xref: /freebsd/sys/contrib/openzfs/cmd/zed/agents/zfs_diagnosis.c (revision 61145dc2b94f12f6a47344fb9aac702321880e43)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 
23 /*
24  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
25  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
26  * Copyright (c) 2016, Intel Corporation.
27  * Copyright (c) 2023, Klara Inc.
28  */
29 
30 #include <stddef.h>
31 #include <string.h>
32 #include <libuutil.h>
33 #include <libzfs.h>
34 #include <sys/types.h>
35 #include <sys/time.h>
36 #include <sys/fs/zfs.h>
37 #include <sys/fm/protocol.h>
38 #include <sys/fm/fs/zfs.h>
39 #include <sys/zio.h>
40 
41 #include "zfs_agents.h"
42 #include "fmd_api.h"
43 
44 /*
45  * Default values for the serd engine when processing checksum or io errors. The
46  * semantics are N <events> in T <seconds>.
47  */
48 #define	DEFAULT_CHECKSUM_N	10	/* events */
49 #define	DEFAULT_CHECKSUM_T	600	/* seconds */
50 #define	DEFAULT_IO_N		10	/* events */
51 #define	DEFAULT_IO_T		600	/* seconds */
52 #define	DEFAULT_SLOW_IO_N	10	/* events */
53 #define	DEFAULT_SLOW_IO_T	30	/* seconds */
54 
55 #define	CASE_GC_TIMEOUT_SECS	43200	/* 12 hours */
56 
57 /*
58  * Our serd engines are named in the following format:
59  *     'zfs_<pool_guid>_<vdev_guid>_{checksum,io,slow_io}'
60  * This #define reserves enough space for two 64-bit hex values plus the
61  * length of the longest string.
62  */
63 #define	MAX_SERDLEN	(16 * 2 + sizeof ("zfs___checksum"))
64 
65 /*
66  * On-disk case structure.  This must maintain backwards compatibility with
67  * previous versions of the DE.  By default, any members appended to the end
68  * will be filled with zeros if they don't exist in a previous version.
69  */
70 typedef struct zfs_case_data {
71 	uint64_t	zc_version;
72 	uint64_t	zc_ena;
73 	uint64_t	zc_pool_guid;
74 	uint64_t	zc_vdev_guid;
75 	uint64_t	zc_parent_guid;
76 	int		zc_pool_state;
77 	char		zc_serd_checksum[MAX_SERDLEN];
78 	char		zc_serd_io[MAX_SERDLEN];
79 	char		zc_serd_slow_io[MAX_SERDLEN];
80 	int		zc_has_remove_timer;
81 } zfs_case_data_t;
82 
83 /*
84  * Time-of-day
85  */
86 typedef struct er_timeval {
87 	uint64_t	ertv_sec;
88 	uint64_t	ertv_nsec;
89 } er_timeval_t;
90 
91 /*
92  * In-core case structure.
93  */
94 typedef struct zfs_case {
95 	boolean_t	zc_present;
96 	uint32_t	zc_version;
97 	zfs_case_data_t	zc_data;
98 	fmd_case_t	*zc_case;
99 	uu_list_node_t	zc_node;
100 	id_t		zc_remove_timer;
101 	char		*zc_fru;
102 	er_timeval_t	zc_when;
103 } zfs_case_t;
104 
105 #define	CASE_DATA			"data"
106 #define	CASE_FRU			"fru"
107 #define	CASE_DATA_VERSION_INITIAL	1
108 #define	CASE_DATA_VERSION_SERD		2
109 
110 typedef struct zfs_de_stats {
111 	fmd_stat_t	old_drops;
112 	fmd_stat_t	dev_drops;
113 	fmd_stat_t	vdev_drops;
114 	fmd_stat_t	import_drops;
115 	fmd_stat_t	resource_drops;
116 } zfs_de_stats_t;
117 
118 zfs_de_stats_t zfs_stats = {
119 	{ "old_drops", FMD_TYPE_UINT64, "ereports dropped (from before load)" },
120 	{ "dev_drops", FMD_TYPE_UINT64, "ereports dropped (dev during open)"},
121 	{ "vdev_drops", FMD_TYPE_UINT64, "ereports dropped (weird vdev types)"},
122 	{ "import_drops", FMD_TYPE_UINT64, "ereports dropped (during import)" },
123 	{ "resource_drops", FMD_TYPE_UINT64, "resource related ereports" }
124 };
125 
126 /* wait 15 seconds after a removal */
127 static hrtime_t zfs_remove_timeout = SEC2NSEC(15);
128 
129 uu_list_pool_t *zfs_case_pool;
130 uu_list_t *zfs_cases;
131 
132 #define	ZFS_MAKE_RSRC(type)	\
133     FM_RSRC_CLASS "." ZFS_ERROR_CLASS "." type
134 #define	ZFS_MAKE_EREPORT(type)	\
135     FM_EREPORT_CLASS "." ZFS_ERROR_CLASS "." type
136 
137 static void zfs_purge_cases(fmd_hdl_t *hdl);
138 
139 /*
140  * Write out the persistent representation of an active case.
141  */
142 static void
zfs_case_serialize(zfs_case_t * zcp)143 zfs_case_serialize(zfs_case_t *zcp)
144 {
145 	zcp->zc_data.zc_version = CASE_DATA_VERSION_SERD;
146 }
147 
148 /*
149  * Read back the persistent representation of an active case.
150  */
151 static zfs_case_t *
zfs_case_unserialize(fmd_hdl_t * hdl,fmd_case_t * cp)152 zfs_case_unserialize(fmd_hdl_t *hdl, fmd_case_t *cp)
153 {
154 	zfs_case_t *zcp;
155 
156 	zcp = fmd_hdl_zalloc(hdl, sizeof (zfs_case_t), FMD_SLEEP);
157 	zcp->zc_case = cp;
158 
159 	fmd_buf_read(hdl, cp, CASE_DATA, &zcp->zc_data,
160 	    sizeof (zcp->zc_data));
161 
162 	if (zcp->zc_data.zc_version > CASE_DATA_VERSION_SERD) {
163 		fmd_hdl_free(hdl, zcp, sizeof (zfs_case_t));
164 		return (NULL);
165 	}
166 
167 	/*
168 	 * fmd_buf_read() will have already zeroed out the remainder of the
169 	 * buffer, so we don't have to do anything special if the version
170 	 * doesn't include the SERD engine name.
