xref: /freebsd/sys/contrib/openzfs/cmd/zed/agents/zfs_mod.c (revision 2008043f386721d58158e37e0d7e50df8095942d)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012 by Delphix. All rights reserved.
24  * Copyright 2014 Nexenta Systems, Inc. All rights reserved.
25  * Copyright (c) 2016, 2017, Intel Corporation.
26  * Copyright (c) 2017 Open-E, Inc. All Rights Reserved.
27  * Copyright (c) 2023, Klara Inc.
28  */
29 
30 /*
31  * ZFS syseventd module.
32  *
33  * file origin: openzfs/usr/src/cmd/syseventd/modules/zfs_mod/zfs_mod.c
34  *
35  * The purpose of this module is to identify when devices are added to the
36  * system, and appropriately online or replace the affected vdevs.
37  *
38  * When a device is added to the system:
39  *
40  * 	1. Search for any vdevs whose devid matches that of the newly added
41  *	   device.
42  *
43  * 	2. If no vdevs are found, then search for any vdevs whose udev path
44  *	   matches that of the new device.
45  *
46  *	3. If no vdevs match by either method, then ignore the event.
47  *
48  * 	4. Attempt to online the device with a flag to indicate that it should
49  *	   be unspared when resilvering completes.  If this succeeds, then the
50  *	   same device was inserted and we should continue normally.
51  *
52  *	5. If the pool does not have the 'autoreplace' property set, attempt to
53  *	   online the device again without the unspare flag, which will
54  *	   generate a FMA fault.
55  *
56  *	6. If the pool has the 'autoreplace' property set, and the matching vdev
57  *	   is a whole disk, then label the new disk and attempt a 'zpool
58  *	   replace'.
59  *
60  * The module responds to EC_DEV_ADD events.  The special ESC_ZFS_VDEV_CHECK
61  * event indicates that a device failed to open during pool load, but the
62  * autoreplace property was set.  In this case, we deferred the associated
63  * FMA fault until our module had a chance to process the autoreplace logic.
64  * If the device could not be replaced, then the second online attempt will
65  * trigger the FMA fault that we skipped earlier.
66  *
67  * On Linux udev provides a disk insert for both the disk and the partition.
68  */
69 
70 #include <ctype.h>
71 #include <fcntl.h>
72 #include <libnvpair.h>
73 #include <libzfs.h>
74 #include <libzutil.h>
75 #include <limits.h>
76 #include <stddef.h>
77 #include <stdlib.h>
78 #include <string.h>
79 #include <syslog.h>
80 #include <sys/list.h>
81 #include <sys/sunddi.h>
82 #include <sys/sysevent/eventdefs.h>
83 #include <sys/sysevent/dev.h>
84 #include <thread_pool.h>
85 #include <pthread.h>
86 #include <unistd.h>
87 #include <errno.h>
88 #include "zfs_agents.h"
89 #include "../zed_log.h"
90 
91 #define	DEV_BYID_PATH	"/dev/disk/by-id/"
92 #define	DEV_BYPATH_PATH	"/dev/disk/by-path/"
93 #define	DEV_BYVDEV_PATH	"/dev/disk/by-vdev/"
94 
95 typedef void (*zfs_process_func_t)(zpool_handle_t *, nvlist_t *, boolean_t);
96 
97 libzfs_handle_t *g_zfshdl;
98 list_t g_pool_list;	/* list of unavailable pools at initialization */
99 list_t g_device_list;	/* list of disks with asynchronous label request */
100 tpool_t *g_tpool;
101 boolean_t g_enumeration_done;
102 pthread_t g_zfs_tid;	/* zfs_enum_pools() thread */
103 
104 typedef struct unavailpool {
105 	zpool_handle_t	*uap_zhp;
106 	list_node_t	uap_node;
107 } unavailpool_t;
108 
109 typedef struct pendingdev {
110 	char		pd_physpath[128];
111 	list_node_t	pd_node;
112 } pendingdev_t;
113 
114 static int
115 zfs_toplevel_state(zpool_handle_t *zhp)
116 {
117 	nvlist_t *nvroot;
118 	vdev_stat_t *vs;
119 	unsigned int c;
120 
121 	verify(nvlist_lookup_nvlist(zpool_get_config(zhp, NULL),
122 	    ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0);
123 	verify(nvlist_lookup_uint64_array(nvroot, ZPOOL_CONFIG_VDEV_STATS,
124 	    (uint64_t **)&vs, &c) == 0);
125 	return (vs->vs_state);
126 }
127 
128 static int
129 zfs_unavail_pool(zpool_handle_t *zhp, void *data)
130 {
131 	zed_log_msg(LOG_INFO, "zfs_unavail_pool: examining '%s' (state %d)",
132 	    zpool_get_name(zhp), (int)zfs_toplevel_state(zhp));
133 
134 	if (zfs_toplevel_state(zhp) < VDEV_STATE_DEGRADED) {
135 		unavailpool_t *uap;
136 		uap = malloc(sizeof (unavailpool_t));
137 		if (uap == NULL) {
138 			perror("malloc");
139 			exit(EXIT_FAILURE);
140 		}
141 
142 		uap->uap_zhp = zhp;
143 		list_insert_tail((list_t *)data, uap);
144 	} else {
145 		zpool_close(zhp);
146 	}
147 	return (0);
148 }
149 
150 /*
151  * Write an array of strings to the zed log
152  */
153 static void lines_to_zed_log_msg(char **lines, int lines_cnt)
154 {
155 	int i;
156 	for (i = 0; i < lines_cnt; i++) {
157 		zed_log_msg(LOG_INFO, "%s", lines[i]);
158 	}
159 }
160 
161 /*
162  * Two stage replace on Linux
163  * since we get disk notifications
164  * we can wait for partitioned disk slice to show up!
165  *
166  * First stage tags the disk, initiates async partitioning, and returns
167  * Second stage finds the tag and proceeds to ZFS labeling/replace
168  *
169  * disk-add --> label-disk + tag-disk --> partition-add --> zpool_vdev_attach
170  *
171  * 1. physical match with no fs, no partition
172  *	tag it top, partition disk
173  *
174  * 2. physical match again, see partition and tag
175  *
176  */
177 
178 /*
179  * The device associated with the given vdev (either by devid or physical path)
180  * has been added to the system.  If 'isdisk' is set, then we only attempt a
181  * replacement if it's a whole disk.  This also implies that we should label the
182  * disk first.
183  *
184  * First, we attempt to online the device (making sure to undo any spare
185  * operation when finished).  If this succeeds, then we're done.  If it fails,
186  * and the new state is VDEV_CANT_OPEN, it indicates that the device was opened,
187  * but that the label was not what we expected.  If the 'autoreplace' property
188  * is enabled, then we relabel the disk (if specified), and attempt a 'zpool
189  * replace'.  If the online is successful, but the new state is something else
190  * (REMOVED or FAULTED), it indicates that we're out of sync or in some sort of
191  * race, and we should avoid attempting to relabel the disk.
