xref: /freebsd/sys/contrib/openzfs/module/zfs/zfs_fm.c (revision 6132212808e8dccedc9e5d85fea4390c2f38059a)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * Copyright (c) 2012 by Delphix. All rights reserved.
28  */
29 
30 #include <sys/spa.h>
31 #include <sys/spa_impl.h>
32 #include <sys/vdev.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio.h>
35 #include <sys/zio_checksum.h>
36 
37 #include <sys/fm/fs/zfs.h>
38 #include <sys/fm/protocol.h>
39 #include <sys/fm/util.h>
40 #include <sys/sysevent.h>
41 
42 /*
43  * This general routine is responsible for generating all the different ZFS
44  * ereports.  The payload is dependent on the class, and which arguments are
45  * supplied to the function:
46  *
47  * 	EREPORT			POOL	VDEV	IO
48  * 	block			X	X	X
49  * 	data			X		X
50  * 	device			X	X
51  * 	pool			X
52  *
53  * If we are in a loading state, all errors are chained together by the same
54  * SPA-wide ENA (Error Numeric Association).
55  *
56  * For isolated I/O requests, we get the ENA from the zio_t. The propagation
57  * gets very complicated due to RAID-Z, gang blocks, and vdev caching.  We want
58  * to chain together all ereports associated with a logical piece of data.  For
59  * read I/Os, there  are basically three 'types' of I/O, which form a roughly
60  * layered diagram:
61  *
62  *      +---------------+
63  * 	| Aggregate I/O |	No associated logical data or device
64  * 	+---------------+
65  *              |
66  *              V
67  * 	+---------------+	Reads associated with a piece of logical data.
68  * 	|   Read I/O    |	This includes reads on behalf of RAID-Z,
69  * 	+---------------+       mirrors, gang blocks, retries, etc.
70  *              |
71  *              V
72  * 	+---------------+	Reads associated with a particular device, but
73  * 	| Physical I/O  |	no logical data.  Issued as part of vdev caching
74  * 	+---------------+	and I/O aggregation.
75  *
76  * Note that 'physical I/O' here is not the same terminology as used in the rest
77  * of ZIO.  Typically, 'physical I/O' simply means that there is no attached
78  * blockpointer.  But I/O with no associated block pointer can still be related
79  * to a logical piece of data (i.e. RAID-Z requests).
80  *
81  * Purely physical I/O always have unique ENAs.  They are not related to a
82  * particular piece of logical data, and therefore cannot be chained together.
83  * We still generate an ereport, but the DE doesn't correlate it with any
84  * logical piece of data.  When such an I/O fails, the delegated I/O requests
85  * will issue a retry, which will trigger the 'real' ereport with the correct
86  * ENA.
87  *
88  * We keep track of the ENA for a ZIO chain through the 'io_logical' member.
89  * When a new logical I/O is issued, we set this to point to itself.  Child I/Os
90  * then inherit this pointer, so that when it is first set subsequent failures
91  * will use the same ENA.  For vdev cache fill and queue aggregation I/O,
92  * this pointer is set to NULL, and no ereport will be generated (since it
93  * doesn't actually correspond to any particular device or piece of data,
94  * and the caller will always retry without caching or queueing anyway).
95  *
96  * For checksum errors, we want to include more information about the actual
97  * error which occurs.  Accordingly, we build an ereport when the error is
98  * noticed, but instead of sending it in immediately, we hang it off of the
99  * io_cksum_report field of the logical IO.  When the logical IO completes
100  * (successfully or not), zfs_ereport_finish_checksum() is called with the
101  * good and bad versions of the buffer (if available), and we annotate the
102  * ereport with information about the differences.
103  */
104 #ifdef _KERNEL
105 void
106 zfs_zevent_post_cb(nvlist_t *nvl, nvlist_t *detector)
107 {
108 	if (nvl)
109 		fm_nvlist_destroy(nvl, FM_NVA_FREE);
110 
111 	if (detector)
112 		fm_nvlist_destroy(detector, FM_NVA_FREE);
113 }
114 
115 /*
116  * We want to rate limit ZIO delay and checksum events so as to not
117  * flood ZED when a disk is acting up.
118  *
119  * Returns 1 if we're ratelimiting, 0 if not.
120  */
121 static int
122 zfs_is_ratelimiting_event(const char *subclass, vdev_t *vd)
123 {
124 	int rc = 0;
125 	/*
126 	 * __ratelimit() returns 1 if we're *not* ratelimiting and 0 if we
127 	 * are.  Invert it to get our return value.
128 	 */
129 	if (strcmp(subclass, FM_EREPORT_ZFS_DELAY) == 0) {
130 		rc = !zfs_ratelimit(&vd->vdev_delay_rl);
131 	} else if (strcmp(subclass, FM_EREPORT_ZFS_CHECKSUM) == 0) {
132 		rc = !zfs_ratelimit(&vd->vdev_checksum_rl);
133 	}
134 
135 	if (rc)	{
136 		/* We're rate limiting */
137 		fm_erpt_dropped_increment();
138 	}
139 
140 	return (rc);
141 }
142 
143 /*
144  * Return B_TRUE if the event actually posted, B_FALSE if not.