171 	 */
172 
173 	if (zcp->zc_data.zc_has_remove_timer)
174 		zcp->zc_remove_timer = fmd_timer_install(hdl, zcp,
175 		    NULL, zfs_remove_timeout);
176 
177 	uu_list_node_init(zcp, &zcp->zc_node, zfs_case_pool);
178 	(void) uu_list_insert_before(zfs_cases, NULL, zcp);
179 
180 	fmd_case_setspecific(hdl, cp, zcp);
181 
182 	return (zcp);
183 }
184 
185 /*
186  * Return count of other unique SERD cases under same vdev parent
187  */
188 static uint_t
zfs_other_serd_cases(fmd_hdl_t * hdl,const zfs_case_data_t * zfs_case)189 zfs_other_serd_cases(fmd_hdl_t *hdl, const zfs_case_data_t *zfs_case)
190 {
191 	zfs_case_t *zcp;
192 	uint_t cases = 0;
193 	static hrtime_t next_check = 0;
194 
195 	/*
196 	 * Note that plumbing in some external GC would require adding locking,
197 	 * since most of this module code is not thread safe and assumes there
198 	 * is only one thread running against the module. So we perform GC here
199 	 * inline periodically so that future delay induced faults will be
200 	 * possible once the issue causing multiple vdev delays is resolved.
201 	 */
202 	if (gethrestime_sec() > next_check) {
203 		/* Periodically purge old SERD entries and stale cases */
204 		fmd_serd_gc(hdl);
205 		zfs_purge_cases(hdl);
206 		next_check = gethrestime_sec() + CASE_GC_TIMEOUT_SECS;
207 	}
208 
209 	for (zcp = uu_list_first(zfs_cases); zcp != NULL;
210 	    zcp = uu_list_next(zfs_cases, zcp)) {
211 		zfs_case_data_t *zcd = &zcp->zc_data;
212 
213 		/*
214 		 * must be same pool and parent vdev but different leaf vdev
215 		 */
216 		if (zcd->zc_pool_guid != zfs_case->zc_pool_guid ||
217 		    zcd->zc_parent_guid != zfs_case->zc_parent_guid ||
218 		    zcd->zc_vdev_guid == zfs_case->zc_vdev_guid) {
219 			continue;
220 		}
221 
222 		/*
223 		 * Check if there is another active serd case besides zfs_case
224 		 *
225 		 * Only one serd engine will be assigned to the case
226 		 */
227 		if (zcd->zc_serd_checksum[0] == zfs_case->zc_serd_checksum[0] &&
228 		    fmd_serd_active(hdl, zcd->zc_serd_checksum)) {
229 			cases++;
230 		}
231 		if (zcd->zc_serd_io[0] == zfs_case->zc_serd_io[0] &&
232 		    fmd_serd_active(hdl, zcd->zc_serd_io)) {
233 			cases++;
234 		}
235 		if (zcd->zc_serd_slow_io[0] == zfs_case->zc_serd_slow_io[0] &&
236 		    fmd_serd_active(hdl, zcd->zc_serd_slow_io)) {
237 			cases++;
238 		}
239 	}
240 	return (cases);
241 }
242 
243 /*
244  * Iterate over any active cases.  If any cases are associated with a pool or
245  * vdev which is no longer present on the system, close the associated case.
246  */
247 static void
zfs_mark_vdev(uint64_t pool_guid,nvlist_t * vd,er_timeval_t * loaded)248 zfs_mark_vdev(uint64_t pool_guid, nvlist_t *vd, er_timeval_t *loaded)
249 {
250 	uint64_t vdev_guid = 0;
251 	uint_t c, children;
252 	nvlist_t **child;
253 	zfs_case_t *zcp;
254 
255 	(void) nvlist_lookup_uint64(vd, ZPOOL_CONFIG_GUID, &vdev_guid);
256 
257 	/*
258 	 * Mark any cases associated with this (pool, vdev) pair.
259 	 */
260 	for (zcp = uu_list_first(zfs_cases); zcp != NULL;
261 	    zcp = uu_list_next(zfs_cases, zcp)) {
262 		if (zcp->zc_data.zc_pool_guid == pool_guid &&
263 		    zcp->zc_data.zc_vdev_guid == vdev_guid) {
264 			zcp->zc_present = B_TRUE;
265 			zcp->zc_when = *loaded;
266 		}
267 	}
268 
269 	/*
270 	 * Iterate over all children.
271 	 */
272 	if (nvlist_lookup_nvlist_array(vd, ZPOOL_CONFIG_CHILDREN, &child,
273 	    &children) == 0) {
274 		for (c = 0; c < children; c++)
275 			zfs_mark_vdev(pool_guid, child[c], loaded);
276 	}
277 
278 	if (nvlist_lookup_nvlist_array(vd, ZPOOL_CONFIG_L2CACHE, &child,
279 	    &children) == 0) {
280 		for (c = 0; c < children; c++)
281 			zfs_mark_vdev(pool_guid, child[c], loaded);
282 	}
283 
284 	if (nvlist_lookup_nvlist_array(vd, ZPOOL_CONFIG_SPARES, &child,
285 	    &children) == 0) {
286 		for (c = 0; c < children; c++)
287 			zfs_mark_vdev(pool_guid, child[c], loaded);
288 	}
289 }
290 
291 static int
zfs_mark_pool(zpool_handle_t * zhp,void * unused)292 zfs_mark_pool(zpool_handle_t *zhp, void *unused)
293 {
294 	(void) unused;
295 	zfs_case_t *zcp;
296 	uint64_t pool_guid;
297 	uint64_t *tod;
298 	er_timeval_t loaded = { 0 };
299 	nvlist_t *config, *vd;
300 	uint_t nelem = 0;
301 	int ret;
302 
303 	pool_guid = zpool_get_prop_int(zhp, ZPOOL_PROP_GUID, NULL);
304 	/*
305 	 * Mark any cases associated with just this pool.