192  *
193  * Also can arrive here from a ESC_ZFS_VDEV_CHECK event
194  */
195 static void
196 zfs_process_add(zpool_handle_t *zhp, nvlist_t *vdev, boolean_t labeled)
197 {
198 	const char *path;
199 	vdev_state_t newstate;
200 	nvlist_t *nvroot, *newvd;
201 	pendingdev_t *device;
202 	uint64_t wholedisk = 0ULL;
203 	uint64_t offline = 0ULL, faulted = 0ULL;
204 	uint64_t guid = 0ULL;
205 	uint64_t is_spare = 0;
206 	const char *physpath = NULL, *new_devid = NULL, *enc_sysfs_path = NULL;
207 	char rawpath[PATH_MAX], fullpath[PATH_MAX];
208 	char pathbuf[PATH_MAX];
209 	int ret;
210 	int online_flag = ZFS_ONLINE_CHECKREMOVE | ZFS_ONLINE_UNSPARE;
211 	boolean_t is_sd = B_FALSE;
212 	boolean_t is_mpath_wholedisk = B_FALSE;
213 	uint_t c;
214 	vdev_stat_t *vs;
215 	char **lines = NULL;
216 	int lines_cnt = 0;
217 
218 	/*
219 	 * Get the persistent path, typically under the '/dev/disk/by-id' or
220 	 * '/dev/disk/by-vdev' directories.  Note that this path can change
221 	 * when a vdev is replaced with a new disk.
222 	 */
223 	if (nvlist_lookup_string(vdev, ZPOOL_CONFIG_PATH, &path) != 0)
224 		return;
225 
226 	/* Skip healthy disks */
227 	verify(nvlist_lookup_uint64_array(vdev, ZPOOL_CONFIG_VDEV_STATS,
228 	    (uint64_t **)&vs, &c) == 0);
229 	if (vs->vs_state == VDEV_STATE_HEALTHY) {
230 		zed_log_msg(LOG_INFO, "%s: %s is already healthy, skip it.",
231 		    __func__, path);
232 		return;
233 	}
234 
235 	(void) nvlist_lookup_string(vdev, ZPOOL_CONFIG_PHYS_PATH, &physpath);
236 	(void) nvlist_lookup_string(vdev, ZPOOL_CONFIG_VDEV_ENC_SYSFS_PATH,
237 	    &enc_sysfs_path);
238 	(void) nvlist_lookup_uint64(vdev, ZPOOL_CONFIG_WHOLE_DISK, &wholedisk);
239 	(void) nvlist_lookup_uint64(vdev, ZPOOL_CONFIG_OFFLINE, &offline);
240 	(void) nvlist_lookup_uint64(vdev, ZPOOL_CONFIG_FAULTED, &faulted);
241 
242 	(void) nvlist_lookup_uint64(vdev, ZPOOL_CONFIG_GUID, &guid);
243 	(void) nvlist_lookup_uint64(vdev, ZPOOL_CONFIG_IS_SPARE, &is_spare);
244 
245 	/*
246 	 * Special case:
247 	 *
248 	 * We've seen times where a disk won't have a ZPOOL_CONFIG_PHYS_PATH
249 	 * entry in their config. For example, on this force-faulted disk:
250 	 *
251 	 *	children[0]:
252 	 *	   type: 'disk'
253 	 *	   id: 0
254 	 *	   guid: 14309659774640089719
255 	 *        path: '/dev/disk/by-vdev/L28'
256 	 *        whole_disk: 0
257 	 *        DTL: 654
258 	 *        create_txg: 4
259 	 *        com.delphix:vdev_zap_leaf: 1161
260 	 *        faulted: 1
261 	 *        aux_state: 'external'
262 	 *	children[1]:
263 	 *        type: 'disk'
264 	 *        id: 1
265 	 *        guid: 16002508084177980912
266 	 *        path: '/dev/disk/by-vdev/L29'
267 	 *        devid: 'dm-uuid-mpath-35000c500a61d68a3'
268 	 *        phys_path: 'L29'
269 	 *        vdev_enc_sysfs_path: '/sys/class/enclosure/0:0:1:0/SLOT 30 32'
270 	 *        whole_disk: 0
271 	 *        DTL: 1028
272 	 *        create_txg: 4
273 	 *        com.delphix:vdev_zap_leaf: 131
274 	 *
275 	 * If the disk's path is a /dev/disk/by-vdev/ path, then we can infer
276 	 * the ZPOOL_CONFIG_PHYS_PATH from the by-vdev disk name.
277 	 */
278 	if (physpath == NULL && path != NULL) {
279 		/* If path begins with "/dev/disk/by-vdev/" ... */
280 		if (strncmp(path, DEV_BYVDEV_PATH,
281 		    strlen(DEV_BYVDEV_PATH)) == 0) {
282 			/* Set physpath to the char after "/dev/disk/by-vdev" */
283 			physpath = &path[strlen(DEV_BYVDEV_PATH)];
284 		}
285 	}
286 
287 	/*
288 	 * We don't want to autoreplace offlined disks.  However, we do want to
289 	 * replace force-faulted disks (`zpool offline -f`).  Force-faulted
290 	 * disks have both offline=1 and faulted=1 in the nvlist.
291 	 */
292 	if (offline && !faulted) {
293 		zed_log_msg(LOG_INFO, "%s: %s is offline, skip autoreplace",
294 		    __func__, path);
295 		return;
296 	}
297 
298 	is_mpath_wholedisk = is_mpath_whole_disk(path);
299 	zed_log_msg(LOG_INFO, "zfs_process_add: pool '%s' vdev '%s', phys '%s'"
300 	    " %s blank disk, %s mpath blank disk, %s labeled, enc sysfs '%s', "
301 	    "(guid %llu)",
302 	    zpool_get_name(zhp), path,
303 	    physpath ? physpath : "NULL",
304 	    wholedisk ? "is" : "not",
305 	    is_mpath_wholedisk? "is" : "not",
306 	    labeled ? "is" : "not",
307 	    enc_sysfs_path,
308 	    (long long unsigned int)guid);
309 
310 	/*
311 	 * The VDEV guid is preferred for identification (gets passed in path)
312 	 */
313 	if (guid != 0) {
314 		(void) snprintf(fullpath, sizeof (fullpath), "%llu",
315 		    (long long unsigned int)guid);
316 	} else {
317 		/*
318 		 * otherwise use path sans partition suffix for whole disks
319 		 */
320 		(void) strlcpy(fullpath, path, sizeof (fullpath));
321 		if (wholedisk) {
322 			char *spath = zfs_strip_partition(fullpath);
323 			if (!spath) {
324 				zed_log_msg(LOG_INFO, "%s: Can't alloc",
325 				    __func__);
326 				return;
327 			}
328 
329 			(void) strlcpy(fullpath, spath, sizeof (fullpath));
330 			free(spath);
331 		}
332 	}
333 
334 	if (is_spare)
335 		online_flag |= ZFS_ONLINE_SPARE;
336 
337 	/*
338 	 * Attempt to online the device.
339 	 */
340 	if (zpool_vdev_online(zhp, fullpath, online_flag, &newstate) == 0 &&
341 	    (newstate == VDEV_STATE_HEALTHY ||
342 	    newstate == VDEV_STATE_DEGRADED)) {
343 		zed_log_msg(LOG_INFO,
344 		    "  zpool_vdev_online: vdev '%s' ('%s') is "
345 		    "%s", fullpath, physpath, (newstate == VDEV_STATE_HEALTHY) ?
