */
667 if (record->zi_cmd == ZINJECT_DELAY_IMPORT)
668 errmsg = "import delay injector not supported";
669 break;
670 default:
671 break;
673 (void) fprintf(stderr, "failed to add handler: %s\n", errmsg);
674 return (1);
677 if (flags & ZINJECT_NULL)
678 return (0);
680 if (quiet) {
681 (void) printf("%llu\n", (u_longlong_t)zc.zc_guid);
682 } else {
683 (void) printf("Added handler %llu with the following "
684 "properties:\n", (u_longlong_t)zc.zc_guid);
685 (void) printf(" pool: %s\n", pool);
686 if (record->zi_guid) {
687 (void) printf(" vdev: %llx\n",
688 (u_longlong_t)record->zi_guid);
689 } else if (record->zi_func[0] != '\0') {
690 (void) printf(" panic function: %s\n",
691 record->zi_func);
692 } else if (record->zi_duration > 0) {
693 (void) printf(" time: %lld seconds\n",
694 (u_longlong_t)record->zi_duration);
695 } else if (record->zi_duration < 0) {
696 (void) printf(" txgs: %lld \n",
697 (u_longlong_t)-record->zi_duration);
698 } else if (record->zi_timer > 0) {
699 (void) printf(" timer: %lld ms\n",
700 (u_longlong_t)NSEC2MSEC(record->zi_timer));
701 } else {
702 (void) printf("objset: %llu\n",
703 (u_longlong_t)record->zi_objset);
704 (void) printf("object: %llu\n",
705 (u_longlong_t)record->zi_object);
706 (void) printf(" type: %llu\n",
707 (u_longlong_t)record->zi_type);
708 (void) printf(" level: %d\n", record->zi_level);
709 if (record->zi_start == 0 &&
710 record->zi_end == -1ULL)
711 (void) printf(" range: all\n");
712 else
713 (void) printf(" range: [%llu, %llu)\n",
714 (u_longlong_t)record->zi_start,
715 (u_longlong_t)record->zi_end);
716 (void) printf(" dvas: 0x%x\n", record->zi_dvas);
720 return (0);
723 static int
724 perform_action(const char *pool, zinject_record_t *record, int cmd)
726 zfs_cmd_t zc = {"\0"};
728 ASSERT(cmd == VDEV_STATE_DEGRADED || cmd == VDEV_STATE_FAULTED);
729 (void) strlcpy(zc.zc_name, pool, sizeof (zc.zc_name));
730 zc.zc_guid = record->zi_guid;
731 zc.zc_cookie = cmd;
733 if (zfs_ioctl(g_zfs, ZFS_IOC_VDEV_SET_STATE, &zc) == 0)
734 return (0);
736 return (1);
739 static int
740 parse_delay(char *str, uint64_t *delay, uint64_t *nlanes)
742 unsigned long scan_delay;
743 unsigned long scan_nlanes;
745 if (sscanf(str, "%lu:%lu", &scan_delay, &scan_nlanes) != 2)
746 return (1);
749 * We explicitly disallow a delay of zero here, because we key
750 * off this value being non-zero in translate_device(), to
751 * determine if the fault is a ZINJECT_DELAY_IO fault or not.
753 if (scan_delay == 0)
754 return (1);
757 * The units for the CLI delay parameter is milliseconds, but
758 * the data passed to the kernel is interpreted as nanoseconds.
759 * Thus we scale the milliseconds to nanoseconds here, and this
760 * nanosecond value is used to pass the delay to the kernel.
762 *delay = MSEC2NSEC(scan_delay);
763 *nlanes = scan_nlanes;
765 return (0);
768 static int
769 parse_frequency(const char *str, uint32_t *percent)
771 double val;
772 char *post;
774 val = strtod(str, &post);
775 if (post == NULL || *post != '\0')
776 return (EINVAL);
778 /* valid range is [0.0001, 100.0] */
779 val /= 100.0f;
780 if (val < 0.000001f || val > 1.0f)
781 return (ERANGE);
783 /* convert to an integer for use by kernel */
784 *percent = ((uint32_t)(val * ZI_PERCENTAGE_MAX));
786 return (0);
790 * This function converts a string specifier for DVAs into a bit mask.