306 	 */
307 	for (zcp = uu_list_first(zfs_cases); zcp != NULL;
308 	    zcp = uu_list_next(zfs_cases, zcp)) {
309 		if (zcp->zc_data.zc_pool_guid == pool_guid &&
310 		    zcp->zc_data.zc_vdev_guid == 0)
311 			zcp->zc_present = B_TRUE;
312 	}
313 
314 	if ((config = zpool_get_config(zhp, NULL)) == NULL) {
315 		zpool_close(zhp);
316 		return (-1);
317 	}
318 
319 	(void) nvlist_lookup_uint64_array(config, ZPOOL_CONFIG_LOADED_TIME,
320 	    &tod, &nelem);
321 	if (nelem == 2) {
322 		loaded.ertv_sec = tod[0];
323 		loaded.ertv_nsec = tod[1];
324 		for (zcp = uu_list_first(zfs_cases); zcp != NULL;
325 		    zcp = uu_list_next(zfs_cases, zcp)) {
326 			if (zcp->zc_data.zc_pool_guid == pool_guid &&
327 			    zcp->zc_data.zc_vdev_guid == 0) {
328 				zcp->zc_when = loaded;
329 			}
330 		}
331 	}
332 
333 	ret = nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &vd);
334 	if (ret) {
335 		zpool_close(zhp);
336 		return (-1);
337 	}
338 
339 	zfs_mark_vdev(pool_guid, vd, &loaded);
340 
341 	zpool_close(zhp);
342 
343 	return (0);
344 }
345 
346 struct load_time_arg {
347 	uint64_t lt_guid;
348 	er_timeval_t *lt_time;
349 	boolean_t lt_found;
350 };
351 
352 static int
zpool_find_load_time(zpool_handle_t * zhp,void * arg)353 zpool_find_load_time(zpool_handle_t *zhp, void *arg)
354 {
355 	struct load_time_arg *lta = arg;
356 	uint64_t pool_guid;
357 	uint64_t *tod;
358 	nvlist_t *config;
359 	uint_t nelem;
360 
361 	if (lta->lt_found) {
362 		zpool_close(zhp);
363 		return (0);
364 	}
365 
366 	pool_guid = zpool_get_prop_int(zhp, ZPOOL_PROP_GUID, NULL);
367 	if (pool_guid != lta->lt_guid) {
368 		zpool_close(zhp);
369 		return (0);
370 	}
371 
372 	if ((config = zpool_get_config(zhp, NULL)) == NULL) {
373 		zpool_close(zhp);
374 		return (-1);
375 	}
376 
377 	if (nvlist_lookup_uint64_array(config, ZPOOL_CONFIG_LOADED_TIME,
378 	    &tod, &nelem) == 0 && nelem == 2) {
379 		lta->lt_found = B_TRUE;
380 		lta->lt_time->ertv_sec = tod[0];
381 		lta->lt_time->ertv_nsec = tod[1];
382 	}
383 
384 	zpool_close(zhp);
385 
386 	return (0);
387 }
388 
389 static void
zfs_purge_cases(fmd_hdl_t * hdl)390 zfs_purge_cases(fmd_hdl_t *hdl)
391 {
392 	zfs_case_t *zcp;
393 	uu_list_walk_t *walk;
394 	libzfs_handle_t *zhdl = fmd_hdl_getspecific(hdl);
395 
396 	/*
397 	 * There is no way to open a pool by GUID, or lookup a vdev by GUID.  No
398 	 * matter what we do, we're going to have to stomach an O(vdevs * cases)
399 	 * algorithm.  In reality, both quantities are likely so small that
400 	 * neither will matter. Given that iterating over pools is more
401 	 * expensive than iterating over the in-memory case list, we opt for a
402 	 * 'present' flag in each case that starts off cleared.  We then iterate
403 	 * over all pools, marking those that are still present, and removing
404 	 * those that aren't found.
405 	 *
406 	 * Note that we could also construct an FMRI and rely on
407 	 * fmd_nvl_fmri_present(), but this would end up doing the same search.
408 	 */
409 
410 	/*
411 	 * Mark the cases as not present.
412 	 */
413 	for (zcp = uu_list_first(zfs_cases); zcp != NULL;
414 	    zcp = uu_list_next(zfs_cases, zcp))
415 		zcp->zc_present = B_FALSE;
416 
417 	/*
418 	 * Iterate over all pools and mark the pools and vdevs found.  If this
419 	 * fails (most probably because we're out of memory), then don't close
420 	 * any of the cases and we cannot be sure they are accurate.
421 	 */
422 	if (zpool_iter(zhdl, zfs_mark_pool, NULL) != 0)
423 		return;
424 
425 	/*
426 	 * Remove those cases which were not found.
427 	 */
428 	walk = uu_list_walk_start(zfs_cases, UU_WALK_ROBUST);
429 	while ((zcp = uu_list_walk_next(walk)) != NULL) {
430 		if (!zcp->zc_present)
431 			fmd_case_close(hdl, zcp->zc_case);
432 	}
433 	uu_list_walk_end(walk);
434 }
435 
436 /*
437  * Construct the name of a serd engine given the pool/vdev GUID and type (io or
438  * checksum).
439  */
440 static void
zfs_serd_name(char * buf,uint64_t pool_guid,uint64_t vdev_guid,const char * type)441 zfs_serd_name(char *buf, uint64_t pool_guid, uint64_t vdev_guid,
442     const char *type)
443 {
444 	(void) snprintf(buf, MAX_SERDLEN, "zfs_%llx_%llx_%s",
445 	    (long long unsigned int)pool_guid,
446 	    (long long unsigned int)vdev_guid, type);
447 }
448 
449 static void
zfs_case_retire(fmd_hdl_t * hdl,zfs_case_t * zcp)450 zfs_case_retire(fmd_hdl_t *hdl, zfs_case_t *zcp)
451 {
452 	fmd_hdl_debug(hdl, "retiring case");
453 
454 	fmd_case_close(hdl, zcp->zc_case);
455 }
456 
457 /*
458  * Solve a given ZFS case.  This first checks to make sure the diagnosis is
459  * still valid, as well as cleaning up any pending timer associated with the
460  * case.
461  */
462 static void
zfs_case_solve(fmd_hdl_t * hdl,zfs_case_t * zcp,const char * faultname)463 zfs_case_solve(fmd_hdl_t *hdl, zfs_case_t *zcp, const char *faultname)
464 {
465 	nvlist_t *detector, *fault;
466 	boolean_t serialize;
467 	nvlist_t *fru = NULL;
468 	fmd_hdl_debug(hdl, "solving fault '%s'", faultname);
469 
470 	/*
471 	 * Construct the detector from the case data.  The detector is in the
472 	 * ZFS scheme, and is either the pool or the vdev, depending on whether
473 	 * this is a vdev or pool fault.