145  */
146 static boolean_t
147 zfs_ereport_start(nvlist_t **ereport_out, nvlist_t **detector_out,
148     const char *subclass, spa_t *spa, vdev_t *vd, const zbookmark_phys_t *zb,
149     zio_t *zio, uint64_t stateoroffset, uint64_t size)
150 {
151 	nvlist_t *ereport, *detector;
152 
153 	uint64_t ena;
154 	char class[64];
155 
156 	if (!zfs_ereport_is_valid(subclass, spa, vd, zio))
157 		return (B_FALSE);
158 
159 	if ((ereport = fm_nvlist_create(NULL)) == NULL)
160 		return (B_FALSE);
161 
162 	if ((detector = fm_nvlist_create(NULL)) == NULL) {
163 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
164 		return (B_FALSE);
165 	}
166 
167 	/*
168 	 * Serialize ereport generation
169 	 */
170 	mutex_enter(&spa->spa_errlist_lock);
171 
172 	/*
173 	 * Determine the ENA to use for this event.  If we are in a loading
174 	 * state, use a SPA-wide ENA.  Otherwise, if we are in an I/O state, use
175 	 * a root zio-wide ENA.  Otherwise, simply use a unique ENA.
176 	 */
177 	if (spa_load_state(spa) != SPA_LOAD_NONE) {
178 		if (spa->spa_ena == 0)
179 			spa->spa_ena = fm_ena_generate(0, FM_ENA_FMT1);
180 		ena = spa->spa_ena;
181 	} else if (zio != NULL && zio->io_logical != NULL) {
182 		if (zio->io_logical->io_ena == 0)
183 			zio->io_logical->io_ena =
184 			    fm_ena_generate(0, FM_ENA_FMT1);
185 		ena = zio->io_logical->io_ena;
186 	} else {
187 		ena = fm_ena_generate(0, FM_ENA_FMT1);
188 	}
189 
190 	/*
191 	 * Construct the full class, detector, and other standard FMA fields.
192 	 */
193 	(void) snprintf(class, sizeof (class), "%s.%s",
194 	    ZFS_ERROR_CLASS, subclass);
195 
196 	fm_fmri_zfs_set(detector, FM_ZFS_SCHEME_VERSION, spa_guid(spa),
197 	    vd != NULL ? vd->vdev_guid : 0);
198 
199 	fm_ereport_set(ereport, FM_EREPORT_VERSION, class, ena, detector, NULL);
200 
201 	/*
202 	 * Construct the per-ereport payload, depending on which parameters are
203 	 * passed in.
204 	 */
205 
206 	/*
207 	 * Generic payload members common to all ereports.
208 	 */
209 	fm_payload_set(ereport,
210 	    FM_EREPORT_PAYLOAD_ZFS_POOL, DATA_TYPE_STRING, spa_name(spa),
211 	    FM_EREPORT_PAYLOAD_ZFS_POOL_GUID, DATA_TYPE_UINT64, spa_guid(spa),
212 	    FM_EREPORT_PAYLOAD_ZFS_POOL_STATE, DATA_TYPE_UINT64,
213 	    (uint64_t)spa_state(spa),
214 	    FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, DATA_TYPE_INT32,
215 	    (int32_t)spa_load_state(spa), NULL);
216 
217 	fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL_FAILMODE,
218 	    DATA_TYPE_STRING,
219 	    spa_get_failmode(spa) == ZIO_FAILURE_MODE_WAIT ?
220 	    FM_EREPORT_FAILMODE_WAIT :
221 	    spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE ?
222 	    FM_EREPORT_FAILMODE_CONTINUE : FM_EREPORT_FAILMODE_PANIC,
223 	    NULL);
224 
225 	if (vd != NULL) {
226 		vdev_t *pvd = vd->vdev_parent;
227 		vdev_queue_t *vq = &vd->vdev_queue;
228 		vdev_stat_t *vs = &vd->vdev_stat;
229 		vdev_t *spare_vd;
230 		uint64_t *spare_guids;
231 		char **spare_paths;
232 		int i, spare_count;
233 
234 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID,
235 		    DATA_TYPE_UINT64, vd->vdev_guid,
236 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_TYPE,
237 		    DATA_TYPE_STRING, vd->vdev_ops->vdev_op_type, NULL);
238 		if (vd->vdev_path != NULL)
239 			fm_payload_set(ereport,
240 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_PATH,
241 			    DATA_TYPE_STRING, vd->vdev_path, NULL);
242 		if (vd->vdev_devid != NULL)
243 			fm_payload_set(ereport,
244 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DEVID,
245 			    DATA_TYPE_STRING, vd->vdev_devid, NULL);
246 		if (vd->vdev_fru != NULL)
247 			fm_payload_set(ereport,
248 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_FRU,
249 			    DATA_TYPE_STRING, vd->vdev_fru, NULL);
250 		if (vd->vdev_enc_sysfs_path != NULL)
251 			fm_payload_set(ereport,
252 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_ENC_SYSFS_PATH,
253 			    DATA_TYPE_STRING, vd->vdev_enc_sysfs_path, NULL);
254 		if (vd->vdev_ashift)
255 			fm_payload_set(ereport,
256 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_ASHIFT,
257 			    DATA_TYPE_UINT64, vd->vdev_ashift, NULL);
258 
259 		if (vq != NULL) {
260 			fm_payload_set(ereport,
261 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_COMP_TS,
262 			    DATA_TYPE_UINT64, vq->vq_io_complete_ts, NULL);
263 			fm_payload_set(ereport,
264 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DELTA_TS,
265 			    DATA_TYPE_UINT64, vq->vq_io_delta_ts, NULL);
266 		}
267 
268 		if (vs != NULL) {
269 			fm_payload_set(ereport,
270 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_READ_ERRORS,
271 			    DATA_TYPE_UINT64, vs->vs_read_errors,
272 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_WRITE_ERRORS,
273 			    DATA_TYPE_UINT64, vs->vs_write_errors,
274 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_CKSUM_ERRORS,
275 			    DATA_TYPE_UINT64, vs->vs_checksum_errors,
276 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DELAYS,