346 		    "HEALTHY" : "DEGRADED");
347 		return;
348 	}
349 
350 	/*
351 	 * vdev_id alias rule for using scsi_debug devices (FMA automated
352 	 * testing)
353 	 */
354 	if (physpath != NULL && strcmp("scsidebug", physpath) == 0)
355 		is_sd = B_TRUE;
356 
357 	/*
358 	 * If the pool doesn't have the autoreplace property set, then use
359 	 * vdev online to trigger a FMA fault by posting an ereport.
360 	 */
361 	if (!zpool_get_prop_int(zhp, ZPOOL_PROP_AUTOREPLACE, NULL) ||
362 	    !(wholedisk || is_mpath_wholedisk) || (physpath == NULL)) {
363 		(void) zpool_vdev_online(zhp, fullpath, ZFS_ONLINE_FORCEFAULT,
364 		    &newstate);
365 		zed_log_msg(LOG_INFO, "Pool's autoreplace is not enabled or "
366 		    "not a blank disk for '%s' ('%s')", fullpath,
367 		    physpath);
368 		return;
369 	}
370 
371 	/*
372 	 * Convert physical path into its current device node.  Rawpath
373 	 * needs to be /dev/disk/by-vdev for a scsi_debug device since
374 	 * /dev/disk/by-path will not be present.
375 	 */
376 	(void) snprintf(rawpath, sizeof (rawpath), "%s%s",
377 	    is_sd ? DEV_BYVDEV_PATH : DEV_BYPATH_PATH, physpath);
378 
379 	if (realpath(rawpath, pathbuf) == NULL && !is_mpath_wholedisk) {
380 		zed_log_msg(LOG_INFO, "  realpath: %s failed (%s)",
381 		    rawpath, strerror(errno));
382 
383 		int err = zpool_vdev_online(zhp, fullpath,
384 		    ZFS_ONLINE_FORCEFAULT, &newstate);
385 
386 		zed_log_msg(LOG_INFO, "  zpool_vdev_online: %s FORCEFAULT (%s) "
387 		    "err %d, new state %d",
388 		    fullpath, libzfs_error_description(g_zfshdl), err,
389 		    err ? (int)newstate : 0);
390 		return;
391 	}
392 
393 	/* Only autoreplace bad disks */
394 	if ((vs->vs_state != VDEV_STATE_DEGRADED) &&
395 	    (vs->vs_state != VDEV_STATE_FAULTED) &&
396 	    (vs->vs_state != VDEV_STATE_REMOVED) &&
397 	    (vs->vs_state != VDEV_STATE_CANT_OPEN)) {
398 		zed_log_msg(LOG_INFO, "  not autoreplacing since disk isn't in "
399 		    "a bad state (currently %llu)", vs->vs_state);
400 		return;
401 	}
402 
403 	nvlist_lookup_string(vdev, "new_devid", &new_devid);
404 
405 	if (is_mpath_wholedisk) {
406 		/* Don't label device mapper or multipath disks. */
407 		zed_log_msg(LOG_INFO,
408 		    "  it's a multipath wholedisk, don't label");
409 		if (zpool_prepare_disk(zhp, vdev, "autoreplace", &lines,
410 		    &lines_cnt) != 0) {
411 			zed_log_msg(LOG_INFO,
412 			    "  zpool_prepare_disk: could not "
413 			    "prepare '%s' (%s)", fullpath,
414 			    libzfs_error_description(g_zfshdl));
415 			if (lines_cnt > 0) {
416 				zed_log_msg(LOG_INFO,
417 				    "  zfs_prepare_disk output:");
418 				lines_to_zed_log_msg(lines, lines_cnt);
419 			}
420 			libzfs_free_str_array(lines, lines_cnt);
421 			return;
422 		}
423 	} else if (!labeled) {
424 		/*
425 		 * we're auto-replacing a raw disk, so label it first
426 		 */
427 		char *leafname;
428 
429 		/*
430 		 * If this is a request to label a whole disk, then attempt to
431 		 * write out the label.  Before we can label the disk, we need
432 		 * to map the physical string that was matched on to the under
433 		 * lying device node.
434 		 *
435 		 * If any part of this process fails, then do a force online
436 		 * to trigger a ZFS fault for the device (and any hot spare
437 		 * replacement).
438 		 */
439 		leafname = strrchr(pathbuf, '/') + 1;
440 
441 		/*
442 		 * If this is a request to label a whole disk, then attempt to
443 		 * write out the label.
444 		 */
445 		if (zpool_prepare_and_label_disk(g_zfshdl, zhp, leafname,
446 		    vdev, "autoreplace", &lines, &lines_cnt) != 0) {
447 			zed_log_msg(LOG_WARNING,
448 			    "  zpool_prepare_and_label_disk: could not "
449 			    "label '%s' (%s)", leafname,
450 			    libzfs_error_description(g_zfshdl));
451 			if (lines_cnt > 0) {
452 				zed_log_msg(LOG_INFO,
453 				"  zfs_prepare_disk output:");
454 				lines_to_zed_log_msg(lines, lines_cnt);
455 			}
456 			libzfs_free_str_array(lines, lines_cnt);
457 
458 			(void) zpool_vdev_online(zhp, fullpath,
459 			    ZFS_ONLINE_FORCEFAULT, &newstate);
460 			return;
461 		}
462 
463 		/*
464 		 * The disk labeling is asynchronous on Linux. Just record
465 		 * this label request and return as there will be another
466 		 * disk add event for the partition after the labeling is
467 		 * completed.
468 		 */
469 		device = malloc(sizeof (pendingdev_t));
470 		if (device == NULL) {
471 			perror("malloc");
472 			exit(EXIT_FAILURE);
473 		}
474 
475 		(void) strlcpy(device->pd_physpath, physpath,
476 		    sizeof (device->pd_physpath));
477 		list_insert_tail(&g_device_list, device);
478 
479 		zed_log_msg(LOG_NOTICE, "  zpool_label_disk: async '%s' (%llu)",
480 		    leafname, (u_longlong_t)guid);
481 
482 		return;	/* resumes at EC_DEV_ADD.ESC_DISK for partition */
483 
484 	} else /* labeled */ {
485 		boolean_t found = B_FALSE;
486 		/*
487 		 * match up with request above to label the disk
488 		 */
489 		for (device = list_head(&g_device_list); device != NULL;
490 		    device = list_next(&g_device_list, device)) {
491 			if (strcmp(physpath, device->pd_physpath) == 0) {
492 				list_remove(&g_device_list, device);
493 				free(device);
494 				found = B_TRUE;
495 				break;
496 			}
497 			zed_log_msg(LOG_INFO, "zpool_label_disk: %s != %s",
498 			    physpath, device->pd_physpath);
499 		}
500 		if (!found) {
501 			/* unexpected partition slice encountered */
502 			zed_log_msg(LOG_WARNING, "labeled disk %s was "
503 			    "unexpected here", fullpath);
504 			(void) zpool_vdev_online(zhp, fullpath,
505 			    ZFS_ONLINE_FORCEFAULT, &newstate);
506 			return;
507 		}
508 
509 		zed_log_msg(LOG_INFO, "  zpool_label_disk: resume '%s' (%llu)",
510 		    physpath, (u_longlong_t)guid);
511 
512 		/*
513 		 * Paths that begin with '/dev/disk/by-id/' will change and so
514 		 * they must be updated before calling zpool_vdev_attach().