474 	 */
475 	detector = fmd_nvl_alloc(hdl, FMD_SLEEP);
476 
477 	(void) nvlist_add_uint8(detector, FM_VERSION, ZFS_SCHEME_VERSION0);
478 	(void) nvlist_add_string(detector, FM_FMRI_SCHEME, FM_FMRI_SCHEME_ZFS);
479 	(void) nvlist_add_uint64(detector, FM_FMRI_ZFS_POOL,
480 	    zcp->zc_data.zc_pool_guid);
481 	if (zcp->zc_data.zc_vdev_guid != 0) {
482 		(void) nvlist_add_uint64(detector, FM_FMRI_ZFS_VDEV,
483 		    zcp->zc_data.zc_vdev_guid);
484 	}
485 
486 	fault = fmd_nvl_create_fault(hdl, faultname, 100, detector,
487 	    fru, detector);
488 	fmd_case_add_suspect(hdl, zcp->zc_case, fault);
489 
490 	nvlist_free(fru);
491 
492 	fmd_case_solve(hdl, zcp->zc_case);
493 
494 	serialize = B_FALSE;
495 	if (zcp->zc_data.zc_has_remove_timer) {
496 		fmd_timer_remove(hdl, zcp->zc_remove_timer);
497 		zcp->zc_data.zc_has_remove_timer = 0;
498 		serialize = B_TRUE;
499 	}
500 	if (serialize)
501 		zfs_case_serialize(zcp);
502 
503 	nvlist_free(detector);
504 }
505 
506 static boolean_t
timeval_earlier(er_timeval_t * a,er_timeval_t * b)507 timeval_earlier(er_timeval_t *a, er_timeval_t *b)
508 {
509 	return (a->ertv_sec < b->ertv_sec ||
510 	    (a->ertv_sec == b->ertv_sec && a->ertv_nsec < b->ertv_nsec));
511 }
512 
513 static void
zfs_ereport_when(fmd_hdl_t * hdl,nvlist_t * nvl,er_timeval_t * when)514 zfs_ereport_when(fmd_hdl_t *hdl, nvlist_t *nvl, er_timeval_t *when)
515 {
516 	(void) hdl;
517 	int64_t *tod;
518 	uint_t	nelem;
519 
520 	if (nvlist_lookup_int64_array(nvl, FM_EREPORT_TIME, &tod,
521 	    &nelem) == 0 && nelem == 2) {
522 		when->ertv_sec = tod[0];
523 		when->ertv_nsec = tod[1];
524 	} else {
525 		when->ertv_sec = when->ertv_nsec = UINT64_MAX;
526 	}
527 }
528 
529 /*
530  * Record the specified event in the SERD engine and return a
531  * boolean value indicating whether or not the engine fired as
532  * the result of inserting this event.
533  *
534  * When the pool has similar active cases on other vdevs, then
535  * the fired state is disregarded and the case is retired.
536  */
537 static int
zfs_fm_serd_record(fmd_hdl_t * hdl,const char * name,fmd_event_t * ep,zfs_case_t * zcp,const char * err_type)538 zfs_fm_serd_record(fmd_hdl_t *hdl, const char *name, fmd_event_t *ep,
539     zfs_case_t *zcp, const char *err_type)
540 {
541 	int fired = fmd_serd_record(hdl, name, ep);
542 	int peers = 0;
543 
544 	if (fired && (peers = zfs_other_serd_cases(hdl, &zcp->zc_data)) > 0) {
545 		fmd_hdl_debug(hdl, "pool %llu is tracking %d other %s cases "
546 		    "-- skip faulting the vdev %llu",
547 		    (u_longlong_t)zcp->zc_data.zc_pool_guid,
548 		    peers, err_type,
549 		    (u_longlong_t)zcp->zc_data.zc_vdev_guid);
550 		zfs_case_retire(hdl, zcp);
551 		fired = 0;
552 	}
553 
554 	return (fired);
555 }
556 
557 /*
558  * Main fmd entry point.
559  */
560 static void
zfs_fm_recv(fmd_hdl_t * hdl,fmd_event_t * ep,nvlist_t * nvl,const char * class)561 zfs_fm_recv(fmd_hdl_t *hdl, fmd_event_t *ep, nvlist_t *nvl, const char *class)
562 {
563 	zfs_case_t *zcp, *dcp;
564 	int32_t pool_state;
565 	uint64_t ena, pool_guid, vdev_guid, parent_guid;
566 	uint64_t checksum_n, checksum_t;
567 	uint64_t io_n, io_t;
568 	er_timeval_t pool_load;
569 	er_timeval_t er_when;
570 	nvlist_t *detector;
571 	boolean_t pool_found = B_FALSE;
572 	boolean_t isresource;
573 	const char *type;
574 
575 	/*
576 	 * We subscribe to notifications for vdev or pool removal.  In these
577 	 * cases, there may be cases that no longer apply.  Purge any cases
578 	 * that no longer apply.
579 	 */
580 	if (fmd_nvl_class_match(hdl, nvl, "sysevent.fs.zfs.*")) {
581 		fmd_hdl_debug(hdl, "purging orphaned cases from %s",
582 		    strrchr(class, '.') + 1);
583 		zfs_purge_cases(hdl);
584 		zfs_stats.resource_drops.fmds_value.ui64++;
585 		return;
586 	}
587 
588 	isresource = fmd_nvl_class_match(hdl, nvl, "resource.fs.zfs.*");
589 
590 	if (isresource) {
591 		/*
592 		 * For resources, we don't have a normal payload.
593 		 */
594 		if (nvlist_lookup_uint64(nvl, FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID,
595 		    &vdev_guid) != 0)
596 			pool_state = SPA_LOAD_OPEN;
597 		else
598 			pool_state = SPA_LOAD_NONE;
599 		detector = NULL;
600 	} else {
601 		(void) nvlist_lookup_nvlist(nvl,
602 		    FM_EREPORT_DETECTOR, &detector);
603 		(void) nvlist_lookup_int32(nvl,
604 		    FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, &pool_state);
605 	}
606 
607 	/*
608 	 * We also ignore all ereports generated during an import of a pool,
609 	 * since the only possible fault (.pool) would result in import failure,
610 	 * and hence no persistent fault.  Some day we may want to do something
611 	 * with these ereports, so we continue generating them internally.
612 	 */
613 	if (pool_state == SPA_LOAD_IMPORT) {
614 		zfs_stats.import_drops.fmds_value.ui64++;
615 		fmd_hdl_debug(hdl, "ignoring '%s' during import", class);
616 		return;
617 	}
618 
619 	/*
620 	 * Device I/O errors are ignored during pool open.