277 			    DATA_TYPE_UINT64, vs->vs_slow_ios,
278 			    NULL);
279 		}
280 
281 		if (pvd != NULL) {
282 			fm_payload_set(ereport,
283 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_GUID,
284 			    DATA_TYPE_UINT64, pvd->vdev_guid,
285 			    FM_EREPORT_PAYLOAD_ZFS_PARENT_TYPE,
286 			    DATA_TYPE_STRING, pvd->vdev_ops->vdev_op_type,
287 			    NULL);
288 			if (pvd->vdev_path)
289 				fm_payload_set(ereport,
290 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_PATH,
291 				    DATA_TYPE_STRING, pvd->vdev_path, NULL);
292 			if (pvd->vdev_devid)
293 				fm_payload_set(ereport,
294 				    FM_EREPORT_PAYLOAD_ZFS_PARENT_DEVID,
295 				    DATA_TYPE_STRING, pvd->vdev_devid, NULL);
296 		}
297 
298 		spare_count = spa->spa_spares.sav_count;
299 		spare_paths = kmem_zalloc(sizeof (char *) * spare_count,
300 		    KM_SLEEP);
301 		spare_guids = kmem_zalloc(sizeof (uint64_t) * spare_count,
302 		    KM_SLEEP);
303 
304 		for (i = 0; i < spare_count; i++) {
305 			spare_vd = spa->spa_spares.sav_vdevs[i];
306 			if (spare_vd) {
307 				spare_paths[i] = spare_vd->vdev_path;
308 				spare_guids[i] = spare_vd->vdev_guid;
309 			}
310 		}
311 
312 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_VDEV_SPARE_PATHS,
313 		    DATA_TYPE_STRING_ARRAY, spare_count, spare_paths,
314 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_SPARE_GUIDS,
315 		    DATA_TYPE_UINT64_ARRAY, spare_count, spare_guids, NULL);
316 
317 		kmem_free(spare_guids, sizeof (uint64_t) * spare_count);
318 		kmem_free(spare_paths, sizeof (char *) * spare_count);
319 	}
320 
321 	if (zio != NULL) {
322 		/*
323 		 * Payload common to all I/Os.
324 		 */
325 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_ERR,
326 		    DATA_TYPE_INT32, zio->io_error, NULL);
327 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_FLAGS,
328 		    DATA_TYPE_INT32, zio->io_flags, NULL);
329 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_STAGE,
330 		    DATA_TYPE_UINT32, zio->io_stage, NULL);
331 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_PIPELINE,
332 		    DATA_TYPE_UINT32, zio->io_pipeline, NULL);
333 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_DELAY,
334 		    DATA_TYPE_UINT64, zio->io_delay, NULL);
335 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_TIMESTAMP,
336 		    DATA_TYPE_UINT64, zio->io_timestamp, NULL);
337 		fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_ZIO_DELTA,
338 		    DATA_TYPE_UINT64, zio->io_delta, NULL);
339 
340 		/*
341 		 * If the 'size' parameter is non-zero, it indicates this is a
342 		 * RAID-Z or other I/O where the physical offset and length are
343 		 * provided for us, instead of within the zio_t.
344 		 */
345 		if (vd != NULL) {
346 			if (size)
347 				fm_payload_set(ereport,
348 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
349 				    DATA_TYPE_UINT64, stateoroffset,
350 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
351 				    DATA_TYPE_UINT64, size, NULL);
352 			else
353 				fm_payload_set(ereport,
354 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_OFFSET,
355 				    DATA_TYPE_UINT64, zio->io_offset,
356 				    FM_EREPORT_PAYLOAD_ZFS_ZIO_SIZE,
357 				    DATA_TYPE_UINT64, zio->io_size, NULL);
358 		}
359 	} else if (vd != NULL) {
360 		/*
361 		 * If we have a vdev but no zio, this is a device fault, and the
362 		 * 'stateoroffset' parameter indicates the previous state of the
363 		 * vdev.
364 		 */
365 		fm_payload_set(ereport,
366 		    FM_EREPORT_PAYLOAD_ZFS_PREV_STATE,
367 		    DATA_TYPE_UINT64, stateoroffset, NULL);
368 	}
369 
370 	/*
371 	 * Payload for I/Os with corresponding logical information.
372 	 */
373 	if (zb != NULL && (zio == NULL || zio->io_logical != NULL)) {
374 		fm_payload_set(ereport,
375 		    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJSET,
376 		    DATA_TYPE_UINT64, zb->zb_objset,
377 		    FM_EREPORT_PAYLOAD_ZFS_ZIO_OBJECT,
378 		    DATA_TYPE_UINT64, zb->zb_object,
379 		    FM_EREPORT_PAYLOAD_ZFS_ZIO_LEVEL,
380 		    DATA_TYPE_INT64, zb->zb_level,
381 		    FM_EREPORT_PAYLOAD_ZFS_ZIO_BLKID,
382 		    DATA_TYPE_UINT64, zb->zb_blkid, NULL);
383 	}
384 
385 	mutex_exit(&spa->spa_errlist_lock);
386 
387 	*ereport_out = ereport;
388 	*detector_out = detector;
389 	return (B_TRUE);
390 }
391 
392 /* if it's <= 128 bytes, save the corruption directly */
393 #define	ZFM_MAX_INLINE		(128 / sizeof (uint64_t))
394 
395 #define	MAX_RANGES		16
396 
397 typedef struct zfs_ecksum_info {
398 	/* histograms of set and cleared bits by bit number in a 64-bit word */
399 	uint32_t zei_histogram_set[sizeof (uint64_t) * NBBY];
400 	uint32_t zei_histogram_cleared[sizeof (uint64_t) * NBBY];
401 
402 	/* inline arrays of bits set and cleared. */
403 	uint64_t zei_bits_set[ZFM_MAX_INLINE];
404 	uint64_t zei_bits_cleared[ZFM_MAX_INLINE];
405 
406 	/*
407 	 * for each range, the number of bits set and cleared.  The Hamming
408 	 * distance between the good and bad buffers is the sum of them all.
