xref: /illumos-gate/usr/src/uts/common/fs/zfs/zio_inject.c (revision 8a2b682e57a046b828f37bcde1776f131ef4629f)
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, 2015 by Delphix. All rights reserved.
24  */
25 
26 /*
27  * ZFS fault injection
28  *
29  * To handle fault injection, we keep track of a series of zinject_record_t
30  * structures which describe which logical block(s) should be injected with a
31  * fault.  These are kept in a global list.  Each record corresponds to a given
32  * spa_t and maintains a special hold on the spa_t so that it cannot be deleted
33  * or exported while the injection record exists.
34  *
35  * Device level injection is done using the 'zi_guid' field.  If this is set, it
36  * means that the error is destined for a particular device, not a piece of
37  * data.
38  *
39  * This is a rather poor data structure and algorithm, but we don't expect more
40  * than a few faults at any one time, so it should be sufficient for our needs.
41  */
42 
43 #include <sys/arc.h>
44 #include <sys/zio_impl.h>
45 #include <sys/zfs_ioctl.h>
46 #include <sys/vdev_impl.h>
47 #include <sys/dmu_objset.h>
48 #include <sys/fs/zfs.h>
49 
50 uint32_t zio_injection_enabled;
51 
52 /*
53  * Data describing each zinject handler registered on the system, and
54  * contains the list node linking the handler in the global zinject
55  * handler list.
56  */
57 typedef struct inject_handler {
58 	int			zi_id;
59 	spa_t			*zi_spa;
60 	zinject_record_t	zi_record;
61 	uint64_t		*zi_lanes;
62 	int			zi_next_lane;
63 	list_node_t		zi_link;
64 } inject_handler_t;
65 
66 /*
67  * List of all zinject handlers registered on the system, protected by
68  * the inject_lock defined below.
69  */
70 static list_t inject_handlers;
71 
72 /*
73  * This protects insertion into, and traversal of, the inject handler
74  * list defined above; as well as the inject_delay_count. Any time a
75  * handler is inserted or removed from the list, this lock should be
76  * taken as a RW_WRITER; and any time traversal is done over the list
77  * (without modification to it) this lock should be taken as a RW_READER.
78  */
79 static krwlock_t inject_lock;
80 
81 /*
82  * This holds the number of zinject delay handlers that have been
83  * registered on the system. It is protected by the inject_lock defined
84  * above. Thus modifications to this count must be a RW_WRITER of the
85  * inject_lock, and reads of this count must be (at least) a RW_READER
86  * of the lock.
87  */
88 static int inject_delay_count = 0;
89 
90 /*
91  * This lock is used only in zio_handle_io_delay(), refer to the comment
92  * in that function for more details.
93  */
94 static kmutex_t inject_delay_mtx;
95 
96 /*
97  * Used to assign unique identifying numbers to each new zinject handler.
98  */
99 static int inject_next_id = 1;
100 
101 /*
102  * Returns true if the given record matches the I/O in progress.
103  */
104 static boolean_t
105 zio_match_handler(zbookmark_phys_t *zb, uint64_t type, int dva,
106     zinject_record_t *record, int error)
107 {
108 	/*
109 	 * Check for a match against the MOS, which is based on type
110 	 */
111 	if (zb->zb_objset == DMU_META_OBJSET &&
112 	    record->zi_objset == DMU_META_OBJSET &&
113 	    record->zi_object == DMU_META_DNODE_OBJECT) {
114 		if (record->zi_type == DMU_OT_NONE ||
115 		    type == record->zi_type)
116 			return (record->zi_freq == 0 ||
117 			    spa_get_random(100) < record->zi_freq);
118 		else
119 			return (B_FALSE);
120 	}
121 
122 	/*
123 	 * Check for an exact match.
124 	 */
125 	if (zb->zb_objset == record->zi_objset &&
126 	    zb->zb_object == record->zi_object &&
127 	    zb->zb_level == record->zi_level &&
128 	    zb->zb_blkid >= record->zi_start &&
129 	    zb->zb_blkid <= record->zi_end &&
130 	    (record->zi_dvas == 0 || (record->zi_dvas & (1ULL << dva))) &&
131 	    error == record->zi_error) {
132 		return (record->zi_freq == 0 ||
133 		    spa_get_random(100) < record->zi_freq);
134 	}
135 
136 	return (B_FALSE);
137 }
138 
139 /*
140  * Panic the system when a config change happens in the function
141  * specified by tag.