515 		 */
516 		if (strncmp(path, DEV_BYID_PATH, strlen(DEV_BYID_PATH)) == 0) {
517 			(void) snprintf(pathbuf, sizeof (pathbuf), "%s%s",
518 			    DEV_BYID_PATH, new_devid);
519 			zed_log_msg(LOG_INFO, "  zpool_label_disk: path '%s' "
520 			    "replaced by '%s'", path, pathbuf);
521 			path = pathbuf;
522 		}
523 	}
524 
525 	libzfs_free_str_array(lines, lines_cnt);
526 
527 	/*
528 	 * Construct the root vdev to pass to zpool_vdev_attach().  While adding
529 	 * the entire vdev structure is harmless, we construct a reduced set of
530 	 * path/physpath/wholedisk to keep it simple.
531 	 */
532 	if (nvlist_alloc(&nvroot, NV_UNIQUE_NAME, 0) != 0) {
533 		zed_log_msg(LOG_WARNING, "zfs_mod: nvlist_alloc out of memory");
534 		return;
535 	}
536 	if (nvlist_alloc(&newvd, NV_UNIQUE_NAME, 0) != 0) {
537 		zed_log_msg(LOG_WARNING, "zfs_mod: nvlist_alloc out of memory");
538 		nvlist_free(nvroot);
539 		return;
540 	}
541 
542 	if (nvlist_add_string(newvd, ZPOOL_CONFIG_TYPE, VDEV_TYPE_DISK) != 0 ||
543 	    nvlist_add_string(newvd, ZPOOL_CONFIG_PATH, path) != 0 ||
544 	    nvlist_add_string(newvd, ZPOOL_CONFIG_DEVID, new_devid) != 0 ||
545 	    (physpath != NULL && nvlist_add_string(newvd,
546 	    ZPOOL_CONFIG_PHYS_PATH, physpath) != 0) ||
547 	    (enc_sysfs_path != NULL && nvlist_add_string(newvd,
548 	    ZPOOL_CONFIG_VDEV_ENC_SYSFS_PATH, enc_sysfs_path) != 0) ||
549 	    nvlist_add_uint64(newvd, ZPOOL_CONFIG_WHOLE_DISK, wholedisk) != 0 ||
550 	    nvlist_add_string(nvroot, ZPOOL_CONFIG_TYPE, VDEV_TYPE_ROOT) != 0 ||
551 	    nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
552 	    (const nvlist_t **)&newvd, 1) != 0) {
553 		zed_log_msg(LOG_WARNING, "zfs_mod: unable to add nvlist pairs");
554 		nvlist_free(newvd);
555 		nvlist_free(nvroot);
556 		return;
557 	}
558 
559 	nvlist_free(newvd);
560 
561 	/*
562 	 * Wait for udev to verify the links exist, then auto-replace
563 	 * the leaf disk at same physical location.
564 	 */
565 	if (zpool_label_disk_wait(path, DISK_LABEL_WAIT) != 0) {
566 		zed_log_msg(LOG_WARNING, "zfs_mod: pool '%s', after labeling "
567 		    "replacement disk, the expected disk partition link '%s' "
568 		    "is missing after waiting %u ms",
569 		    zpool_get_name(zhp), path, DISK_LABEL_WAIT);
570 		nvlist_free(nvroot);
571 		return;
572 	}
573 
574 	/*
575 	 * Prefer sequential resilvering when supported (mirrors and dRAID),
576 	 * otherwise fallback to a traditional healing resilver.
577 	 */
578 	ret = zpool_vdev_attach(zhp, fullpath, path, nvroot, B_TRUE, B_TRUE);
579 	if (ret != 0) {
580 		ret = zpool_vdev_attach(zhp, fullpath, path, nvroot,
581 		    B_TRUE, B_FALSE);
582 	}
583 
584 	zed_log_msg(LOG_WARNING, "  zpool_vdev_replace: %s with %s (%s)",
585 	    fullpath, path, (ret == 0) ? "no errors" :
586 	    libzfs_error_description(g_zfshdl));
587 
588 	nvlist_free(nvroot);
589 }
590 
591 /*
592  * Utility functions to find a vdev matching given criteria.
593  */
594 typedef struct dev_data {
595 	const char		*dd_compare;
596 	const char		*dd_prop;
597 	zfs_process_func_t	dd_func;
598 	boolean_t		dd_found;
599 	boolean_t		dd_islabeled;
600 	uint64_t		dd_pool_guid;
601 	uint64_t		dd_vdev_guid;
602 	uint64_t		dd_new_vdev_guid;
603 	const char		*dd_new_devid;
604 	uint64_t		dd_num_spares;
605 } dev_data_t;
606 
607 static void
608 zfs_iter_vdev(zpool_handle_t *zhp, nvlist_t *nvl, void *data)
609 {
610 	dev_data_t *dp = data;
611 	const char *path = NULL;
612 	uint_t c, children;
613 	nvlist_t **child;
614 	uint64_t guid = 0;
615 	uint64_t isspare = 0;
616 
617 	/*
618 	 * First iterate over any children.
619 	 */
620 	if (nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_CHILDREN,
621 	    &child, &children) == 0) {
622 		for (c = 0; c < children; c++)
623 			zfs_iter_vdev(zhp, child[c], data);
624 	}
625 
626 	/*
627 	 * Iterate over any spares and cache devices
628 	 */
629 	if (nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_SPARES,
630 	    &child, &children) == 0) {
631 		for (c = 0; c < children; c++)
632 			zfs_iter_vdev(zhp, child[c], data);
633 	}
634 	if (nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_L2CACHE,
635 	    &child, &children) == 0) {
636 		for (c = 0; c < children; c++)
637 			zfs_iter_vdev(zhp, child[c], data);
638 	}
639 
640 	/* once a vdev was matched and processed there is nothing left to do */
641 	if (dp->dd_found && dp->dd_num_spares == 0)
642 		return;
643 	(void) nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_GUID, &guid);
644 
645 	/*
646 	 * Match by GUID if available otherwise fallback to devid or physical
647 	 */
648 	if (dp->dd_vdev_guid != 0) {
649 		if (guid != dp->dd_vdev_guid)
650 			return;
651 		zed_log_msg(LOG_INFO, "  zfs_iter_vdev: matched on %llu", guid);
652 		dp->dd_found = B_TRUE;
653 
654 	} else if (dp->dd_compare != NULL) {
655 		/*
656 		 * NOTE: On Linux there is an event for partition, so unlike
657 		 * illumos, substring matching is not required to accommodate
658 		 * the partition suffix. An exact match will be present in
659 		 * the dp->dd_compare value.
660 		 * If the attached disk already contains a vdev GUID, it means
661 		 * the disk is not clean. In such a scenario, the physical path
662 		 * would be a match that makes the disk faulted when trying to
663 		 * online it. So, we would only want to proceed if either GUID
664 		 * matches with the last attached disk or the disk is in clean
665 		 * state.