409 	 */
410 	uint32_t zei_range_sets[MAX_RANGES];
411 	uint32_t zei_range_clears[MAX_RANGES];
412 
413 	struct zei_ranges {
414 		uint32_t	zr_start;
415 		uint32_t	zr_end;
416 	} zei_ranges[MAX_RANGES];
417 
418 	size_t	zei_range_count;
419 	uint32_t zei_mingap;
420 	uint32_t zei_allowed_mingap;
421 
422 } zfs_ecksum_info_t;
423 
424 static void
425 update_histogram(uint64_t value_arg, uint32_t *hist, uint32_t *count)
426 {
427 	size_t i;
428 	size_t bits = 0;
429 	uint64_t value = BE_64(value_arg);
430 
431 	/* We store the bits in big-endian (largest-first) order */
432 	for (i = 0; i < 64; i++) {
433 		if (value & (1ull << i)) {
434 			hist[63 - i]++;
435 			++bits;
436 		}
437 	}
438 	/* update the count of bits changed */
439 	*count += bits;
440 }
441 
442 /*
443  * We've now filled up the range array, and need to increase "mingap" and
444  * shrink the range list accordingly.  zei_mingap is always the smallest
445  * distance between array entries, so we set the new_allowed_gap to be
446  * one greater than that.  We then go through the list, joining together
447  * any ranges which are closer than the new_allowed_gap.
448  *
449  * By construction, there will be at least one.  We also update zei_mingap
450  * to the new smallest gap, to prepare for our next invocation.
451  */
452 static void
453 zei_shrink_ranges(zfs_ecksum_info_t *eip)
454 {
455 	uint32_t mingap = UINT32_MAX;
456 	uint32_t new_allowed_gap = eip->zei_mingap + 1;
457 
458 	size_t idx, output;
459 	size_t max = eip->zei_range_count;
460 
461 	struct zei_ranges *r = eip->zei_ranges;
462 
463 	ASSERT3U(eip->zei_range_count, >, 0);
464 	ASSERT3U(eip->zei_range_count, <=, MAX_RANGES);
465 
466 	output = idx = 0;
467 	while (idx < max - 1) {
468 		uint32_t start = r[idx].zr_start;
469 		uint32_t end = r[idx].zr_end;
470 
471 		while (idx < max - 1) {
472 			idx++;
473 
474 			uint32_t nstart = r[idx].zr_start;
475 			uint32_t nend = r[idx].zr_end;
476 
477 			uint32_t gap = nstart - end;
478 			if (gap < new_allowed_gap) {
479 				end = nend;
480 				continue;
481 			}
482 			if (gap < mingap)
483 				mingap = gap;
484 			break;
485 		}
486 		r[output].zr_start = start;
487 		r[output].zr_end = end;
488 		output++;
489 	}
490 	ASSERT3U(output, <, eip->zei_range_count);
491 	eip->zei_range_count = output;
492 	eip->zei_mingap = mingap;
493 	eip->zei_allowed_mingap = new_allowed_gap;
494 }
495 
496 static void
497 zei_add_range(zfs_ecksum_info_t *eip, int start, int end)
498 {
499 	struct zei_ranges *r = eip->zei_ranges;
500 	size_t count = eip->zei_range_count;
501 
502 	if (count >= MAX_RANGES) {
503 		zei_shrink_ranges(eip);
504 		count = eip->zei_range_count;
505 	}
506 	if (count == 0) {
507 		eip->zei_mingap = UINT32_MAX;
508 		eip->zei_allowed_mingap = 1;
509 	} else {
510 		int gap = start - r[count - 1].zr_end;
511 
512 		if (gap < eip->zei_allowed_mingap) {
513 			r[count - 1].zr_end = end;
514 			return;
515 		}
516 		if (gap < eip->zei_mingap)
517 			eip->zei_mingap = gap;
518 	}
519 	r[count].zr_start = start;
520 	r[count].zr_end = end;
521 	eip->zei_range_count++;
522 }
523 
524 static size_t
525 zei_range_total_size(zfs_ecksum_info_t *eip)
526 {
527 	struct zei_ranges *r = eip->zei_ranges;
528 	size_t count = eip->zei_range_count;
529 	size_t result = 0;
530 	size_t idx;
531 
532 	for (idx = 0; idx < count; idx++)
533 		result += (r[idx].zr_end - r[idx].zr_start);
534 
535 	return (result);
536 }
537 
538 static zfs_ecksum_info_t *
539 annotate_ecksum(nvlist_t *ereport, zio_bad_cksum_t *info,
540     const abd_t *goodabd, const abd_t *badabd, size_t size,
541     boolean_t drop_if_identical)
542 {
543 	const uint64_t *good;
544 	const uint64_t *bad;
545 
546 	uint64_t allset = 0;
547 	uint64_t allcleared = 0;
548 
549 	size_t nui64s = size / sizeof (uint64_t);
550 
551 	size_t inline_size;
552 	int no_inline = 0;
553 	size_t idx;
554 	size_t range;
555 
556 	size_t offset = 0;
557 	ssize_t start = -1;
558 
559 	zfs_ecksum_info_t *eip = kmem_zalloc(sizeof (*eip), KM_SLEEP);
560 
561 	/* don't do any annotation for injected checksum errors */
562 	if (info != NULL && info->zbc_injected)
563 		return (eip);
564 
565 	if (info != NULL && info->zbc_has_cksum) {
566 		fm_payload_set(ereport,
567 		    FM_EREPORT_PAYLOAD_ZFS_CKSUM_EXPECTED,
568 		    DATA_TYPE_UINT64_ARRAY,
569 		    sizeof (info->zbc_expected) / sizeof (uint64_t),
570 		    (uint64_t *)&info->zbc_expected,
571 		    FM_EREPORT_PAYLOAD_ZFS_CKSUM_ACTUAL,
572 		    DATA_TYPE_UINT64_ARRAY,
573 		    sizeof (info->zbc_actual) / sizeof (uint64_t),
574 		    (uint64_t *)&info->zbc_actual,