142  */
143 void
144 zio_handle_panic_injection(spa_t *spa, char *tag, uint64_t type)
145 {
146 	inject_handler_t *handler;
147 
148 	rw_enter(&inject_lock, RW_READER);
149 
150 	for (handler = list_head(&inject_handlers); handler != NULL;
151 	    handler = list_next(&inject_handlers, handler)) {
152 
153 		if (spa != handler->zi_spa)
154 			continue;
155 
156 		if (handler->zi_record.zi_type == type &&
157 		    strcmp(tag, handler->zi_record.zi_func) == 0)
158 			panic("Panic requested in function %s\n", tag);
159 	}
160 
161 	rw_exit(&inject_lock);
162 }
163 
164 
165 /*
166  * If this is a physical I/O for a vdev child determine which DVA it is
167  * for. We iterate backwards through the DVAs matching on the offset so
168  * that we end up with ZI_NO_DVA (-1) if we don't find a match.
169  */
170 static int
171 zio_match_dva(zio_t *zio)
172 {
173 	int i = ZI_NO_DVA;
174 
175 	if (zio->io_bp != NULL && zio->io_vd != NULL &&
176 	    zio->io_child_type == ZIO_CHILD_VDEV) {
177 		for (i = BP_GET_NDVAS(zio->io_bp) - 1; i >= 0; i--) {
178 			dva_t *dva = &zio->io_bp->blk_dva[i];
179 			uint64_t off = DVA_GET_OFFSET(dva);
180 			vdev_t *vd = vdev_lookup_top(zio->io_spa,
181 			    DVA_GET_VDEV(dva));
182 
183 			/* Compensate for vdev label added to leaves */
184 			if (zio->io_vd->vdev_ops->vdev_op_leaf)
185 				off += VDEV_LABEL_START_SIZE;
186 
187 			if (zio->io_vd == vd && zio->io_offset == off)
188 				break;
189 		}
190 	}
191 
192 	return (i);
193 }
194 
195 
196 /*
197  * Inject a decryption failure. Decryption failures can occur in
198  * both the ARC and the ZIO layers.
199  */
200 int
201 zio_handle_decrypt_injection(spa_t *spa, const zbookmark_phys_t *zb,
202     uint64_t type, int error)
203 {
204 	int ret = 0;
205 	inject_handler_t *handler;
206 
207 	rw_enter(&inject_lock, RW_READER);
208 
209 	for (handler = list_head(&inject_handlers); handler != NULL;
210 	    handler = list_next(&inject_handlers, handler)) {
211 
212 		if (spa != handler->zi_spa ||
213 		    handler->zi_record.zi_cmd != ZINJECT_DECRYPT_FAULT)
214 			continue;
215 
216 		if (zio_match_handler((zbookmark_phys_t *)zb, type, ZI_NO_DVA,
217 		    &handler->zi_record, error)) {
218 			ret = error;
219 			break;
220 		}
221 	}
222 
223 	rw_exit(&inject_lock);
224 	return (ret);
225 }
226 
227 /*
228  * Determine if the I/O in question should return failure.  Returns the errno
229  * to be returned to the caller.
230  */
231 int
232 zio_handle_fault_injection(zio_t *zio, int error)
233 {
234 	int ret = 0;
235 	inject_handler_t *handler;
236 
237 	/*
238 	 * Ignore I/O not associated with any logical data.
239 	 */
240 	if (zio->io_logical == NULL)
241 		return (0);
242 
243 	/*
244 	 * Currently, we only support fault injection on reads.