666 		 */
667 		if (nvlist_lookup_string(nvl, dp->dd_prop, &path) != 0 ||
668 		    strcmp(dp->dd_compare, path) != 0) {
669 			return;
670 		}
671 		if (dp->dd_new_vdev_guid != 0 && dp->dd_new_vdev_guid != guid) {
672 			zed_log_msg(LOG_INFO, "  %s: no match (GUID:%llu"
673 			    " != vdev GUID:%llu)", __func__,
674 			    dp->dd_new_vdev_guid, guid);
675 			return;
676 		}
677 
678 		zed_log_msg(LOG_INFO, "  zfs_iter_vdev: matched %s on %s",
679 		    dp->dd_prop, path);
680 		dp->dd_found = B_TRUE;
681 
682 		/* pass the new devid for use by auto-replacing code */
683 		if (dp->dd_new_devid != NULL) {
684 			(void) nvlist_add_string(nvl, "new_devid",
685 			    dp->dd_new_devid);
686 		}
687 	}
688 
689 	if (dp->dd_found == B_TRUE && nvlist_lookup_uint64(nvl,
690 	    ZPOOL_CONFIG_IS_SPARE, &isspare) == 0 && isspare)
691 		dp->dd_num_spares++;
692 
693 	(dp->dd_func)(zhp, nvl, dp->dd_islabeled);
694 }
695 
696 static void
697 zfs_enable_ds(void *arg)
698 {
699 	unavailpool_t *pool = (unavailpool_t *)arg;
700 
701 	(void) zpool_enable_datasets(pool->uap_zhp, NULL, 0);
702 	zpool_close(pool->uap_zhp);
703 	free(pool);
704 }
705 
706 static int
707 zfs_iter_pool(zpool_handle_t *zhp, void *data)
708 {
709 	nvlist_t *config, *nvl;
710 	dev_data_t *dp = data;
711 	uint64_t pool_guid;
712 	unavailpool_t *pool;
713 
714 	zed_log_msg(LOG_INFO, "zfs_iter_pool: evaluating vdevs on %s (by %s)",
715 	    zpool_get_name(zhp), dp->dd_vdev_guid ? "GUID" : dp->dd_prop);
716 
717 	/*
718 	 * For each vdev in this pool, look for a match to apply dd_func
719 	 */
720 	if ((config = zpool_get_config(zhp, NULL)) != NULL) {
721 		if (dp->dd_pool_guid == 0 ||
722 		    (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID,
723 		    &pool_guid) == 0 && pool_guid == dp->dd_pool_guid)) {
724 			(void) nvlist_lookup_nvlist(config,
725 			    ZPOOL_CONFIG_VDEV_TREE, &nvl);
726 			zfs_iter_vdev(zhp, nvl, data);
727 		}
728 	} else {
729 		zed_log_msg(LOG_INFO, "%s: no config\n", __func__);
730 	}
731 
732 	/*
733 	 * if this pool was originally unavailable,
734 	 * then enable its datasets asynchronously
735 	 */
736 	if (g_enumeration_done)  {
737 		for (pool = list_head(&g_pool_list); pool != NULL;
738 		    pool = list_next(&g_pool_list, pool)) {
739 
740 			if (strcmp(zpool_get_name(zhp),
741 			    zpool_get_name(pool->uap_zhp)))
742 				continue;
743 			if (zfs_toplevel_state(zhp) >= VDEV_STATE_DEGRADED) {
744 				list_remove(&g_pool_list, pool);
745 				(void) tpool_dispatch(g_tpool, zfs_enable_ds,
746 				    pool);
747 				break;
748 			}
749 		}
750 	}
751 
752 	zpool_close(zhp);
753 
754 	/* cease iteration after a match */
755 	return (dp->dd_found && dp->dd_num_spares == 0);
756 }
757 
758 /*
759  * Given a physical device location, iterate over all
760  * (pool, vdev) pairs which correspond to that location.
761  */
762 static boolean_t
763 devphys_iter(const char *physical, const char *devid, zfs_process_func_t func,
764     boolean_t is_slice, uint64_t new_vdev_guid)
765 {
766 	dev_data_t data = { 0 };
767 
768 	data.dd_compare = physical;
769 	data.dd_func = func;
770 	data.dd_prop = ZPOOL_CONFIG_PHYS_PATH;
771 	data.dd_found = B_FALSE;
772 	data.dd_islabeled = is_slice;
773 	data.dd_new_devid = devid;	/* used by auto replace code */
774 	data.dd_new_vdev_guid = new_vdev_guid;
775 
776 	(void) zpool_iter(g_zfshdl, zfs_iter_pool, &data);
777 
778 	return (data.dd_found);
779 }
780 
781 /*
782  * Given a device identifier, find any vdevs with a matching by-vdev
783  * path.  Normally we shouldn't need this as the comparison would be
784  * made earlier in the devphys_iter().  For example, if we were replacing
785  * /dev/disk/by-vdev/L28, normally devphys_iter() would match the
786  * ZPOOL_CONFIG_PHYS_PATH of "L28" from the old disk config to "L28"
787  * of the new disk config.  However, we've seen cases where
788  * ZPOOL_CONFIG_PHYS_PATH was not in the config for the old disk.  Here's
789  * an example of a real 2-disk mirror pool where one disk was force
790  * faulted:
791  *
792  *       com.delphix:vdev_zap_top: 129
793  *           children[0]:
794  *               type: 'disk'
795  *               id: 0
796  *               guid: 14309659774640089719
797  *               path: '/dev/disk/by-vdev/L28'
798  *               whole_disk: 0
799  *               DTL: 654
800  *               create_txg: 4
801  *               com.delphix:vdev_zap_leaf: 1161
802  *               faulted: 1
803  *               aux_state: 'external'
804  *           children[1]:
805  *               type: 'disk'
806  *               id: 1
807  *               guid: 16002508084177980912
808  *               path: '/dev/disk/by-vdev/L29'
809  *               devid: 'dm-uuid-mpath-35000c500a61d68a3'
810  *               phys_path: 'L29'
811  *               vdev_enc_sysfs_path: '/sys/class/enclosure/0:0:1:0/SLOT 30 32'
812  *               whole_disk: 0
813  *               DTL: 1028
814  *               create_txg: 4
815  *               com.delphix:vdev_zap_leaf: 131
816  *
817  * So in the case above, the only thing we could compare is the path.
818  *
819  * We can do this because we assume by-vdev paths are authoritative as physical
820  * paths.  We could not assume this for normal paths like /dev/sda since the
821  * physical location /dev/sda points to could change over time.
822  */
823 static boolean_t
824 by_vdev_path_iter(const char *by_vdev_path, const char *devid,
825     zfs_process_func_t func, boolean_t is_slice)
826 {
827 	dev_data_t data = { 0 };
828 
829 	data.dd_compare = by_vdev_path;
830 	data.dd_func = func;
831 	data.dd_prop = ZPOOL_CONFIG_PATH;
832 	data.dd_found = B_FALSE;
833 	data.dd_islabeled = is_slice;
834 	data.dd_new_devid = devid;
835 
836 	if (strncmp(by_vdev_path, DEV_BYVDEV_PATH,
837 	    strlen(DEV_BYVDEV_PATH)) != 0) {
838 		/* by_vdev_path doesn't start with "/dev/disk/by-vdev/" */
839 		return (B_FALSE);
840 	}
841 
842 	(void) zpool_iter(g_zfshdl, zfs_iter_pool, &data);
843 
844 	return (data.dd_found);
845 }
846 
847 /*
848  * Given a device identifier, find any vdevs with a matching devid.