575 		    FM_EREPORT_PAYLOAD_ZFS_CKSUM_ALGO,
576 		    DATA_TYPE_STRING,
577 		    info->zbc_checksum_name,
578 		    NULL);
579 
580 		if (info->zbc_byteswapped) {
581 			fm_payload_set(ereport,
582 			    FM_EREPORT_PAYLOAD_ZFS_CKSUM_BYTESWAP,
583 			    DATA_TYPE_BOOLEAN, 1,
584 			    NULL);
585 		}
586 	}
587 
588 	if (badabd == NULL || goodabd == NULL)
589 		return (eip);
590 
591 	ASSERT3U(nui64s, <=, UINT32_MAX);
592 	ASSERT3U(size, ==, nui64s * sizeof (uint64_t));
593 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
594 	ASSERT3U(size, <=, UINT32_MAX);
595 
596 	good = (const uint64_t *) abd_borrow_buf_copy((abd_t *)goodabd, size);
597 	bad = (const uint64_t *) abd_borrow_buf_copy((abd_t *)badabd, size);
598 
599 	/* build up the range list by comparing the two buffers. */
600 	for (idx = 0; idx < nui64s; idx++) {
601 		if (good[idx] == bad[idx]) {
602 			if (start == -1)
603 				continue;
604 
605 			zei_add_range(eip, start, idx);
606 			start = -1;
607 		} else {
608 			if (start != -1)
609 				continue;
610 
611 			start = idx;
612 		}
613 	}
614 	if (start != -1)
615 		zei_add_range(eip, start, idx);
616 
617 	/* See if it will fit in our inline buffers */
618 	inline_size = zei_range_total_size(eip);
619 	if (inline_size > ZFM_MAX_INLINE)
620 		no_inline = 1;
621 
622 	/*
623 	 * If there is no change and we want to drop if the buffers are
624 	 * identical, do so.
625 	 */
626 	if (inline_size == 0 && drop_if_identical) {
627 		kmem_free(eip, sizeof (*eip));
628 		abd_return_buf((abd_t *)goodabd, (void *)good, size);
629 		abd_return_buf((abd_t *)badabd, (void *)bad, size);
630 		return (NULL);
631 	}
632 
633 	/*
634 	 * Now walk through the ranges, filling in the details of the
635 	 * differences.  Also convert our uint64_t-array offsets to byte
636 	 * offsets.
637 	 */
638 	for (range = 0; range < eip->zei_range_count; range++) {
639 		size_t start = eip->zei_ranges[range].zr_start;
640 		size_t end = eip->zei_ranges[range].zr_end;
641 
642 		for (idx = start; idx < end; idx++) {
643 			uint64_t set, cleared;
644 
645 			// bits set in bad, but not in good
646 			set = ((~good[idx]) & bad[idx]);
647 			// bits set in good, but not in bad
648 			cleared = (good[idx] & (~bad[idx]));
649 
650 			allset |= set;
651 			allcleared |= cleared;
652 
653 			if (!no_inline) {
654 				ASSERT3U(offset, <, inline_size);
655 				eip->zei_bits_set[offset] = set;
656 				eip->zei_bits_cleared[offset] = cleared;
657 				offset++;
658 			}
659 
660 			update_histogram(set, eip->zei_histogram_set,
661 			    &eip->zei_range_sets[range]);
662 			update_histogram(cleared, eip->zei_histogram_cleared,
663 			    &eip->zei_range_clears[range]);
664 		}
665 
666 		/* convert to byte offsets */
667 		eip->zei_ranges[range].zr_start	*= sizeof (uint64_t);
668 		eip->zei_ranges[range].zr_end	*= sizeof (uint64_t);
669 	}
670 
671 	abd_return_buf((abd_t *)goodabd, (void *)good, size);
672 	abd_return_buf((abd_t *)badabd, (void *)bad, size);
673 
674 	eip->zei_allowed_mingap	*= sizeof (uint64_t);
675 	inline_size		*= sizeof (uint64_t);
676 
677 	/* fill in ereport */
678 	fm_payload_set(ereport,
679 	    FM_EREPORT_PAYLOAD_ZFS_BAD_OFFSET_RANGES,
680 	    DATA_TYPE_UINT32_ARRAY, 2 * eip->zei_range_count,
681 	    (uint32_t *)eip->zei_ranges,
682 	    FM_EREPORT_PAYLOAD_ZFS_BAD_RANGE_MIN_GAP,
683 	    DATA_TYPE_UINT32, eip->zei_allowed_mingap,
684 	    FM_EREPORT_PAYLOAD_ZFS_BAD_RANGE_SETS,
685 	    DATA_TYPE_UINT32_ARRAY, eip->zei_range_count, eip->zei_range_sets,
686 	    FM_EREPORT_PAYLOAD_ZFS_BAD_RANGE_CLEARS,
687 	    DATA_TYPE_UINT32_ARRAY, eip->zei_range_count, eip->zei_range_clears,
688 	    NULL);
689 
690 	if (!no_inline) {
691 		fm_payload_set(ereport,
692 		    FM_EREPORT_PAYLOAD_ZFS_BAD_SET_BITS,
693 		    DATA_TYPE_UINT8_ARRAY,
694 		    inline_size, (uint8_t *)eip->zei_bits_set,
695 		    FM_EREPORT_PAYLOAD_ZFS_BAD_CLEARED_BITS,
696 		    DATA_TYPE_UINT8_ARRAY,
697 		    inline_size, (uint8_t *)eip->zei_bits_cleared,
698 		    NULL);
699 	} else {
700 		fm_payload_set(ereport,
701 		    FM_EREPORT_PAYLOAD_ZFS_BAD_SET_HISTOGRAM,
702 		    DATA_TYPE_UINT32_ARRAY,
703 		    NBBY * sizeof (uint64_t), eip->zei_histogram_set,
704 		    FM_EREPORT_PAYLOAD_ZFS_BAD_CLEARED_HISTOGRAM,
705 		    DATA_TYPE_UINT32_ARRAY,
706 		    NBBY * sizeof (uint64_t), eip->zei_histogram_cleared,
707 		    NULL);
708 	}
709 	return (eip);
710 }
711 #endif
712 
713 /*
714  * Make sure our event is still valid for the given zio/vdev/pool.  For example,
715  * we don't want to keep logging events for a faulted or missing vdev.