791 * The dva's provided by the user should be 0 indexed and separated by
792 * a comma. For example:
793 * "1" -> 0b0010 (0x2)
794 * "0,1" -> 0b0011 (0x3)
795 * "0,1,2" -> 0b0111 (0x7)
797 static int
798 parse_dvas(const char *str, uint32_t *dvas_out)
800 const char *c = str;
801 uint32_t mask = 0;
802 boolean_t need_delim = B_FALSE;
804 /* max string length is 5 ("0,1,2") */
805 if (strlen(str) > 5 || strlen(str) == 0)
806 return (EINVAL);
808 while (*c != '\0') {
809 switch (*c) {
810 case '0':
811 case '1':
812 case '2':
813 /* check for pipe between DVAs */
814 if (need_delim)
815 return (EINVAL);
817 /* check if this DVA has been set already */
818 if (mask & (1 << ((*c) - '0')))
819 return (EINVAL);
821 mask |= (1 << ((*c) - '0'));
822 need_delim = B_TRUE;
823 break;
824 case ',':
825 need_delim = B_FALSE;
826 break;
827 default:
828 /* check for invalid character */
829 return (EINVAL);
831 c++;
834 /* check for dangling delimiter */
835 if (!need_delim)
836 return (EINVAL);
838 *dvas_out = mask;
839 return (0);
843 main(int argc, char **argv)
845 int c;
846 char *range = NULL;
847 char *cancel = NULL;
848 char *end;
849 char *raw = NULL;
850 char *device = NULL;
851 int level = 0;
852 int quiet = 0;
853 int error = 0;
854 int domount = 0;
855 int io_type = ZIO_TYPES;
856 int action = VDEV_STATE_UNKNOWN;
857 err_type_t type = TYPE_INVAL;
858 err_type_t label = TYPE_INVAL;
859 zinject_record_t record = { 0 };
860 char pool[MAXNAMELEN] = "";
861 char dataset[MAXNAMELEN] = "";
862 zfs_handle_t *zhp = NULL;
863 int nowrites = 0;
864 int dur_txg = 0;
865 int dur_secs = 0;
866 int ret;
867 int flags = 0;
868 uint32_t dvas = 0;
870 if ((g_zfs = libzfs_init()) == NULL) {
871 (void) fprintf(stderr, "%s\n", libzfs_error_init(errno));
872 return (1);
875 libzfs_print_on_error(g_zfs, B_TRUE);
877 if ((zfs_fd = open(ZFS_DEV, O_RDWR)) < 0) {
878 (void) fprintf(stderr, "failed to open ZFS device\n");
879 libzfs_fini(g_zfs);
880 return (1);
883 if (argc == 1) {
885 * No arguments. Print the available handlers. If there are no
886 * available handlers, direct the user to '-h' for help
887 * information.
889 if (print_all_handlers() == 0) {
890 (void) printf("No handlers registered.\n");
891 (void) printf("Run 'zinject -h' for usage "
892 "information.\n");
894 libzfs_fini(g_zfs);
895 return (0);
898 while ((c = getopt(argc, argv,
899 ":aA:b:C:d:D:f:Fg:qhIc:t:T:l:mr:s:e:uL:p:P:")) != -1) {
900 switch (c) {
901 case 'a':
902 flags |= ZINJECT_FLUSH_ARC;
903 break;
904 case 'A':
905 if (strcasecmp(optarg, "degrade") == 0) {
906 action = VDEV_STATE_DEGRADED;
907 } else if (strcasecmp(optarg, "fault") == 0) {
908 action = VDEV_STATE_FAULTED;
909 } else {
910 (void) fprintf(stderr, "invalid action '%s': "
911 "must be 'degrade' or 'fault'\n", optarg);
912 usage();
913 libzfs_fini(g_zfs);
914 return (1);
916 break;
917 case 'b':
918 raw = optarg;
919 break;
920 case 'c':