245 	 */
246 	if (zio->io_type != ZIO_TYPE_READ)
247 		return (0);
248 
249 	rw_enter(&inject_lock, RW_READER);
250 
251 	for (handler = list_head(&inject_handlers); handler != NULL;
252 	    handler = list_next(&inject_handlers, handler)) {
253 
254 		if (zio->io_spa != handler->zi_spa ||
255 		    handler->zi_record.zi_cmd != ZINJECT_DATA_FAULT)
256 			continue;
257 
258 		/* If this handler matches, return the specified error */
259 		if (zio_match_handler(&zio->io_logical->io_bookmark,
260 		    zio->io_bp ? BP_GET_TYPE(zio->io_bp) : DMU_OT_NONE,
261 		    zio_match_dva(zio), &handler->zi_record, error)) {
262 			ret = error;
263 			break;
264 		}
265 	}
266 
267 	rw_exit(&inject_lock);
268 
269 	return (ret);
270 }
271 
272 /*
273  * Determine if the zio is part of a label update and has an injection
274  * handler associated with that portion of the label. Currently, we
275  * allow error injection in either the nvlist or the uberblock region of
276  * of the vdev label.
277  */
278 int
279 zio_handle_label_injection(zio_t *zio, int error)
280 {
281 	inject_handler_t *handler;
282 	vdev_t *vd = zio->io_vd;
283 	uint64_t offset = zio->io_offset;
284 	int label;
285 	int ret = 0;
286 
287 	if (offset >= VDEV_LABEL_START_SIZE &&
288 	    offset < vd->vdev_psize - VDEV_LABEL_END_SIZE)
289 		return (0);
290 
291 	rw_enter(&inject_lock, RW_READER);
292 
293 	for (handler = list_head(&inject_handlers); handler != NULL;
294 	    handler = list_next(&inject_handlers, handler)) {
295 		uint64_t start = handler->zi_record.zi_start;
296 		uint64_t end = handler->zi_record.zi_end;
297 
298 		if (handler->zi_record.zi_cmd != ZINJECT_LABEL_FAULT)
299 			continue;
300 
301 		/*
302 		 * The injection region is the relative offsets within a
303 		 * vdev label. We must determine the label which is being
304 		 * updated and adjust our region accordingly.
305 		 */
306 		label = vdev_label_number(vd->vdev_psize, offset);
307 		start = vdev_label_offset(vd->vdev_psize, label, start);
308 		end = vdev_label_offset(vd->vdev_psize, label, end);
309 
310 		if (zio->io_vd->vdev_guid == handler->zi_record.zi_guid &&
311 		    (offset >= start && offset <= end)) {
312 			ret = error;
313 			break;
314 		}
315 	}
316 	rw_exit(&inject_lock);
317 	return (ret);
318 }
319 
320 
321 int
322 zio_handle_device_injection(vdev_t *vd, zio_t *zio, int error)
323 {
324 	inject_handler_t *handler;
325 	int ret = 0;
326 
327 	/*
328 	 * We skip over faults in the labels unless it's during
329 	 * device open (i.e. zio == NULL).
330 	 */
331 	if (zio != NULL) {
332 		uint64_t offset = zio->io_offset;
333 
334 		if (offset < VDEV_LABEL_START_SIZE ||
335 		    offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE)
336 			return (0);
337 	}
338 
339 	rw_enter(&inject_lock, RW_READER);
340 
341 	for (handler = list_head(&inject_handlers); handler != NULL;
342 	    handler = list_next(&inject_handlers, handler)) {
343 
344 		if (handler->zi_record.zi_cmd != ZINJECT_DEVICE_FAULT)
345 			continue;
346 
347 		if (vd->vdev_guid == handler->zi_record.zi_guid) {
348 			if (handler->zi_record.zi_failfast &&
349 			    (zio == NULL || (zio->io_flags &
350 			    (ZIO_FLAG_IO_RETRY | ZIO_FLAG_TRYHARD)))) {
351 				continue;
352 			}
353 
354 			/* Handle type specific I/O failures */
355 			if (zio != NULL &&
356 			    handler->zi_record.zi_iotype != ZIO_TYPES &&
357 			    handler->zi_record.zi_iotype != zio->io_type)
358 				continue;
359 
360 			if (handler->zi_record.zi_error == error) {
361 				/*
362 				 * For a failed open, pretend like the device
363 				 * has gone away.
364 				 */
365 				if (error == ENXIO)
366 					vd->vdev_stat.vs_aux =
367 					    VDEV_AUX_OPEN_FAILED;
368 
369 				/*
370 				 * Treat these errors as if they had been
371 				 * retried so that all the appropriate stats
372 				 * and FMA events are generated.