621 	 */
622 	if (pool_state == SPA_LOAD_OPEN &&
623 	    (fmd_nvl_class_match(hdl, nvl,
624 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_CHECKSUM)) ||
625 	    fmd_nvl_class_match(hdl, nvl,
626 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_IO)) ||
627 	    fmd_nvl_class_match(hdl, nvl,
628 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_PROBE_FAILURE)))) {
629 		fmd_hdl_debug(hdl, "ignoring '%s' during pool open", class);
630 		zfs_stats.dev_drops.fmds_value.ui64++;
631 		return;
632 	}
633 
634 	/*
635 	 * We ignore ereports for anything except disks and files.
636 	 */
637 	if (nvlist_lookup_string(nvl, FM_EREPORT_PAYLOAD_ZFS_VDEV_TYPE,
638 	    &type) == 0) {
639 		if (strcmp(type, VDEV_TYPE_DISK) != 0 &&
640 		    strcmp(type, VDEV_TYPE_FILE) != 0) {
641 			zfs_stats.vdev_drops.fmds_value.ui64++;
642 			return;
643 		}
644 	}
645 
646 	/*
647 	 * Determine if this ereport corresponds to an open case.
648 	 * Each vdev or pool can have a single case.
649 	 */
650 	(void) nvlist_lookup_uint64(nvl,
651 	    FM_EREPORT_PAYLOAD_ZFS_POOL_GUID, &pool_guid);
652 	if (nvlist_lookup_uint64(nvl,
653 	    FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID, &vdev_guid) != 0)
654 		vdev_guid = 0;
655 	if (nvlist_lookup_uint64(nvl,
656 	    FM_EREPORT_PAYLOAD_ZFS_PARENT_GUID, &parent_guid) != 0)
657 		parent_guid = 0;
658 	if (nvlist_lookup_uint64(nvl, FM_EREPORT_ENA, &ena) != 0)
659 		ena = 0;
660 
661 	zfs_ereport_when(hdl, nvl, &er_when);
662 
663 	for (zcp = uu_list_first(zfs_cases); zcp != NULL;
664 	    zcp = uu_list_next(zfs_cases, zcp)) {
665 		if (zcp->zc_data.zc_pool_guid == pool_guid) {
666 			pool_found = B_TRUE;
667 			pool_load = zcp->zc_when;
668 		}
669 		if (zcp->zc_data.zc_vdev_guid == vdev_guid)
670 			break;
671 	}
672 
673 	/*
674 	 * Avoid falsely accusing a pool of being faulty.  Do so by
675 	 * not replaying ereports that were generated prior to the
676 	 * current import.  If the failure that generated them was
677 	 * transient because the device was actually removed but we
678 	 * didn't receive the normal asynchronous notification, we
679 	 * don't want to mark it as faulted and potentially panic. If
680 	 * there is still a problem we'd expect not to be able to
681 	 * import the pool, or that new ereports will be generated
682 	 * once the pool is used.
683 	 */
684 	if (pool_found && timeval_earlier(&er_when, &pool_load)) {
685 		fmd_hdl_debug(hdl, "ignoring pool %llx, "
686 		    "ereport time %lld.%lld, pool load time = %lld.%lld",
687 		    pool_guid, er_when.ertv_sec, er_when.ertv_nsec,
688 		    pool_load.ertv_sec, pool_load.ertv_nsec);
689 		zfs_stats.old_drops.fmds_value.ui64++;
690 		return;
691 	}
692 
693 	if (!pool_found) {
694 		/*
695 		 * Haven't yet seen this pool, but same situation
696 		 * may apply.
697 		 */
698 		libzfs_handle_t *zhdl = fmd_hdl_getspecific(hdl);
699 		struct load_time_arg la;
700 
701 		la.lt_guid = pool_guid;
702 		la.lt_time = &pool_load;
703 		la.lt_found = B_FALSE;
704 
705 		if (zhdl != NULL &&
706 		    zpool_iter(zhdl, zpool_find_load_time, &la) == 0 &&
707 		    la.lt_found == B_TRUE) {
708 			pool_found = B_TRUE;
709 
710 			if (timeval_earlier(&er_when, &pool_load)) {
711 				fmd_hdl_debug(hdl, "ignoring pool %llx, "
712 				    "ereport time %lld.%lld, "
713 				    "pool load time = %lld.%lld",
714 				    pool_guid, er_when.ertv_sec,
715 				    er_when.ertv_nsec, pool_load.ertv_sec,
716 				    pool_load.ertv_nsec);
717 				zfs_stats.old_drops.fmds_value.ui64++;
718 				return;
719 			}
720 		}
721 	}
722 
723 	if (zcp == NULL) {
724 		fmd_case_t *cs;
725 		zfs_case_data_t data = { 0 };
726 
727 		/*
728 		 * If this is one of our 'fake' resource ereports, and there is
729 		 * no case open, simply discard it.
730 		 */
731 		if (isresource) {
732 			zfs_stats.resource_drops.fmds_value.ui64++;
733 			fmd_hdl_debug(hdl, "discarding '%s for vdev %llu",
734 			    class, vdev_guid);
735 			return;
736 		}
737 
738 		/*
739 		 * Skip tracking some ereports
740 		 */
741 		if (strcmp(class,
742 		    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_DATA)) == 0 ||
743 		    strcmp(class,
744 		    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_CONFIG_CACHE_WRITE)) == 0) {
745 			zfs_stats.resource_drops.fmds_value.ui64++;
746 			return;
747 		}
748 
749 		/*
750 		 * Open a new case.
751 		 */
752 		cs = fmd_case_open(hdl, NULL);
753 
754 		fmd_hdl_debug(hdl, "opening case for vdev %llu due to '%s'",
755 		    vdev_guid, class);
756 
757 		/*
758 		 * Initialize the case buffer.  To commonize code, we actually
759 		 * create the buffer with existing data, and then call
760 		 * zfs_case_unserialize() to instantiate the in-core structure.
761 		 */
762 		fmd_buf_create(hdl, cs, CASE_DATA, sizeof (zfs_case_data_t));
763 
764 		data.zc_version = CASE_DATA_VERSION_SERD;
765 		data.zc_ena = ena;
766 		data.zc_pool_guid = pool_guid;
767 		data.zc_vdev_guid = vdev_guid;
768 		data.zc_parent_guid = parent_guid;
769 		data.zc_pool_state = (int)pool_state;
770 
771 		fmd_buf_write(hdl, cs, CASE_DATA, &data, sizeof (data));
772 
773 		zcp = zfs_case_unserialize(hdl, cs);
774 		assert(zcp != NULL);
775 		if (pool_found)
776 			zcp->zc_when = pool_load;
777 	}
778 
779 	if (isresource) {
780 		fmd_hdl_debug(hdl, "resource event '%s'", class);
781 
782 		if (fmd_nvl_class_match(hdl, nvl,
783 		    ZFS_MAKE_RSRC(FM_RESOURCE_AUTOREPLACE))) {
784 			/*
785 			 * The 'resource.fs.zfs.autoreplace' event indicates
786 			 * that the pool was loaded with the 'autoreplace'
787 			 * property set.  In this case, any pending device
788 			 * failures should be ignored, as the asynchronous
789 			 * autoreplace handling will take care of them.