921 cancel = optarg;
922 break;
923 case 'C':
924 ret = parse_dvas(optarg, &dvas);
925 if (ret != 0) {
926 (void) fprintf(stderr, "invalid DVA list '%s': "
927 "DVAs should be 0 indexed and separated by "
928 "commas.\n", optarg);
929 usage();
930 libzfs_fini(g_zfs);
931 return (1);
933 break;
934 case 'd':
935 device = optarg;
936 break;
937 case 'D':
938 errno = 0;
939 ret = parse_delay(optarg, &record.zi_timer,
940 &record.zi_nlanes);
941 if (ret != 0) {
943 (void) fprintf(stderr, "invalid i/o delay "
944 "value: '%s'\n", optarg);
945 usage();
946 libzfs_fini(g_zfs);
947 return (1);
949 break;
950 case 'e':
951 error = str_to_err(optarg);
952 if (error < 0) {
953 (void) fprintf(stderr, "invalid error type "
954 "'%s': must be one of: io decompress "
955 "decrypt nxio dtl corrupt noop\n",
956 optarg);
957 usage();
958 libzfs_fini(g_zfs);
959 return (1);
961 break;
962 case 'f':
963 ret = parse_frequency(optarg, &record.zi_freq);
964 if (ret != 0) {
965 (void) fprintf(stderr, "%sfrequency value must "
966 "be in the range [0.0001, 100.0]\n",
967 ret == EINVAL ? "invalid value: " :
968 ret == ERANGE ? "out of range: " : "");
969 libzfs_fini(g_zfs);
970 return (1);
972 break;
973 case 'F':
974 record.zi_failfast = B_TRUE;
975 break;
976 case 'g':
977 dur_txg = 1;
978 record.zi_duration = (int)strtol(optarg, &end, 10);
979 if (record.zi_duration <= 0 || *end != '\0') {
980 (void) fprintf(stderr, "invalid duration '%s': "
981 "must be a positive integer\n", optarg);
982 usage();
983 libzfs_fini(g_zfs);
984 return (1);
986 /* store duration of txgs as its negative */
987 record.zi_duration *= -1;
988 break;
989 case 'h':
990 usage();
991 libzfs_fini(g_zfs);
992 return (0);
993 case 'I':
994 /* default duration, if one hasn't yet been defined */
995 nowrites = 1;
996 if (dur_secs == 0 && dur_txg == 0)
997 record.zi_duration = 30;
998 break;
999 case 'l':
1000 level = (int)strtol(optarg, &end, 10);
1001 if (*end != '\0') {
1002 (void) fprintf(stderr, "invalid level '%s': "
1003 "must be an integer\n", optarg);
1004 usage();
1005 libzfs_fini(g_zfs);
1006 return (1);
1008 break;
1009 case 'm':
1010 domount = 1;
1011 break;
1012 case 'p':
1013 (void) strlcpy(record.zi_func, optarg,
1014 sizeof (record.zi_func));
1015 record.zi_cmd = ZINJECT_PANIC;
1016 break;
1017 case 'P':
1018 if (strcasecmp(optarg, "import") == 0) {
1019 record.zi_cmd = ZINJECT_DELAY_IMPORT;
1020 } else if (strcasecmp(optarg, "export") == 0) {
1021 record.zi_cmd = ZINJECT_DELAY_EXPORT;
1022 } else {
1023 (void) fprintf(stderr, "invalid command '%s': "
1024 "must be 'import' or 'export'\n", optarg);
1025 usage();
1026 libzfs_fini(g_zfs);
1027 return (1);
1029 break;
1030 case 'q':
1031 quiet = 1;
1032 break;
1033 case 'r':
1034 range = optarg;
1035 flags |= ZINJECT_CALC_RANGE;
1036 break;
1037 case 's':
1038 dur_secs = 1;
1039 record.zi_duration = (int)strtol(optarg, &end, 10);