716  */
717 boolean_t
718 zfs_ereport_is_valid(const char *subclass, spa_t *spa, vdev_t *vd, zio_t *zio)
719 {
720 #ifdef _KERNEL
721 	/*
722 	 * If we are doing a spa_tryimport() or in recovery mode,
723 	 * ignore errors.
724 	 */
725 	if (spa_load_state(spa) == SPA_LOAD_TRYIMPORT ||
726 	    spa_load_state(spa) == SPA_LOAD_RECOVER)
727 		return (B_FALSE);
728 
729 	/*
730 	 * If we are in the middle of opening a pool, and the previous attempt
731 	 * failed, don't bother logging any new ereports - we're just going to
732 	 * get the same diagnosis anyway.
733 	 */
734 	if (spa_load_state(spa) != SPA_LOAD_NONE &&
735 	    spa->spa_last_open_failed)
736 		return (B_FALSE);
737 
738 	if (zio != NULL) {
739 		/*
740 		 * If this is not a read or write zio, ignore the error.  This
741 		 * can occur if the DKIOCFLUSHWRITECACHE ioctl fails.
742 		 */
743 		if (zio->io_type != ZIO_TYPE_READ &&
744 		    zio->io_type != ZIO_TYPE_WRITE)
745 			return (B_FALSE);
746 
747 		if (vd != NULL) {
748 			/*
749 			 * If the vdev has already been marked as failing due
750 			 * to a failed probe, then ignore any subsequent I/O
751 			 * errors, as the DE will automatically fault the vdev
752 			 * on the first such failure.  This also catches cases
753 			 * where vdev_remove_wanted is set and the device has
754 			 * not yet been asynchronously placed into the REMOVED
755 			 * state.
756 			 */
757 			if (zio->io_vd == vd && !vdev_accessible(vd, zio))
758 				return (B_FALSE);
759 
760 			/*
761 			 * Ignore checksum errors for reads from DTL regions of
762 			 * leaf vdevs.
763 			 */
764 			if (zio->io_type == ZIO_TYPE_READ &&
765 			    zio->io_error == ECKSUM &&
766 			    vd->vdev_ops->vdev_op_leaf &&
767 			    vdev_dtl_contains(vd, DTL_MISSING, zio->io_txg, 1))
768 				return (B_FALSE);
769 		}
770 	}
771 
772 	/*
773 	 * For probe failure, we want to avoid posting ereports if we've
774 	 * already removed the device in the meantime.
775 	 */
776 	if (vd != NULL &&
777 	    strcmp(subclass, FM_EREPORT_ZFS_PROBE_FAILURE) == 0 &&
778 	    (vd->vdev_remove_wanted || vd->vdev_state == VDEV_STATE_REMOVED))
779 		return (B_FALSE);
780 
781 	/* Ignore bogus delay events (like from ioctls or unqueued IOs) */
782 	if ((strcmp(subclass, FM_EREPORT_ZFS_DELAY) == 0) &&
783 	    (zio != NULL) && (!zio->io_timestamp)) {
784 		return (B_FALSE);
785 	}
786 #endif
787 	return (B_TRUE);
788 }
789 
790 /*
791  * Return 0 if event was posted, EINVAL if there was a problem posting it or
792  * EBUSY if the event was rate limited.