849  * On Linux we can match devid directly which is always a whole disk.
850  */
851 static boolean_t
852 devid_iter(const char *devid, zfs_process_func_t func, boolean_t is_slice)
853 {
854 	dev_data_t data = { 0 };
855 
856 	data.dd_compare = devid;
857 	data.dd_func = func;
858 	data.dd_prop = ZPOOL_CONFIG_DEVID;
859 	data.dd_found = B_FALSE;
860 	data.dd_islabeled = is_slice;
861 	data.dd_new_devid = devid;
862 
863 	(void) zpool_iter(g_zfshdl, zfs_iter_pool, &data);
864 
865 	return (data.dd_found);
866 }
867 
868 /*
869  * Given a device guid, find any vdevs with a matching guid.
870  */
871 static boolean_t
872 guid_iter(uint64_t pool_guid, uint64_t vdev_guid, const char *devid,
873     zfs_process_func_t func, boolean_t is_slice)
874 {
875 	dev_data_t data = { 0 };
876 
877 	data.dd_func = func;
878 	data.dd_found = B_FALSE;
879 	data.dd_pool_guid = pool_guid;
880 	data.dd_vdev_guid = vdev_guid;
881 	data.dd_islabeled = is_slice;
882 	data.dd_new_devid = devid;
883 
884 	(void) zpool_iter(g_zfshdl, zfs_iter_pool, &data);
885 
886 	return (data.dd_found);
887 }
888 
889 /*
890  * Handle a EC_DEV_ADD.ESC_DISK event.
891  *
892  * illumos
893  *	Expects: DEV_PHYS_PATH string in schema
894  *	Matches: vdev's ZPOOL_CONFIG_PHYS_PATH or ZPOOL_CONFIG_DEVID
895  *
896  *      path: '/dev/dsk/c0t1d0s0' (persistent)
897  *     devid: 'id1,sd@SATA_____Hitachi_HDS72101______JP2940HZ3H74MC/a'
898  * phys_path: '/pci@0,0/pci103c,1609@11/disk@1,0:a'
899  *
900  * linux
901  *	provides: DEV_PHYS_PATH and DEV_IDENTIFIER strings in schema
902  *	Matches: vdev's ZPOOL_CONFIG_PHYS_PATH or ZPOOL_CONFIG_DEVID
903  *
904  *      path: '/dev/sdc1' (not persistent)
905  *     devid: 'ata-SAMSUNG_HD204UI_S2HGJD2Z805891-part1'
906  * phys_path: 'pci-0000:04:00.0-sas-0x4433221106000000-lun-0'
907  */
908 static int
909 zfs_deliver_add(nvlist_t *nvl)
910 {
911 	const char *devpath = NULL, *devid = NULL;
912 	uint64_t pool_guid = 0, vdev_guid = 0;
913 	boolean_t is_slice;
914 
915 	/*
916 	 * Expecting a devid string and an optional physical location and guid
917 	 */
918 	if (nvlist_lookup_string(nvl, DEV_IDENTIFIER, &devid) != 0) {
919 		zed_log_msg(LOG_INFO, "%s: no dev identifier\n", __func__);
920 		return (-1);
921 	}
922 
923 	(void) nvlist_lookup_string(nvl, DEV_PHYS_PATH, &devpath);
924 	(void) nvlist_lookup_uint64(nvl, ZFS_EV_POOL_GUID, &pool_guid);
925 	(void) nvlist_lookup_uint64(nvl, ZFS_EV_VDEV_GUID, &vdev_guid);
926 
927 	is_slice = (nvlist_lookup_boolean(nvl, DEV_IS_PART) == 0);
928 
929 	zed_log_msg(LOG_INFO, "zfs_deliver_add: adding %s (%s) (is_slice %d)",
930 	    devid, devpath ? devpath : "NULL", is_slice);
931 
932 	/*
933 	 * Iterate over all vdevs looking for a match in the following order:
934 	 * 1. ZPOOL_CONFIG_DEVID (identifies the unique disk)
935 	 * 2. ZPOOL_CONFIG_PHYS_PATH (identifies disk physical location).
936 	 * 3. ZPOOL_CONFIG_GUID (identifies unique vdev).
937 	 * 4. ZPOOL_CONFIG_PATH for /dev/disk/by-vdev devices only (since
938 	 *    by-vdev paths represent physical paths).
939 	 */
940 	if (devid_iter(devid, zfs_process_add, is_slice))
941 		return (0);
942 	if (devpath != NULL && devphys_iter(devpath, devid, zfs_process_add,
943 	    is_slice, vdev_guid))
944 		return (0);
945 	if (vdev_guid != 0)
946 		(void) guid_iter(pool_guid, vdev_guid, devid, zfs_process_add,
947 		    is_slice);
948 
949 	if (devpath != NULL) {
950 		/* Can we match a /dev/disk/by-vdev/ path? */
951 		char by_vdev_path[MAXPATHLEN];
952 		snprintf(by_vdev_path, sizeof (by_vdev_path),
953 		    "/dev/disk/by-vdev/%s", devpath);
954 		if (by_vdev_path_iter(by_vdev_path, devid, zfs_process_add,
955 		    is_slice))
956 			return (0);
957 	}
958 
959 	return (0);
960 }
961 
962 /*
963  * Called when we receive a VDEV_CHECK event, which indicates a device could not
964  * be opened during initial pool open, but the autoreplace property was set on
965  * the pool.  In this case, we treat it as if it were an add event.
966  */
967 static int
968 zfs_deliver_check(nvlist_t *nvl)
969 {
970 	dev_data_t data = { 0 };
971 
972 	if (nvlist_lookup_uint64(nvl, ZFS_EV_POOL_GUID,
973 	    &data.dd_pool_guid) != 0 ||
974 	    nvlist_lookup_uint64(nvl, ZFS_EV_VDEV_GUID,
975 	    &data.dd_vdev_guid) != 0 ||
976 	    data.dd_vdev_guid == 0)
977 		return (0);
978 
979 	zed_log_msg(LOG_INFO, "zfs_deliver_check: pool '%llu', vdev %llu",
980 	    data.dd_pool_guid, data.dd_vdev_guid);
981 
982 	data.dd_func = zfs_process_add;
983 
984 	(void) zpool_iter(g_zfshdl, zfs_iter_pool, &data);
985 
986 	return (0);
987 }
988 
989 /*
990  * Given a path to a vdev, lookup the vdev's physical size from its
991  * config nvlist.
992  *
993  * Returns the vdev's physical size in bytes on success, 0 on error.
994  */
995 static uint64_t
996 vdev_size_from_config(zpool_handle_t *zhp, const char *vdev_path)
997 {
998 	nvlist_t *nvl = NULL;
999 	boolean_t avail_spare, l2cache, log;
1000 	vdev_stat_t *vs = NULL;
1001 	uint_t c;
1002 
1003 	nvl = zpool_find_vdev(zhp, vdev_path, &avail_spare, &l2cache, &log);
1004 	if (!nvl)
1005 		return (0);
1006 
1007 	verify(nvlist_lookup_uint64_array(nvl, ZPOOL_CONFIG_VDEV_STATS,
1008 	    (uint64_t **)&vs, &c) == 0);
1009 	if (!vs) {
1010 		zed_log_msg(LOG_INFO, "%s: no nvlist for '%s'", __func__,
1011 		    vdev_path);
1012 		return (0);
1013 	}
1014 
1015 	return (vs->vs_pspace);
1016 }
1017 
1018 /*
1019  * Given a path to a vdev, lookup if the vdev is a "whole disk" in the
1020  * config nvlist.  "whole disk" means that ZFS was passed a whole disk
1021  * at pool creation time, which it partitioned up and has full control over.