790 			 */
791 			fmd_case_close(hdl, zcp->zc_case);
792 		} else if (fmd_nvl_class_match(hdl, nvl,
793 		    ZFS_MAKE_RSRC(FM_RESOURCE_REMOVED))) {
794 			/*
795 			 * The 'resource.fs.zfs.removed' event indicates that
796 			 * device removal was detected, and the device was
797 			 * closed asynchronously.  If this is the case, we
798 			 * assume that any recent I/O errors were due to the
799 			 * device removal, not any fault of the device itself.
800 			 * We reset the SERD engine, and cancel any pending
801 			 * timers.
802 			 */
803 			if (zcp->zc_data.zc_has_remove_timer) {
804 				fmd_timer_remove(hdl, zcp->zc_remove_timer);
805 				zcp->zc_data.zc_has_remove_timer = 0;
806 				zfs_case_serialize(zcp);
807 			}
808 			if (zcp->zc_data.zc_serd_io[0] != '\0')
809 				fmd_serd_reset(hdl, zcp->zc_data.zc_serd_io);
810 			if (zcp->zc_data.zc_serd_checksum[0] != '\0')
811 				fmd_serd_reset(hdl,
812 				    zcp->zc_data.zc_serd_checksum);
813 			if (zcp->zc_data.zc_serd_slow_io[0] != '\0')
814 				fmd_serd_reset(hdl,
815 				    zcp->zc_data.zc_serd_slow_io);
816 		} else if (fmd_nvl_class_match(hdl, nvl,
817 		    ZFS_MAKE_RSRC(FM_RESOURCE_STATECHANGE))) {
818 			uint64_t state = 0;
819 
820 			if (zcp != NULL &&
821 			    nvlist_lookup_uint64(nvl,
822 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_STATE, &state) == 0 &&
823 			    state == VDEV_STATE_HEALTHY) {
824 				fmd_hdl_debug(hdl, "closing case after a "
825 				    "device statechange to healthy");
826 				fmd_case_close(hdl, zcp->zc_case);
827 			}
828 		}
829 		zfs_stats.resource_drops.fmds_value.ui64++;
830 		return;
831 	}
832 
833 	/*
834 	 * Associate the ereport with this case.
835 	 */
836 	fmd_case_add_ereport(hdl, zcp->zc_case, ep);
837 
838 	/*
839 	 * Don't do anything else if this case is already solved.
840 	 */
841 	if (fmd_case_solved(hdl, zcp->zc_case))
842 		return;
843 
844 	if (vdev_guid)
845 		fmd_hdl_debug(hdl, "error event '%s', vdev %llu", class,
846 		    vdev_guid);
847 	else
848 		fmd_hdl_debug(hdl, "error event '%s'", class);
849 
850 	/*
851 	 * Determine if we should solve the case and generate a fault.  We solve
852 	 * a case if:
853 	 *
854 	 * 	a. A pool failed to open (ereport.fs.zfs.pool)
855 	 * 	b. A device failed to open (ereport.fs.zfs.pool) while a pool
856 	 *	   was up and running.
857 	 *
858 	 * We may see a series of ereports associated with a pool open, all
859 	 * chained together by the same ENA.  If the pool open succeeds, then
860 	 * we'll see no further ereports.  To detect when a pool open has
861 	 * succeeded, we associate a timer with the event.  When it expires, we
862 	 * close the case.
863 	 */
864 	if (fmd_nvl_class_match(hdl, nvl,
865 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_POOL))) {
866 		/*
867 		 * Pool level fault.  Before solving the case, go through and
868 		 * close any open device cases that may be pending.
869 		 */
870 		for (dcp = uu_list_first(zfs_cases); dcp != NULL;
871 		    dcp = uu_list_next(zfs_cases, dcp)) {
872 			if (dcp->zc_data.zc_pool_guid ==
873 			    zcp->zc_data.zc_pool_guid &&
874 			    dcp->zc_data.zc_vdev_guid != 0)
875 				fmd_case_close(hdl, dcp->zc_case);
876 		}
877 
878 		zfs_case_solve(hdl, zcp, "fault.fs.zfs.pool");
879 	} else if (fmd_nvl_class_match(hdl, nvl,
880 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_LOG_REPLAY))) {
881 		/*
882 		 * Pool level fault for reading the intent logs.
883 		 */
884 		zfs_case_solve(hdl, zcp, "fault.fs.zfs.log_replay");
885 	} else if (fmd_nvl_class_match(hdl, nvl, "ereport.fs.zfs.vdev.*")) {
886 		/*
887 		 * Device fault.
888 		 */
889 		zfs_case_solve(hdl, zcp, "fault.fs.zfs.device");
890 	} else if (fmd_nvl_class_match(hdl, nvl,
891 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_IO)) ||
892 	    fmd_nvl_class_match(hdl, nvl,
893 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_CHECKSUM)) ||
894 	    fmd_nvl_class_match(hdl, nvl,
895 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_IO_FAILURE)) ||
896 	    fmd_nvl_class_match(hdl, nvl,
897 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_DELAY)) ||
898 	    fmd_nvl_class_match(hdl, nvl,
899 	    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_PROBE_FAILURE))) {
900 		const char *failmode = NULL;
901 		boolean_t checkremove = B_FALSE;
902 		uint32_t pri = 0;
903 
904 		/*
905 		 * If this is a checksum or I/O error, then toss it into the
906 		 * appropriate SERD engine and check to see if it has fired.
907 		 * Ideally, we want to do something more sophisticated,
908 		 * (persistent errors for a single data block, etc).  For now,
909 		 * a single SERD engine is sufficient.