793  */
794 int
795 zfs_ereport_post(const char *subclass, spa_t *spa, vdev_t *vd,
796     const zbookmark_phys_t *zb, zio_t *zio, uint64_t stateoroffset,
797     uint64_t size)
798 {
799 	int rc = 0;
800 #ifdef _KERNEL
801 	nvlist_t *ereport = NULL;
802 	nvlist_t *detector = NULL;
803 
804 	if (zfs_is_ratelimiting_event(subclass, vd))
805 		return (SET_ERROR(EBUSY));
806 
807 	if (!zfs_ereport_start(&ereport, &detector, subclass, spa, vd,
808 	    zb, zio, stateoroffset, size))
809 		return (SET_ERROR(EINVAL));	/* couldn't post event */
810 
811 	if (ereport == NULL)
812 		return (SET_ERROR(EINVAL));
813 
814 	/* Cleanup is handled by the callback function */
815 	rc = zfs_zevent_post(ereport, detector, zfs_zevent_post_cb);
816 #endif
817 	return (rc);
818 }
819 
820 void
821 zfs_ereport_start_checksum(spa_t *spa, vdev_t *vd, const zbookmark_phys_t *zb,
822     struct zio *zio, uint64_t offset, uint64_t length, void *arg,
823     zio_bad_cksum_t *info)
824 {
825 	zio_cksum_report_t *report;
826 
827 #ifdef _KERNEL
828 	if (zfs_is_ratelimiting_event(FM_EREPORT_ZFS_CHECKSUM, vd))
829 		return;
830 #endif
831 
832 	report = kmem_zalloc(sizeof (*report), KM_SLEEP);
833 
834 	if (zio->io_vsd != NULL)
835 		zio->io_vsd_ops->vsd_cksum_report(zio, report, arg);
836 	else
837 		zio_vsd_default_cksum_report(zio, report, arg);
838 
839 	/* copy the checksum failure information if it was provided */
840 	if (info != NULL) {
841 		report->zcr_ckinfo = kmem_zalloc(sizeof (*info), KM_SLEEP);
842 		bcopy(info, report->zcr_ckinfo, sizeof (*info));
843 	}
844 
845 	report->zcr_align = 1ULL << vd->vdev_top->vdev_ashift;
846 	report->zcr_length = length;
847 
848 #ifdef _KERNEL
849 	(void) zfs_ereport_start(&report->zcr_ereport, &report->zcr_detector,
850 	    FM_EREPORT_ZFS_CHECKSUM, spa, vd, zb, zio, offset, length);
851 
852 	if (report->zcr_ereport == NULL) {
853 		zfs_ereport_free_checksum(report);
854 		return;
855 	}
856 #endif
857 
858 	mutex_enter(&spa->spa_errlist_lock);
859 	report->zcr_next = zio->io_logical->io_cksum_report;
860 	zio->io_logical->io_cksum_report = report;
861 	mutex_exit(&spa->spa_errlist_lock);
862 }
863 
864 void
865 zfs_ereport_finish_checksum(zio_cksum_report_t *report, const abd_t *good_data,
866     const abd_t *bad_data, boolean_t drop_if_identical)
867 {
868 #ifdef _KERNEL
869 	zfs_ecksum_info_t *info;
870 
871 	info = annotate_ecksum(report->zcr_ereport, report->zcr_ckinfo,
872 	    good_data, bad_data, report->zcr_length, drop_if_identical);
873 	if (info != NULL)
874 		zfs_zevent_post(report->zcr_ereport,
875 		    report->zcr_detector, zfs_zevent_post_cb);
876 	else
877 		zfs_zevent_post_cb(report->zcr_ereport, report->zcr_detector);
878 
879 	report->zcr_ereport = report->zcr_detector = NULL;
880 	if (info != NULL)
881 		kmem_free(info, sizeof (*info));
882 #endif
883 }
884 
885 void
886 zfs_ereport_free_checksum(zio_cksum_report_t *rpt)
887 {
888 #ifdef _KERNEL
889 	if (rpt->zcr_ereport != NULL) {
890 		fm_nvlist_destroy(rpt->zcr_ereport,
891 		    FM_NVA_FREE);
892 		fm_nvlist_destroy(rpt->zcr_detector,
893 		    FM_NVA_FREE);
894 	}
895 #endif
896 	rpt->zcr_free(rpt->zcr_cbdata, rpt->zcr_cbinfo);
897 
898 	if (rpt->zcr_ckinfo != NULL)
899 		kmem_free(rpt->zcr_ckinfo, sizeof (*rpt->zcr_ckinfo));
900 
901 	kmem_free(rpt, sizeof (*rpt));
902 }
903 
904 
905 int
906 zfs_ereport_post_checksum(spa_t *spa, vdev_t *vd, const zbookmark_phys_t *zb,
907     struct zio *zio, uint64_t offset, uint64_t length,
908     const abd_t *good_data, const abd_t *bad_data, zio_bad_cksum_t *zbc)
909 {
910 	int rc = 0;
911 #ifdef _KERNEL
912 	nvlist_t *ereport = NULL;
913 	nvlist_t *detector = NULL;
914 	zfs_ecksum_info_t *info;
915 
916 	if (zfs_is_ratelimiting_event(FM_EREPORT_ZFS_CHECKSUM, vd))
917 		return (EBUSY);
918 
919 	if (!zfs_ereport_start(&ereport, &detector, FM_EREPORT_ZFS_CHECKSUM,
920 	    spa, vd, zb, zio, offset, length) || (ereport == NULL)) {
921 		return (SET_ERROR(EINVAL));
922 	}
923 
924 	info = annotate_ecksum(ereport, zbc, good_data, bad_data, length,
925 	    B_FALSE);
926 
927 	if (info != NULL) {
928 		rc = zfs_zevent_post(ereport, detector, zfs_zevent_post_cb);
929 		kmem_free(info, sizeof (*info));
930 	}
931 #endif
932 	return (rc);
933 }
934 
935 /*
936  * The 'sysevent.fs.zfs.*' events are signals posted to notify user space of
937  * change in the pool.  All sysevents are listed in sys/sysevent/eventdefs.h
938  * and are designed to be consumed by the ZFS Event Daemon (ZED).  For
939  * additional details refer to the zed(8) man page.