373 				 */
374 				if (!handler->zi_record.zi_failfast &&
375 				    zio != NULL)
376 					zio->io_flags |= ZIO_FLAG_IO_RETRY;
377 
378 				ret = error;
379 				break;
380 			}
381 			if (handler->zi_record.zi_error == ENXIO) {
382 				ret = SET_ERROR(EIO);
383 				break;
384 			}
385 		}
386 	}
387 
388 	rw_exit(&inject_lock);
389 
390 	return (ret);
391 }
392 
393 /*
394  * Simulate hardware that ignores cache flushes.  For requested number
395  * of seconds nix the actual writing to disk.
396  */
397 void
398 zio_handle_ignored_writes(zio_t *zio)
399 {
400 	inject_handler_t *handler;
401 
402 	rw_enter(&inject_lock, RW_READER);
403 
404 	for (handler = list_head(&inject_handlers); handler != NULL;
405 	    handler = list_next(&inject_handlers, handler)) {
406 
407 		/* Ignore errors not destined for this pool */
408 		if (zio->io_spa != handler->zi_spa ||
409 		    handler->zi_record.zi_cmd != ZINJECT_IGNORED_WRITES)
410 			continue;
411 
412 		/*
413 		 * Positive duration implies # of seconds, negative
414 		 * a number of txgs
415 		 */
416 		if (handler->zi_record.zi_timer == 0) {
417 			if (handler->zi_record.zi_duration > 0)
418 				handler->zi_record.zi_timer = ddi_get_lbolt64();
419 			else
420 				handler->zi_record.zi_timer = zio->io_txg;
421 		}
422 
423 		/* Have a "problem" writing 60% of the time */
424 		if (spa_get_random(100) < 60)
425 			zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
426 		break;
427 	}
428 
429 	rw_exit(&inject_lock);
430 }
431 
432 void
433 spa_handle_ignored_writes(spa_t *spa)
434 {
435 	inject_handler_t *handler;
436 
437 	if (zio_injection_enabled == 0)
438 		return;
439 
440 	rw_enter(&inject_lock, RW_READER);
441 
442 	for (handler = list_head(&inject_handlers); handler != NULL;
443 	    handler = list_next(&inject_handlers, handler)) {
444 
445 		if (spa != handler->zi_spa ||
446 		    handler->zi_record.zi_cmd != ZINJECT_IGNORED_WRITES)
447 			continue;
448 
449 		if (handler->zi_record.zi_duration > 0) {
450 			VERIFY(handler->zi_record.zi_timer == 0 ||
451 			    handler->zi_record.zi_timer +
452 			    handler->zi_record.zi_duration * hz >
453 			    ddi_get_lbolt64());
454 		} else {
455 			/* duration is negative so the subtraction here adds */
456 			VERIFY(handler->zi_record.zi_timer == 0 ||
457 			    handler->zi_record.zi_timer -
458 			    handler->zi_record.zi_duration >=
459 			    spa_syncing_txg(spa));
460 		}
461 	}
462 
463 	rw_exit(&inject_lock);
464 }
465 
466 hrtime_t
467 zio_handle_io_delay(zio_t *zio)
468 {
469 	vdev_t *vd = zio->io_vd;
470 	inject_handler_t *min_handler = NULL;
471 	hrtime_t min_target = 0;
472 
473 	rw_enter(&inject_lock, RW_READER);
474 
475 	/*
476 	 * inject_delay_count is a subset of zio_injection_enabled that
477 	 * is only incremented for delay handlers. These checks are
478 	 * mainly added to remind the reader why we're not explicitly
479 	 * checking zio_injection_enabled like the other functions.
480 	 */
481 	IMPLY(inject_delay_count > 0, zio_injection_enabled > 0);
482 	IMPLY(zio_injection_enabled == 0, inject_delay_count == 0);
483 
484 	/*
485 	 * If there aren't any inject delay handlers registered, then we
486 	 * can short circuit and simply return 0 here. A value of zero
487 	 * informs zio_delay_interrupt() that this request should not be
488 	 * delayed. This short circuit keeps us from acquiring the
489 	 * inject_delay_mutex unnecessarily.