910 		 */
911 		if (fmd_nvl_class_match(hdl, nvl,
912 		    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_IO))) {
913 			if (zcp->zc_data.zc_serd_io[0] == '\0') {
914 				if (nvlist_lookup_uint64(nvl,
915 				    FM_EREPORT_PAYLOAD_ZFS_VDEV_IO_N,
916 				    &io_n) != 0) {
917 					io_n = DEFAULT_IO_N;
918 				}
919 				if (nvlist_lookup_uint64(nvl,
920 				    FM_EREPORT_PAYLOAD_ZFS_VDEV_IO_T,
921 				    &io_t) != 0) {
922 					io_t = DEFAULT_IO_T;
923 				}
924 				zfs_serd_name(zcp->zc_data.zc_serd_io,
925 				    pool_guid, vdev_guid, "io");
926 				fmd_serd_create(hdl, zcp->zc_data.zc_serd_io,
927 				    io_n,
928 				    SEC2NSEC(io_t));
929 				zfs_case_serialize(zcp);
930 			}
931 			if (zfs_fm_serd_record(hdl, zcp->zc_data.zc_serd_io,
932 			    ep, zcp, "io error")) {
933 				checkremove = B_TRUE;
934 			}
935 		} else if (fmd_nvl_class_match(hdl, nvl,
936 		    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_DELAY))) {
937 			uint64_t slow_io_n, slow_io_t;
938 
939 			/*
940 			 * Create a slow io SERD engine when the VDEV has the
941 			 * 'vdev_slow_io_n' and 'vdev_slow_io_n' properties.
942 			 */
943 			if (zcp->zc_data.zc_serd_slow_io[0] == '\0' &&
944 			    nvlist_lookup_uint64(nvl,
945 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_SLOW_IO_N,
946 			    &slow_io_n) == 0 &&
947 			    nvlist_lookup_uint64(nvl,
948 			    FM_EREPORT_PAYLOAD_ZFS_VDEV_SLOW_IO_T,
949 			    &slow_io_t) == 0) {
950 				zfs_serd_name(zcp->zc_data.zc_serd_slow_io,
951 				    pool_guid, vdev_guid, "slow_io");
952 				fmd_serd_create(hdl,
953 				    zcp->zc_data.zc_serd_slow_io,
954 				    slow_io_n,
955 				    SEC2NSEC(slow_io_t));
956 				zfs_case_serialize(zcp);
957 			}
958 			/* Pass event to SERD engine and see if this triggers */
959 			if (zcp->zc_data.zc_serd_slow_io[0] != '\0' &&
960 			    zfs_fm_serd_record(hdl,
961 			    zcp->zc_data.zc_serd_slow_io, ep, zcp, "slow io")) {
962 				zfs_case_solve(hdl, zcp,
963 				    "fault.fs.zfs.vdev.slow_io");
964 			}
965 		} else if (fmd_nvl_class_match(hdl, nvl,
966 		    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_CHECKSUM))) {
967 			uint64_t flags = 0;
968 			int32_t flags32 = 0;
969 			/*
970 			 * We ignore ereports for checksum errors generated by
971 			 * scrub/resilver I/O to avoid potentially further
972 			 * degrading the pool while it's being repaired.
973 			 *
974 			 * Note that FM_EREPORT_PAYLOAD_ZFS_ZIO_FLAGS used to
975 			 * be int32. To allow newer zed to work on older
976 			 * kernels, if we don't find the flags, we look for
977 			 * the older ones too.
978 			 */
979 			if (((nvlist_lookup_uint32(nvl,
980 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_PRIORITY, &pri) == 0) &&
981 			    (pri == ZIO_PRIORITY_SCRUB ||
982 			    pri == ZIO_PRIORITY_REBUILD)) ||
983 			    ((nvlist_lookup_uint64(nvl,
984 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_FLAGS, &flags) == 0) &&
985 			    (flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER))) ||
986 			    ((nvlist_lookup_int32(nvl,
987 			    FM_EREPORT_PAYLOAD_ZFS_ZIO_FLAGS, &flags32) == 0) &&
988 			    (flags32 & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER)))) {
989 				fmd_hdl_debug(hdl, "ignoring '%s' for "
990 				    "scrub/resilver I/O", class);
991 				return;
992 			}
993 
994 			if (zcp->zc_data.zc_serd_checksum[0] == '\0') {
995 				if (nvlist_lookup_uint64(nvl,
996 				    FM_EREPORT_PAYLOAD_ZFS_VDEV_CKSUM_N,
997 				    &checksum_n) != 0) {
998 					checksum_n = DEFAULT_CHECKSUM_N;
999 				}
1000 				if (nvlist_lookup_uint64(nvl,
1001 				    FM_EREPORT_PAYLOAD_ZFS_VDEV_CKSUM_T,
1002 				    &checksum_t) != 0) {
1003 					checksum_t = DEFAULT_CHECKSUM_T;
1004 				}
1005 
1006 				zfs_serd_name(zcp->zc_data.zc_serd_checksum,
1007 				    pool_guid, vdev_guid, "checksum");
1008 				fmd_serd_create(hdl,
1009 				    zcp->zc_data.zc_serd_checksum,
1010 				    checksum_n,
1011 				    SEC2NSEC(checksum_t));
1012 				zfs_case_serialize(zcp);
1013 			}
1014 			if (zfs_fm_serd_record(hdl,
1015 			    zcp->zc_data.zc_serd_checksum, ep, zcp,
1016 			    "checksum")) {
1017 				zfs_case_solve(hdl, zcp,
1018 				    "fault.fs.zfs.vdev.checksum");
1019 			}
1020 		} else if (fmd_nvl_class_match(hdl, nvl,
1021 		    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_IO_FAILURE)) &&
1022 		    (nvlist_lookup_string(nvl,
1023 		    FM_EREPORT_PAYLOAD_ZFS_POOL_FAILMODE, &failmode) == 0) &&
1024 		    failmode != NULL) {
1025 			if (strncmp(failmode, FM_EREPORT_FAILMODE_CONTINUE,
1026 			    strlen(FM_EREPORT_FAILMODE_CONTINUE)) == 0) {
1027 				zfs_case_solve(hdl, zcp,
1028 				    "fault.fs.zfs.io_failure_continue");
1029 			} else if (strncmp(failmode, FM_EREPORT_FAILMODE_WAIT,
1030 			    strlen(FM_EREPORT_FAILMODE_WAIT)) == 0) {
1031 				zfs_case_solve(hdl, zcp,
1032 				    "fault.fs.zfs.io_failure_wait");
1033 			}
1034 		} else if (fmd_nvl_class_match(hdl, nvl,
1035 		    ZFS_MAKE_EREPORT(FM_EREPORT_ZFS_PROBE_FAILURE))) {
1036 #ifndef __linux__
1037 			/* This causes an unexpected fault diagnosis on linux */
1038 			checkremove = B_TRUE;
1039 #endif
1040 		}
1041 
1042 		/*
1043 		 * Because I/O errors may be due to device removal, we postpone
1044 		 * any diagnosis until we're sure that we aren't about to
1045 		 * receive a 'resource.fs.zfs.removed' event.