940  */
941 nvlist_t *
942 zfs_event_create(spa_t *spa, vdev_t *vd, const char *type, const char *name,
943     nvlist_t *aux)
944 {
945 	nvlist_t *resource = NULL;
946 #ifdef _KERNEL
947 	char class[64];
948 
949 	if (spa_load_state(spa) == SPA_LOAD_TRYIMPORT)
950 		return (NULL);
951 
952 	if ((resource = fm_nvlist_create(NULL)) == NULL)
953 		return (NULL);
954 
955 	(void) snprintf(class, sizeof (class), "%s.%s.%s", type,
956 	    ZFS_ERROR_CLASS, name);
957 	VERIFY0(nvlist_add_uint8(resource, FM_VERSION, FM_RSRC_VERSION));
958 	VERIFY0(nvlist_add_string(resource, FM_CLASS, class));
959 	VERIFY0(nvlist_add_string(resource,
960 	    FM_EREPORT_PAYLOAD_ZFS_POOL, spa_name(spa)));
961 	VERIFY0(nvlist_add_uint64(resource,
962 	    FM_EREPORT_PAYLOAD_ZFS_POOL_GUID, spa_guid(spa)));
963 	VERIFY0(nvlist_add_uint64(resource,
964 	    FM_EREPORT_PAYLOAD_ZFS_POOL_STATE, spa_state(spa)));
965 	VERIFY0(nvlist_add_int32(resource,
966 	    FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, spa_load_state(spa)));
967 
968 	if (vd) {
969 		VERIFY0(nvlist_add_uint64(resource,
970 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID, vd->vdev_guid));
971 		VERIFY0(nvlist_add_uint64(resource,
972 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_STATE, vd->vdev_state));
973 		if (vd->vdev_path != NULL)
974 			VERIFY0(nvlist_add_string(resource,
975 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_PATH, vd->vdev_path));
976 		if (vd->vdev_devid != NULL)
977 			VERIFY0(nvlist_add_string(resource,
978 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_DEVID, vd->vdev_devid));
979 		if (vd->vdev_fru != NULL)
980 			VERIFY0(nvlist_add_string(resource,
981 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_FRU, vd->vdev_fru));
982 		if (vd->vdev_enc_sysfs_path != NULL)
983 			VERIFY0(nvlist_add_string(resource,
984 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_ENC_SYSFS_PATH,
985 			    vd->vdev_enc_sysfs_path));
986 	}
987 
988 	/* also copy any optional payload data */
989 	if (aux) {
990 		nvpair_t *elem = NULL;
991 
992 		while ((elem = nvlist_next_nvpair(aux, elem)) != NULL)
993 			(void) nvlist_add_nvpair(resource, elem);
994 	}
995 
996 #endif
997 	return (resource);
998 }
999 
1000 static void
1001 zfs_post_common(spa_t *spa, vdev_t *vd, const char *type, const char *name,
1002     nvlist_t *aux)
1003 {
1004 #ifdef _KERNEL
1005 	nvlist_t *resource;
1006 
1007 	resource = zfs_event_create(spa, vd, type, name, aux);
1008 	if (resource)
1009 		zfs_zevent_post(resource, NULL, zfs_zevent_post_cb);
1010 #endif
1011 }
1012 
1013 /*
1014  * The 'resource.fs.zfs.removed' event is an internal signal that the given vdev
1015  * has been removed from the system.  This will cause the DE to ignore any
1016  * recent I/O errors, inferring that they are due to the asynchronous device
1017  * removal.
1018  */
1019 void
1020 zfs_post_remove(spa_t *spa, vdev_t *vd)
1021 {
1022 	zfs_post_common(spa, vd, FM_RSRC_CLASS, FM_RESOURCE_REMOVED, NULL);
1023 }
1024 
1025 /*
1026  * The 'resource.fs.zfs.autoreplace' event is an internal signal that the pool
1027  * has the 'autoreplace' property set, and therefore any broken vdevs will be
1028  * handled by higher level logic, and no vdev fault should be generated.
1029  */
1030 void
1031 zfs_post_autoreplace(spa_t *spa, vdev_t *vd)
1032 {
1033 	zfs_post_common(spa, vd, FM_RSRC_CLASS, FM_RESOURCE_AUTOREPLACE, NULL);
1034 }
1035 
1036 /*
1037  * The 'resource.fs.zfs.statechange' event is an internal signal that the
1038  * given vdev has transitioned its state to DEGRADED or HEALTHY.  This will
1039  * cause the retire agent to repair any outstanding fault management cases
1040  * open because the device was not found (fault.fs.zfs.device).
1041  */
1042 void
1043 zfs_post_state_change(spa_t *spa, vdev_t *vd, uint64_t laststate)
1044 {
1045 #ifdef _KERNEL
1046 	nvlist_t *aux;
1047 
1048 	/*
1049 	 * Add optional supplemental keys to payload
1050 	 */
1051 	aux = fm_nvlist_create(NULL);
1052 	if (vd && aux) {
1053 		if (vd->vdev_physpath) {
1054 			(void) nvlist_add_string(aux,
1055 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_PHYSPATH,
1056 			    vd->vdev_physpath);
1057 		}
1058 		if (vd->vdev_enc_sysfs_path) {
1059 			(void) nvlist_add_string(aux,
1060 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_ENC_SYSFS_PATH,
1061 			    vd->vdev_enc_sysfs_path);
1062 		}
1063 
1064 		(void) nvlist_add_uint64(aux,
1065 		    FM_EREPORT_PAYLOAD_ZFS_VDEV_LASTSTATE, laststate);
1066 	}
1067 
1068 	zfs_post_common(spa, vd, FM_RSRC_CLASS, FM_RESOURCE_STATECHANGE,
1069 	    aux);
1070 
1071 	if (aux)
1072 		fm_nvlist_destroy(aux, FM_NVA_FREE);
1073 #endif
1074 }
1075 
1076 #if defined(_KERNEL)
1077 EXPORT_SYMBOL(zfs_ereport_post);
1078 EXPORT_SYMBOL(zfs_ereport_is_valid);
1079 EXPORT_SYMBOL(zfs_ereport_post_checksum);
1080 EXPORT_SYMBOL(zfs_post_remove);
1081 EXPORT_SYMBOL(zfs_post_autoreplace);
1082 EXPORT_SYMBOL(zfs_post_state_change);
1083 #endif /* _KERNEL */
1084