490 	 */
491 	if (inject_delay_count == 0) {
492 		rw_exit(&inject_lock);
493 		return (0);
494 	}
495 
496 	/*
497 	 * Each inject handler has a number of "lanes" associated with
498 	 * it. Each lane is able to handle requests independently of one
499 	 * another, and at a latency defined by the inject handler
500 	 * record's zi_timer field. Thus if a handler in configured with
501 	 * a single lane with a 10ms latency, it will delay requests
502 	 * such that only a single request is completed every 10ms. So,
503 	 * if more than one request is attempted per each 10ms interval,
504 	 * the average latency of the requests will be greater than
505 	 * 10ms; but if only a single request is submitted each 10ms
506 	 * interval the average latency will be 10ms.
507 	 *
508 	 * We need to acquire this mutex to prevent multiple concurrent
509 	 * threads being assigned to the same lane of a given inject
510 	 * handler. The mutex allows us to perform the following two
511 	 * operations atomically:
512 	 *
513 	 *	1. determine the minimum handler and minimum target
514 	 *	   value of all the possible handlers
515 	 *	2. update that minimum handler's lane array
516 	 *
517 	 * Without atomicity, two (or more) threads could pick the same
518 	 * lane in step (1), and then conflict with each other in step
519 	 * (2). This could allow a single lane handler to process
520 	 * multiple requests simultaneously, which shouldn't be possible.
521 	 */
522 	mutex_enter(&inject_delay_mtx);
523 
524 	for (inject_handler_t *handler = list_head(&inject_handlers);
525 	    handler != NULL; handler = list_next(&inject_handlers, handler)) {
526 		if (handler->zi_record.zi_cmd != ZINJECT_DELAY_IO)
527 			continue;
528 
529 		if (vd->vdev_guid != handler->zi_record.zi_guid)
530 			continue;
531 
532 		/*
533 		 * Defensive; should never happen as the array allocation
534 		 * occurs prior to inserting this handler on the list.
535 		 */
536 		ASSERT3P(handler->zi_lanes, !=, NULL);
537 
538 		/*
539 		 * This should never happen, the zinject command should
540 		 * prevent a user from setting an IO delay with zero lanes.
541 		 */
542 		ASSERT3U(handler->zi_record.zi_nlanes, !=, 0);
543 
544 		ASSERT3U(handler->zi_record.zi_nlanes, >,
545 		    handler->zi_next_lane);
546 
547 		/*
548 		 * We want to issue this IO to the lane that will become
549 		 * idle the soonest, so we compare the soonest this
550 		 * specific handler can complete the IO with all other
551 		 * handlers, to find the lowest value of all possible
552 		 * lanes. We then use this lane to submit the request.
553 		 *
554 		 * Since each handler has a constant value for its
555 		 * delay, we can just use the "next" lane for that
556 		 * handler; as it will always be the lane with the
557 		 * lowest value for that particular handler (i.e. the
558 		 * lane that will become idle the soonest). This saves a
559 		 * scan of each handler's lanes array.
560 		 *
561 		 * There's two cases to consider when determining when
562 		 * this specific IO request should complete. If this
563 		 * lane is idle, we want to "submit" the request now so
564 		 * it will complete after zi_timer milliseconds. Thus,
565 		 * we set the target to now + zi_timer.
566 		 *
567 		 * If the lane is busy, we want this request to complete
568 		 * zi_timer milliseconds after the lane becomes idle.
569 		 * Since the 'zi_lanes' array holds the time at which
570 		 * each lane will become idle, we use that value to
571 		 * determine when this request should complete.
572 		 */
573 		hrtime_t idle = handler->zi_record.zi_timer + gethrtime();
574 		hrtime_t busy = handler->zi_record.zi_timer +
575 		    handler->zi_lanes[handler->zi_next_lane];
576 		hrtime_t target = MAX(idle, busy);
577 
578 		if (min_handler == NULL) {
579 			min_handler = handler;
580 			min_target = target;
581 			continue;
582 		}
583 
584 		ASSERT3P(min_handler, !=, NULL);
585 		ASSERT3U(min_target, !=, 0);
586 
587 		/*
588 		 * We don't yet increment the "next lane" variable since
589 		 * we still might find a lower value lane in another
590 		 * handler during any remaining iterations. Once we're
591 		 * sure we've selected the absolute minimum, we'll claim
592 		 * the lane and increment the handler's "next lane"
593 		 * field below.