1046 		 */
1047 		if (checkremove) {
1048 			if (zcp->zc_data.zc_has_remove_timer)
1049 				fmd_timer_remove(hdl, zcp->zc_remove_timer);
1050 			zcp->zc_remove_timer = fmd_timer_install(hdl, zcp, NULL,
1051 			    zfs_remove_timeout);
1052 			if (!zcp->zc_data.zc_has_remove_timer) {
1053 				zcp->zc_data.zc_has_remove_timer = 1;
1054 				zfs_case_serialize(zcp);
1055 			}
1056 		}
1057 	}
1058 }
1059 
1060 /*
1061  * The timeout is fired when we diagnosed an I/O error, and it was not due to
1062  * device removal (which would cause the timeout to be cancelled).
1063  */
1064 static void
zfs_fm_timeout(fmd_hdl_t * hdl,id_t id,void * data)1065 zfs_fm_timeout(fmd_hdl_t *hdl, id_t id, void *data)
1066 {
1067 	zfs_case_t *zcp = data;
1068 
1069 	if (id == zcp->zc_remove_timer)
1070 		zfs_case_solve(hdl, zcp, "fault.fs.zfs.vdev.io");
1071 }
1072 
1073 /*
1074  * The specified case has been closed and any case-specific
1075  * data structures should be deallocated.
1076  */
1077 static void
zfs_fm_close(fmd_hdl_t * hdl,fmd_case_t * cs)1078 zfs_fm_close(fmd_hdl_t *hdl, fmd_case_t *cs)
1079 {
1080 	zfs_case_t *zcp = fmd_case_getspecific(hdl, cs);
1081 
1082 	if (zcp->zc_data.zc_serd_checksum[0] != '\0')
1083 		fmd_serd_destroy(hdl, zcp->zc_data.zc_serd_checksum);
1084 	if (zcp->zc_data.zc_serd_io[0] != '\0')
1085 		fmd_serd_destroy(hdl, zcp->zc_data.zc_serd_io);
1086 	if (zcp->zc_data.zc_serd_slow_io[0] != '\0')
1087 		fmd_serd_destroy(hdl, zcp->zc_data.zc_serd_slow_io);
1088 	if (zcp->zc_data.zc_has_remove_timer)
1089 		fmd_timer_remove(hdl, zcp->zc_remove_timer);
1090 
1091 	uu_list_remove(zfs_cases, zcp);
1092 	uu_list_node_fini(zcp, &zcp->zc_node, zfs_case_pool);
1093 	fmd_hdl_free(hdl, zcp, sizeof (zfs_case_t));
1094 }
1095 
1096 static const fmd_hdl_ops_t fmd_ops = {
1097 	zfs_fm_recv,	/* fmdo_recv */
1098 	zfs_fm_timeout,	/* fmdo_timeout */
1099 	zfs_fm_close,	/* fmdo_close */
1100 	NULL,		/* fmdo_stats */
1101 	NULL,	/* fmdo_gc */
1102 };
1103 
1104 static const fmd_prop_t fmd_props[] = {
1105 	{ NULL, 0, NULL }
1106 };
1107 
1108 static const fmd_hdl_info_t fmd_info = {
1109 	"ZFS Diagnosis Engine", "1.0", &fmd_ops, fmd_props
1110 };
1111 
1112 void
_zfs_diagnosis_init(fmd_hdl_t * hdl)1113 _zfs_diagnosis_init(fmd_hdl_t *hdl)
1114 {
1115 	libzfs_handle_t *zhdl;
1116 
1117 	if ((zhdl = libzfs_init()) == NULL)
1118 		return;
1119 
1120 	if ((zfs_case_pool = uu_list_pool_create("zfs_case_pool",
1121 	    sizeof (zfs_case_t), offsetof(zfs_case_t, zc_node),
1122 	    NULL, UU_LIST_POOL_DEBUG)) == NULL) {
1123 		libzfs_fini(zhdl);
1124 		return;
1125 	}
1126 
1127 	if ((zfs_cases = uu_list_create(zfs_case_pool, NULL,
1128 	    UU_LIST_DEBUG)) == NULL) {
1129 		uu_list_pool_destroy(zfs_case_pool);
1130 		libzfs_fini(zhdl);
1131 		return;
1132 	}
1133 
1134 	if (fmd_hdl_register(hdl, FMD_API_VERSION, &fmd_info) != 0) {
1135 		uu_list_destroy(zfs_cases);
1136 		uu_list_pool_destroy(zfs_case_pool);
1137 		libzfs_fini(zhdl);
1138 		return;
1139 	}
1140 
1141 	fmd_hdl_setspecific(hdl, zhdl);
1142 
1143 	(void) fmd_stat_create(hdl, FMD_STAT_NOALLOC, sizeof (zfs_stats) /
1144 	    sizeof (fmd_stat_t), (fmd_stat_t *)&zfs_stats);
1145 }
1146 
1147 void
_zfs_diagnosis_fini(fmd_hdl_t * hdl)1148 _zfs_diagnosis_fini(fmd_hdl_t *hdl)
1149 {
1150 	zfs_case_t *zcp;
1151 	uu_list_walk_t *walk;
1152 	libzfs_handle_t *zhdl;
1153 
1154 	/*
1155 	 * Remove all active cases.
1156 	 */
1157 	walk = uu_list_walk_start(zfs_cases, UU_WALK_ROBUST);
1158 	while ((zcp = uu_list_walk_next(walk)) != NULL) {
1159 		fmd_hdl_debug(hdl, "removing case ena %llu",
1160 		    (long long unsigned)zcp->zc_data.zc_ena);
1161 		uu_list_remove(zfs_cases, zcp);
1162 		uu_list_node_fini(zcp, &zcp->zc_node, zfs_case_pool);
1163 		fmd_hdl_free(hdl, zcp, sizeof (zfs_case_t));
1164 	}
1165 	uu_list_walk_end(walk);
1166 
1167 	uu_list_destroy(zfs_cases);
1168 	uu_list_pool_destroy(zfs_case_pool);
1169 
1170 	zhdl = fmd_hdl_getspecific(hdl);
1171 	libzfs_fini(zhdl);
1172 }
1173