1040 if (record.zi_duration <= 0 || *end != '\0') {
1041 (void) fprintf(stderr, "invalid duration '%s': "
1042 "must be a positive integer\n", optarg);
1043 usage();
1044 libzfs_fini(g_zfs);
1045 return (1);
1047 break;
1048 case 'T':
1049 if (strcasecmp(optarg, "read") == 0) {
1050 io_type = ZIO_TYPE_READ;
1051 } else if (strcasecmp(optarg, "write") == 0) {
1052 io_type = ZIO_TYPE_WRITE;
1053 } else if (strcasecmp(optarg, "free") == 0) {
1054 io_type = ZIO_TYPE_FREE;
1055 } else if (strcasecmp(optarg, "claim") == 0) {
1056 io_type = ZIO_TYPE_CLAIM;
1057 } else if (strcasecmp(optarg, "flush") == 0) {
1058 io_type = ZIO_TYPE_FLUSH;
1059 } else if (strcasecmp(optarg, "all") == 0) {
1060 io_type = ZIO_TYPES;
1061 } else {
1062 (void) fprintf(stderr, "invalid I/O type "
1063 "'%s': must be 'read', 'write', 'free', "
1064 "'claim', 'flush' or 'all'\n", optarg);
1065 usage();
1066 libzfs_fini(g_zfs);
1067 return (1);
1069 break;
1070 case 't':
1071 if ((type = name_to_type(optarg)) == TYPE_INVAL &&
1072 !MOS_TYPE(type)) {
1073 (void) fprintf(stderr, "invalid type '%s'\n",
1074 optarg);
1075 usage();
1076 libzfs_fini(g_zfs);
1077 return (1);
1079 break;
1080 case 'u':
1081 flags |= ZINJECT_UNLOAD_SPA;
1082 break;
1083 case 'L':
1084 if ((label = name_to_type(optarg)) == TYPE_INVAL &&
1085 !LABEL_TYPE(type)) {
1086 (void) fprintf(stderr, "invalid label type "
1087 "'%s'\n", optarg);
1088 usage();
1089 libzfs_fini(g_zfs);
1090 return (1);
1092 break;
1093 case ':':
1094 (void) fprintf(stderr, "option -%c requires an "
1095 "operand\n", optopt);
1096 usage();
1097 libzfs_fini(g_zfs);
1098 return (1);
1099 case '?':
1100 (void) fprintf(stderr, "invalid option '%c'\n",
1101 optopt);
1102 usage();
1103 libzfs_fini(g_zfs);
1104 return (2);
1108 argc -= optind;
1109 argv += optind;
1111 if (record.zi_duration != 0 && record.zi_cmd == 0)
1112 record.zi_cmd = ZINJECT_IGNORED_WRITES;
1114 if (cancel != NULL) {
1116 * '-c' is invalid with any other options.
1118 if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1119 level != 0 || record.zi_cmd != ZINJECT_UNINITIALIZED ||
1120 record.zi_freq > 0 || dvas != 0) {
1121 (void) fprintf(stderr, "cancel (-c) incompatible with "
1122 "any other options\n");
1123 usage();
1124 libzfs_fini(g_zfs);
1125 return (2);
1127 if (argc != 0) {
1128 (void) fprintf(stderr, "extraneous argument to '-c'\n");
1129 usage();
1130 libzfs_fini(g_zfs);
1131 return (2);
1134 if (strcmp(cancel, "all") == 0) {
1135 return (cancel_all_handlers());
1136 } else {
1137 int id = (int)strtol(cancel, &end, 10);
1138 if (*end != '\0') {
1139 (void) fprintf(stderr, "invalid handle id '%s':"
1140 " must be an integer or 'all'\n", cancel);
1141 usage();
1142 libzfs_fini(g_zfs);
1143 return (1);
1145 return (cancel_handler(id));
1149 if (device != NULL) {
1151 * Device (-d) injection uses a completely different mechanism
1152 * for doing injection, so handle it separately here.