594 		 */
595 
596 		if (target < min_target) {
597 			min_handler = handler;
598 			min_target = target;
599 		}
600 	}
601 
602 	/*
603 	 * 'min_handler' will be NULL if no IO delays are registered for
604 	 * this vdev, otherwise it will point to the handler containing
605 	 * the lane that will become idle the soonest.
606 	 */
607 	if (min_handler != NULL) {
608 		ASSERT3U(min_target, !=, 0);
609 		min_handler->zi_lanes[min_handler->zi_next_lane] = min_target;
610 
611 		/*
612 		 * If we've used all possible lanes for this handler,
613 		 * loop back and start using the first lane again;
614 		 * otherwise, just increment the lane index.
615 		 */
616 		min_handler->zi_next_lane = (min_handler->zi_next_lane + 1) %
617 		    min_handler->zi_record.zi_nlanes;
618 	}
619 
620 	mutex_exit(&inject_delay_mtx);
621 	rw_exit(&inject_lock);
622 
623 	return (min_target);
624 }
625 
626 /*
627  * Create a new handler for the given record.  We add it to the list, adding
628  * a reference to the spa_t in the process.  We increment zio_injection_enabled,
629  * which is the switch to trigger all fault injection.
630  */
631 int
632 zio_inject_fault(char *name, int flags, int *id, zinject_record_t *record)
633 {
634 	inject_handler_t *handler;
635 	int error;
636 	spa_t *spa;
637 
638 	/*
639 	 * If this is pool-wide metadata, make sure we unload the corresponding
640 	 * spa_t, so that the next attempt to load it will trigger the fault.
641 	 * We call spa_reset() to unload the pool appropriately.
642 	 */
643 	if (flags & ZINJECT_UNLOAD_SPA)
644 		if ((error = spa_reset(name)) != 0)
645 			return (error);
646 
647 	if (record->zi_cmd == ZINJECT_DELAY_IO) {
648 		/*
649 		 * A value of zero for the number of lanes or for the
650 		 * delay time doesn't make sense.
651 		 */
652 		if (record->zi_timer == 0 || record->zi_nlanes == 0)
653 			return (SET_ERROR(EINVAL));
654 
655 		/*
656 		 * The number of lanes is directly mapped to the size of
657 		 * an array used by the handler. Thus, to ensure the
658 		 * user doesn't trigger an allocation that's "too large"
659 		 * we cap the number of lanes here.
660 		 */
661 		if (record->zi_nlanes >= UINT16_MAX)
662 			return (SET_ERROR(EINVAL));
663 	}
664 
665 	if (!(flags & ZINJECT_NULL)) {
666 		/*
667 		 * spa_inject_ref() will add an injection reference, which will
668 		 * prevent the pool from being removed from the namespace while
669 		 * still allowing it to be unloaded.
670 		 */
671 		if ((spa = spa_inject_addref(name)) == NULL)
672 			return (SET_ERROR(ENOENT));
673 
674 		handler = kmem_alloc(sizeof (inject_handler_t), KM_SLEEP);
675 
676 		handler->zi_spa = spa;
677 		handler->zi_record = *record;
678 
679 		if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) {
680 			handler->zi_lanes = kmem_zalloc(
681 			    sizeof (*handler->zi_lanes) *
682 			    handler->zi_record.zi_nlanes, KM_SLEEP);
683 			handler->zi_next_lane = 0;
684 		} else {
685 			handler->zi_lanes = NULL;
686 			handler->zi_next_lane = 0;
687 		}
688 
689 		rw_enter(&inject_lock, RW_WRITER);
690 
691 		/*
692 		 * We can't move this increment into the conditional
693 		 * above because we need to hold the RW_WRITER lock of
694 		 * inject_lock, and we don't want to hold that while
695 		 * allocating the handler's zi_lanes array.