1022  * Thus a partition with wholedisk=1 set tells us that zfs created the
1023  * partition at creation time.  A partition without whole disk set would have
1024  * been created by externally (like with fdisk) and passed to ZFS.
1025  *
1026  * Returns the whole disk value (either 0 or 1).
1027  */
1028 static uint64_t
1029 vdev_whole_disk_from_config(zpool_handle_t *zhp, const char *vdev_path)
1030 {
1031 	nvlist_t *nvl = NULL;
1032 	boolean_t avail_spare, l2cache, log;
1033 	uint64_t wholedisk = 0;
1034 
1035 	nvl = zpool_find_vdev(zhp, vdev_path, &avail_spare, &l2cache, &log);
1036 	if (!nvl)
1037 		return (0);
1038 
1039 	(void) nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_WHOLE_DISK, &wholedisk);
1040 
1041 	return (wholedisk);
1042 }
1043 
1044 /*
1045  * If the device size grew more than 1% then return true.
1046  */
1047 #define	DEVICE_GREW(oldsize, newsize) \
1048 		    ((newsize > oldsize) && \
1049 		    ((newsize / (newsize - oldsize)) <= 100))
1050 
1051 static int
1052 zfsdle_vdev_online(zpool_handle_t *zhp, void *data)
1053 {
1054 	boolean_t avail_spare, l2cache;
1055 	nvlist_t *udev_nvl = data;
1056 	nvlist_t *tgt;
1057 	int error;
1058 
1059 	const char *tmp_devname;
1060 	char devname[MAXPATHLEN] = "";
1061 	uint64_t guid;
1062 
1063 	if (nvlist_lookup_uint64(udev_nvl, ZFS_EV_VDEV_GUID, &guid) == 0) {
1064 		sprintf(devname, "%llu", (u_longlong_t)guid);
1065 	} else if (nvlist_lookup_string(udev_nvl, DEV_PHYS_PATH,
1066 	    &tmp_devname) == 0) {
1067 		strlcpy(devname, tmp_devname, MAXPATHLEN);
1068 		zfs_append_partition(devname, MAXPATHLEN);
1069 	} else {
1070 		zed_log_msg(LOG_INFO, "%s: no guid or physpath", __func__);
1071 	}
1072 
1073 	zed_log_msg(LOG_INFO, "zfsdle_vdev_online: searching for '%s' in '%s'",
1074 	    devname, zpool_get_name(zhp));
1075 
1076 	if ((tgt = zpool_find_vdev_by_physpath(zhp, devname,
1077 	    &avail_spare, &l2cache, NULL)) != NULL) {
1078 		const char *path;
1079 		char fullpath[MAXPATHLEN];
1080 		uint64_t wholedisk = 0;
1081 
1082 		error = nvlist_lookup_string(tgt, ZPOOL_CONFIG_PATH, &path);
1083 		if (error) {
1084 			zpool_close(zhp);
1085 			return (0);
1086 		}
1087 
1088 		(void) nvlist_lookup_uint64(tgt, ZPOOL_CONFIG_WHOLE_DISK,
1089 		    &wholedisk);
1090 
1091 		if (wholedisk) {
1092 			char *tmp;
1093 			path = strrchr(path, '/');
1094 			if (path != NULL) {
1095 				tmp = zfs_strip_partition(path + 1);
1096 				if (tmp == NULL) {
1097 					zpool_close(zhp);
1098 					return (0);
1099 				}
1100 			} else {
1101 				zpool_close(zhp);
1102 				return (0);
1103 			}
1104 
1105 			(void) strlcpy(fullpath, tmp, sizeof (fullpath));
1106 			free(tmp);
1107 
1108 			/*
1109 			 * We need to reopen the pool associated with this
1110 			 * device so that the kernel can update the size of
1111 			 * the expanded device.  When expanding there is no
1112 			 * need to restart the scrub from the beginning.
1113 			 */
1114 			boolean_t scrub_restart = B_FALSE;
1115 			(void) zpool_reopen_one(zhp, &scrub_restart);
1116 		} else {
1117 			(void) strlcpy(fullpath, path, sizeof (fullpath));
1118 		}
1119 
1120 		if (zpool_get_prop_int(zhp, ZPOOL_PROP_AUTOEXPAND, NULL)) {
1121 			vdev_state_t newstate;
1122 
1123 			if (zpool_get_state(zhp) != POOL_STATE_UNAVAIL) {
1124 				/*
1125 				 * If this disk size has not changed, then
1126 				 * there's no need to do an autoexpand.  To
1127 				 * check we look at the disk's size in its
1128 				 * config, and compare it to the disk size
1129 				 * that udev is reporting.
1130 				 */
1131 				uint64_t udev_size = 0, conf_size = 0,
1132 				    wholedisk = 0, udev_parent_size = 0;
1133 
1134 				/*
1135 				 * Get the size of our disk that udev is
1136 				 * reporting.
1137 				 */
1138 				if (nvlist_lookup_uint64(udev_nvl, DEV_SIZE,
1139 				    &udev_size) != 0) {
1140 					udev_size = 0;
1141 				}
1142 
1143 				/*
1144 				 * Get the size of our disk's parent device
1145 				 * from udev (where sda1's parent is sda).
1146 				 */
1147 				if (nvlist_lookup_uint64(udev_nvl,
1148 				    DEV_PARENT_SIZE, &udev_parent_size) != 0) {
1149 					udev_parent_size = 0;
1150 				}
1151 
1152 				conf_size = vdev_size_from_config(zhp,
1153 				    fullpath);
1154 
1155 				wholedisk = vdev_whole_disk_from_config(zhp,
1156 				    fullpath);
1157 
1158 				/*
1159 				 * Only attempt an autoexpand if the vdev size
1160 				 * changed.  There are two different cases
1161 				 * to consider.
1162 				 *
1163 				 * 1. wholedisk=1
1164 				 * If you do a 'zpool create' on a whole disk
1165 				 * (like /dev/sda), then zfs will create
1166 				 * partitions on the disk (like /dev/sda1).  In
1167 				 * that case, wholedisk=1 will be set in the
1168 				 * partition's nvlist config.  So zed will need
1169 				 * to see if your parent device (/dev/sda)
1170 				 * expanded in size, and if so, then attempt
1171 				 * the autoexpand.
1172 				 *
1173 				 * 2. wholedisk=0
1174 				 * If you do a 'zpool create' on an existing
1175 				 * partition, or a device that doesn't allow
1176 				 * partitions, then wholedisk=0, and you will
1177 				 * simply need to check if the device itself
1178 				 * expanded in size.