1154 if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1155 level != 0 || record.zi_cmd != ZINJECT_UNINITIALIZED ||
1156 dvas != 0) {
1157 (void) fprintf(stderr, "device (-d) incompatible with "
1158 "data error injection\n");
1159 usage();
1160 libzfs_fini(g_zfs);
1161 return (2);
1164 if (argc != 1) {
1165 (void) fprintf(stderr, "device (-d) injection requires "
1166 "a single pool name\n");
1167 usage();
1168 libzfs_fini(g_zfs);
1169 return (2);
1172 (void) strlcpy(pool, argv[0], sizeof (pool));
1173 dataset[0] = '\0';
1175 if (error == ECKSUM) {
1176 (void) fprintf(stderr, "device error type must be "
1177 "'io', 'nxio' or 'corrupt'\n");
1178 libzfs_fini(g_zfs);
1179 return (1);
1182 if (error == EILSEQ &&
1183 (record.zi_freq == 0 || io_type != ZIO_TYPE_READ)) {
1184 (void) fprintf(stderr, "device corrupt errors require "
1185 "io type read and a frequency value\n");
1186 libzfs_fini(g_zfs);
1187 return (1);
1190 record.zi_iotype = io_type;
1191 if (translate_device(pool, device, label, &record) != 0) {
1192 libzfs_fini(g_zfs);
1193 return (1);
1196 if (record.zi_nlanes) {
1197 switch (io_type) {
1198 case ZIO_TYPE_READ:
1199 case ZIO_TYPE_WRITE:
1200 case ZIO_TYPES:
1201 break;
1202 default:
1203 (void) fprintf(stderr, "I/O type for a delay "
1204 "must be 'read' or 'write'\n");
1205 usage();
1206 libzfs_fini(g_zfs);
1207 return (1);
1211 if (!error)
1212 error = ENXIO;
1214 if (action != VDEV_STATE_UNKNOWN)
1215 return (perform_action(pool, &record, action));
1217 } else if (raw != NULL) {
1218 if (range != NULL || type != TYPE_INVAL || level != 0 ||
1219 record.zi_cmd != ZINJECT_UNINITIALIZED ||
1220 record.zi_freq > 0 || dvas != 0) {
1221 (void) fprintf(stderr, "raw (-b) format with "
1222 "any other options\n");
1223 usage();
1224 libzfs_fini(g_zfs);
1225 return (2);
1228 if (argc != 1) {
1229 (void) fprintf(stderr, "raw (-b) format expects a "
1230 "single pool name\n");
1231 usage();
1232 libzfs_fini(g_zfs);
1233 return (2);
1236 (void) strlcpy(pool, argv[0], sizeof (pool));
1237 dataset[0] = '\0';
1239 if (error == ENXIO) {
1240 (void) fprintf(stderr, "data error type must be "
1241 "'checksum' or 'io'\n");
1242 libzfs_fini(g_zfs);
1243 return (1);
1246 record.zi_cmd = ZINJECT_DATA_FAULT;
1247 if (translate_raw(raw, &record) != 0) {
1248 libzfs_fini(g_zfs);
1249 return (1);
1251 if (!error)
1252 error = EIO;
1253 } else if (record.zi_cmd == ZINJECT_PANIC) {
1254 if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1255 level != 0 || device != NULL || record.zi_freq > 0 ||
1256 dvas != 0) {
1257 (void) fprintf(stderr, "%s incompatible with other "
1258 "options\n", "import|export delay (-P)");
1259 usage();
1260 libzfs_fini(g_zfs);
1261 return (2);
1264 if (argc < 1 || argc > 2) {
1265 (void) fprintf(stderr, "panic (-p) injection requires "
1266 "a single pool name and an optional id\n");
1267 usage();
1268 libzfs_fini(g_zfs);
1269 return (2);
1272 (void) strlcpy(pool, argv[0], sizeof (pool));
1273 if (argv[1] != NULL)
1274 record.zi_type = atoi(argv[1]);
1275 dataset[0] = '\0';
1276 } else if (record.zi_cmd == ZINJECT_DELAY_IMPORT ||
1277 record.zi_cmd == ZINJECT_DELAY_EXPORT) {
1278 if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1279 level != 0 || device != NULL || record.zi_freq > 0 ||
1280 dvas != 0) {
1281 (void) fprintf(stderr, "%s incompatible with other "
1282 "options\n", "import|export delay (-P)");
1283 usage();
1284 libzfs_fini(g_zfs);
1285 return (2);
1288 if (argc != 1 || record.zi_duration <= 0) {
1289 (void) fprintf(stderr, "import|export delay (-P) "