696 		 */
697 		if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) {
698 			ASSERT3S(inject_delay_count, >=, 0);
699 			inject_delay_count++;
700 			ASSERT3S(inject_delay_count, >, 0);
701 		}
702 
703 		*id = handler->zi_id = inject_next_id++;
704 		list_insert_tail(&inject_handlers, handler);
705 		atomic_inc_32(&zio_injection_enabled);
706 
707 		rw_exit(&inject_lock);
708 	}
709 
710 	/*
711 	 * Flush the ARC, so that any attempts to read this data will end up
712 	 * going to the ZIO layer.  Note that this is a little overkill, but
713 	 * we don't have the necessary ARC interfaces to do anything else, and
714 	 * fault injection isn't a performance critical path.
715 	 */
716 	if (flags & ZINJECT_FLUSH_ARC)
717 		/*
718 		 * We must use FALSE to ensure arc_flush returns, since
719 		 * we're not preventing concurrent ARC insertions.
720 		 */
721 		arc_flush(NULL, FALSE);
722 
723 	return (0);
724 }
725 
726 /*
727  * Returns the next record with an ID greater than that supplied to the
728  * function.  Used to iterate over all handlers in the system.
729  */
730 int
731 zio_inject_list_next(int *id, char *name, size_t buflen,
732     zinject_record_t *record)
733 {
734 	inject_handler_t *handler;
735 	int ret;
736 
737 	mutex_enter(&spa_namespace_lock);
738 	rw_enter(&inject_lock, RW_READER);
739 
740 	for (handler = list_head(&inject_handlers); handler != NULL;
741 	    handler = list_next(&inject_handlers, handler))
742 		if (handler->zi_id > *id)
743 			break;
744 
745 	if (handler) {
746 		*record = handler->zi_record;
747 		*id = handler->zi_id;
748 		(void) strncpy(name, spa_name(handler->zi_spa), buflen);
749 		ret = 0;
750 	} else {
751 		ret = SET_ERROR(ENOENT);
752 	}
753 
754 	rw_exit(&inject_lock);
755 	mutex_exit(&spa_namespace_lock);
756 
757 	return (ret);
758 }
759 
760 /*
761  * Clear the fault handler with the given identifier, or return ENOENT if none
762  * exists.
763  */
764 int
765 zio_clear_fault(int id)
766 {
767 	inject_handler_t *handler;
768 
769 	rw_enter(&inject_lock, RW_WRITER);
770 
771 	for (handler = list_head(&inject_handlers); handler != NULL;
772 	    handler = list_next(&inject_handlers, handler))
773 		if (handler->zi_id == id)
774 			break;
775 
776 	if (handler == NULL) {
777 		rw_exit(&inject_lock);
778 		return (SET_ERROR(ENOENT));
779 	}
780 
781 	if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) {
782 		ASSERT3S(inject_delay_count, >, 0);
783 		inject_delay_count--;
784 		ASSERT3S(inject_delay_count, >=, 0);
785 	}
786 
787 	list_remove(&inject_handlers, handler);
788 	rw_exit(&inject_lock);
789 
790 	if (handler->zi_record.zi_cmd == ZINJECT_DELAY_IO) {
791 		ASSERT3P(handler->zi_lanes, !=, NULL);
792 		kmem_free(handler->zi_lanes, sizeof (*handler->zi_lanes) *
793 		    handler->zi_record.zi_nlanes);
794 	} else {
795 		ASSERT3P(handler->zi_lanes, ==, NULL);
796 	}
797 
798 	spa_inject_delref(handler->zi_spa);
799 	kmem_free(handler, sizeof (inject_handler_t));
800 	atomic_dec_32(&zio_injection_enabled);
801 
802 	return (0);
803 }
804 
805 void
806 zio_inject_init(void)
807 {
808 	rw_init(&inject_lock, NULL, RW_DEFAULT, NULL);
809 	mutex_init(&inject_delay_mtx, NULL, MUTEX_DEFAULT, NULL);
810 	list_create(&inject_handlers, sizeof (inject_handler_t),
811 	    offsetof(inject_handler_t, zi_link));
812 }
813 
814 void
815 zio_inject_fini(void)
816 {
817 	list_destroy(&inject_handlers);
818 	mutex_destroy(&inject_delay_mtx);
819 	rw_destroy(&inject_lock);
820 }
821