1179 				 */
1180 				if (DEVICE_GREW(conf_size, udev_size) ||
1181 				    (wholedisk && DEVICE_GREW(conf_size,
1182 				    udev_parent_size))) {
1183 					error = zpool_vdev_online(zhp, fullpath,
1184 					    0, &newstate);
1185 
1186 					zed_log_msg(LOG_INFO,
1187 					    "%s: autoexpanding '%s' from %llu"
1188 					    " to %llu bytes in pool '%s': %d",
1189 					    __func__, fullpath, conf_size,
1190 					    MAX(udev_size, udev_parent_size),
1191 					    zpool_get_name(zhp), error);
1192 				}
1193 			}
1194 		}
1195 		zpool_close(zhp);
1196 		return (1);
1197 	}
1198 	zpool_close(zhp);
1199 	return (0);
1200 }
1201 
1202 /*
1203  * This function handles the ESC_DEV_DLE device change event.  Use the
1204  * provided vdev guid when looking up a disk or partition, when the guid
1205  * is not present assume the entire disk is owned by ZFS and append the
1206  * expected -part1 partition information then lookup by physical path.
1207  */
1208 static int
1209 zfs_deliver_dle(nvlist_t *nvl)
1210 {
1211 	const char *devname;
1212 	char name[MAXPATHLEN];
1213 	uint64_t guid;
1214 
1215 	if (nvlist_lookup_uint64(nvl, ZFS_EV_VDEV_GUID, &guid) == 0) {
1216 		sprintf(name, "%llu", (u_longlong_t)guid);
1217 	} else if (nvlist_lookup_string(nvl, DEV_PHYS_PATH, &devname) == 0) {
1218 		strlcpy(name, devname, MAXPATHLEN);
1219 		zfs_append_partition(name, MAXPATHLEN);
1220 	} else {
1221 		sprintf(name, "unknown");
1222 		zed_log_msg(LOG_INFO, "zfs_deliver_dle: no guid or physpath");
1223 	}
1224 
1225 	if (zpool_iter(g_zfshdl, zfsdle_vdev_online, nvl) != 1) {
1226 		zed_log_msg(LOG_INFO, "zfs_deliver_dle: device '%s' not "
1227 		    "found", name);
1228 		return (1);
1229 	}
1230 
1231 	return (0);
1232 }
1233 
1234 /*
1235  * syseventd daemon module event handler
1236  *
1237  * Handles syseventd daemon zfs device related events:
1238  *
1239  *	EC_DEV_ADD.ESC_DISK
1240  *	EC_DEV_STATUS.ESC_DEV_DLE
1241  *	EC_ZFS.ESC_ZFS_VDEV_CHECK
1242  *
1243  * Note: assumes only one thread active at a time (not thread safe)
1244  */
1245 static int
1246 zfs_slm_deliver_event(const char *class, const char *subclass, nvlist_t *nvl)
1247 {
1248 	int ret;
1249 	boolean_t is_check = B_FALSE, is_dle = B_FALSE;
1250 
1251 	if (strcmp(class, EC_DEV_ADD) == 0) {
1252 		/*
1253 		 * We're mainly interested in disk additions, but we also listen
1254 		 * for new loop devices, to allow for simplified testing.
1255 		 */
1256 		if (strcmp(subclass, ESC_DISK) != 0 &&
1257 		    strcmp(subclass, ESC_LOFI) != 0)
1258 			return (0);
1259 
1260 		is_check = B_FALSE;
1261 	} else if (strcmp(class, EC_ZFS) == 0 &&
1262 	    strcmp(subclass, ESC_ZFS_VDEV_CHECK) == 0) {
1263 		/*
1264 		 * This event signifies that a device failed to open
1265 		 * during pool load, but the 'autoreplace' property was
1266 		 * set, so we should pretend it's just been added.
1267 		 */
1268 		is_check = B_TRUE;
1269 	} else if (strcmp(class, EC_DEV_STATUS) == 0 &&
1270 	    strcmp(subclass, ESC_DEV_DLE) == 0) {
1271 		is_dle = B_TRUE;
1272 	} else {
1273 		return (0);
1274 	}
1275 
1276 	if (is_dle)
1277 		ret = zfs_deliver_dle(nvl);
1278 	else if (is_check)
1279 		ret = zfs_deliver_check(nvl);
1280 	else
1281 		ret = zfs_deliver_add(nvl);
1282 
1283 	return (ret);
1284 }
1285 
1286 static void *
1287 zfs_enum_pools(void *arg)
1288 {
1289 	(void) arg;
1290 
1291 	(void) zpool_iter(g_zfshdl, zfs_unavail_pool, (void *)&g_pool_list);
1292 	/*
1293 	 * Linux - instead of using a thread pool, each list entry
1294 	 * will spawn a thread when an unavailable pool transitions
1295 	 * to available. zfs_slm_fini will wait for these threads.
1296 	 */
1297 	g_enumeration_done = B_TRUE;
1298 	return (NULL);
1299 }
1300 
1301 /*
1302  * called from zed daemon at startup
1303  *
1304  * sent messages from zevents or udev monitor
1305  *
1306  * For now, each agent has its own libzfs instance
1307  */
1308 int
1309 zfs_slm_init(void)
1310 {
1311 	if ((g_zfshdl = libzfs_init()) == NULL)
1312 		return (-1);
1313 
1314 	/*
1315 	 * collect a list of unavailable pools (asynchronously,
1316 	 * since this can take a while)
1317 	 */
1318 	list_create(&g_pool_list, sizeof (struct unavailpool),
1319 	    offsetof(struct unavailpool, uap_node));
1320 
1321 	if (pthread_create(&g_zfs_tid, NULL, zfs_enum_pools, NULL) != 0) {
1322 		list_destroy(&g_pool_list);
1323 		libzfs_fini(g_zfshdl);
1324 		return (-1);
1325 	}
1326 
1327 	pthread_setname_np(g_zfs_tid, "enum-pools");
1328 	list_create(&g_device_list, sizeof (struct pendingdev),
1329 	    offsetof(struct pendingdev, pd_node));
1330 
1331 	return (0);
1332 }
1333 
1334 void
1335 zfs_slm_fini(void)
1336 {
1337 	unavailpool_t *pool;
1338 	pendingdev_t *device;
1339 
1340 	/* wait for zfs_enum_pools thread to complete */
1341 	(void) pthread_join(g_zfs_tid, NULL);
1342 	/* destroy the thread pool */
1343 	if (g_tpool != NULL) {
1344 		tpool_wait(g_tpool);
1345 		tpool_destroy(g_tpool);
1346 	}
1347 
1348 	while ((pool = list_remove_head(&g_pool_list)) != NULL) {
1349 		zpool_close(pool->uap_zhp);
1350 		free(pool);
1351 	}
1352 	list_destroy(&g_pool_list);
1353 
1354 	while ((device = list_remove_head(&g_device_list)) != NULL)
1355 		free(device);
1356 	list_destroy(&g_device_list);
1357 
1358 	libzfs_fini(g_zfshdl);
1359 }
1360 
1361 void
1362 zfs_slm_event(const char *class, const char *subclass, nvlist_t *nvl)
1363 {
1364 	zed_log_msg(LOG_INFO, "zfs_slm_event: %s.%s", class, subclass);
1365 	(void) zfs_slm_deliver_event(class, subclass, nvl);
1366 }
1367