1290 "injection requires a duration (-s) and a single "
1291 "pool name\n");
1292 usage();
1293 libzfs_fini(g_zfs);
1294 return (2);
1297 (void) strlcpy(pool, argv[0], sizeof (pool));
1298 } else if (record.zi_cmd == ZINJECT_IGNORED_WRITES) {
1299 if (raw != NULL || range != NULL || type != TYPE_INVAL ||
1300 level != 0 || record.zi_freq > 0 || dvas != 0) {
1301 (void) fprintf(stderr, "hardware failure (-I) "
1302 "incompatible with other options\n");
1303 usage();
1304 libzfs_fini(g_zfs);
1305 return (2);
1308 if (nowrites == 0) {
1309 (void) fprintf(stderr, "-s or -g meaningless "
1310 "without -I (ignore writes)\n");
1311 usage();
1312 libzfs_fini(g_zfs);
1313 return (2);
1314 } else if (dur_secs && dur_txg) {
1315 (void) fprintf(stderr, "choose a duration either "
1316 "in seconds (-s) or a number of txgs (-g) "
1317 "but not both\n");
1318 usage();
1319 libzfs_fini(g_zfs);
1320 return (2);
1321 } else if (argc != 1) {
1322 (void) fprintf(stderr, "ignore writes (-I) "
1323 "injection requires a single pool name\n");
1324 usage();
1325 libzfs_fini(g_zfs);
1326 return (2);
1329 (void) strlcpy(pool, argv[0], sizeof (pool));
1330 dataset[0] = '\0';
1331 } else if (type == TYPE_INVAL) {
1332 if (flags == 0) {
1333 (void) fprintf(stderr, "at least one of '-b', '-d', "
1334 "'-t', '-a', '-p', '-I' or '-u' "
1335 "must be specified\n");
1336 usage();
1337 libzfs_fini(g_zfs);
1338 return (2);
1341 if (argc == 1 && (flags & ZINJECT_UNLOAD_SPA)) {
1342 (void) strlcpy(pool, argv[0], sizeof (pool));
1343 dataset[0] = '\0';
1344 } else if (argc != 0) {
1345 (void) fprintf(stderr, "extraneous argument for "
1346 "'-f'\n");
1347 usage();
1348 libzfs_fini(g_zfs);
1349 return (2);
1352 flags |= ZINJECT_NULL;
1353 } else {
1354 if (argc != 1) {
1355 (void) fprintf(stderr, "missing object\n");
1356 usage();
1357 libzfs_fini(g_zfs);
1358 return (2);
1361 if (error == ENXIO || error == EILSEQ) {
1362 (void) fprintf(stderr, "data error type must be "
1363 "'checksum' or 'io'\n");
1364 libzfs_fini(g_zfs);
1365 return (1);
1368 if (dvas != 0) {
1369 if (error == EACCES || error == EINVAL) {
1370 (void) fprintf(stderr, "the '-C' option may "
1371 "not be used with logical data errors "
1372 "'decrypt' and 'decompress'\n");
1373 libzfs_fini(g_zfs);
1374 return (1);
1377 record.zi_dvas = dvas;
1380 if (error == EACCES) {
1381 if (type != TYPE_DATA) {
1382 (void) fprintf(stderr, "decryption errors "
1383 "may only be injected for 'data' types\n");
1384 libzfs_fini(g_zfs);
1385 return (1);
1388 record.zi_cmd = ZINJECT_DECRYPT_FAULT;
1390 * Internally, ZFS actually uses ECKSUM for decryption
1391 * errors since EACCES is used to indicate the key was
1392 * not found.
1394 error = ECKSUM;
1395 } else {
1396 record.zi_cmd = ZINJECT_DATA_FAULT;
1399 if (translate_record(type, argv[0], range, level, &record, pool,
1400 dataset) != 0) {
1401 libzfs_fini(g_zfs);
1402 return (1);
1404 if (!error)
1405 error = EIO;
1409 * If this is pool-wide metadata, unmount everything. The ioctl() will
1410 * unload the pool, so that we trigger spa-wide reopen of metadata next
1411 * time we access the pool.
1413 if (dataset[0] != '\0' && domount) {
1414 if ((zhp = zfs_open(g_zfs, dataset,
1415 ZFS_TYPE_DATASET)) == NULL) {
1416 libzfs_fini(g_zfs);
1417 return (1);
1419 if (zfs_unmount(zhp, NULL, 0) != 0) {
1420 libzfs_fini(g_zfs);
1421 return (1);
1425 record.zi_error = error;
1427 ret = register_handler(pool, flags, &record, quiet);
1429 if (dataset[0] != '\0' && domount)
1430 ret = (zfs_mount(zhp, NULL, 0) != 0);
1432 libzfs_fini(g_zfs);
1434 return (ret);