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/*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
25 * Copyright (c) 2017, Intel Corporation.
26 * Copyright 2019 Joyent, Inc.
27 */
28
29/*
30 * Virtual Device Labels
31 * ---------------------
32 *
33 * The vdev label serves several distinct purposes:
34 *
35 *	1. Uniquely identify this device as part of a ZFS pool and confirm its
36 *	   identity within the pool.
37 *
38 *	2. Verify that all the devices given in a configuration are present
39 *         within the pool.
40 *
41 *	3. Determine the uberblock for the pool.
42 *
43 *	4. In case of an import operation, determine the configuration of the
44 *         toplevel vdev of which it is a part.
45 *
46 *	5. If an import operation cannot find all the devices in the pool,
47 *         provide enough information to the administrator to determine which
48 *         devices are missing.
49 *
50 * It is important to note that while the kernel is responsible for writing the
51 * label, it only consumes the information in the first three cases.  The
52 * latter information is only consumed in userland when determining the
53 * configuration to import a pool.
54 *
55 *
56 * Label Organization
57 * ------------------
58 *
59 * Before describing the contents of the label, it's important to understand how
60 * the labels are written and updated with respect to the uberblock.
61 *
62 * When the pool configuration is altered, either because it was newly created
63 * or a device was added, we want to update all the labels such that we can deal
64 * with fatal failure at any point.  To this end, each disk has two labels which
65 * are updated before and after the uberblock is synced.  Assuming we have
66 * labels and an uberblock with the following transaction groups:
67 *
68 *              L1          UB          L2
69 *           +------+    +------+    +------+
70 *           |      |    |      |    |      |
71 *           | t10  |    | t10  |    | t10  |
72 *           |      |    |      |    |      |
73 *           +------+    +------+    +------+
74 *
75 * In this stable state, the labels and the uberblock were all updated within
76 * the same transaction group (10).  Each label is mirrored and checksummed, so
77 * that we can detect when we fail partway through writing the label.
78 *
79 * In order to identify which labels are valid, the labels are written in the
80 * following manner:
81 *
82 *	1. For each vdev, update 'L1' to the new label
83 *	2. Update the uberblock
84 *	3. For each vdev, update 'L2' to the new label
85 *
86 * Given arbitrary failure, we can determine the correct label to use based on
87 * the transaction group.  If we fail after updating L1 but before updating the
88 * UB, we will notice that L1's transaction group is greater than the uberblock,
89 * so L2 must be valid.  If we fail after writing the uberblock but before
90 * writing L2, we will notice that L2's transaction group is less than L1, and
91 * therefore L1 is valid.
92 *
93 * Another added complexity is that not every label is updated when the config
94 * is synced.  If we add a single device, we do not want to have to re-write
95 * every label for every device in the pool.  This means that both L1 and L2 may
96 * be older than the pool uberblock, because the necessary information is stored
97 * on another vdev.
98 *
99 *
100 * On-disk Format
101 * --------------
102 *
103 * The vdev label consists of two distinct parts, and is wrapped within the
104 * vdev_label_t structure.  The label includes 8k of padding to permit legacy
105 * VTOC disk labels, but is otherwise ignored.
106 *
107 * The first half of the label is a packed nvlist which contains pool wide
108 * properties, per-vdev properties, and configuration information.  It is
109 * described in more detail below.
110 *
111 * The latter half of the label consists of a redundant array of uberblocks.
112 * These uberblocks are updated whenever a transaction group is committed,
113 * or when the configuration is updated.  When a pool is loaded, we scan each
114 * vdev for the 'best' uberblock.
115 *
116 *
117 * Configuration Information
118 * -------------------------
119 *
120 * The nvlist describing the pool and vdev contains the following elements:
121 *
122 *	version		ZFS on-disk version
123 *	name		Pool name
124 *	state		Pool state
125 *	txg		Transaction group in which this label was written
126 *	pool_guid	Unique identifier for this pool
127 *	vdev_tree	An nvlist describing vdev tree.
128 *	features_for_read
129 *			An nvlist of the features necessary for reading the MOS.
130 *
131 * Each leaf device label also contains the following:
132 *
133 *	top_guid	Unique ID for top-level vdev in which this is contained
134 *	guid		Unique ID for the leaf vdev
135 *
136 * The 'vs' configuration follows the format described in 'spa_config.c'.
137 */
138
139#include <sys/zfs_context.h>
140#include <sys/spa.h>
141#include <sys/spa_impl.h>
142#include <sys/dmu.h>
143#include <sys/zap.h>
144#include <sys/vdev.h>
145#include <sys/vdev_impl.h>
146#include <sys/uberblock_impl.h>
147#include <sys/metaslab.h>
148#include <sys/metaslab_impl.h>
149#include <sys/zio.h>
150#include <sys/dsl_scan.h>
151#include <sys/abd.h>
152#include <sys/fs/zfs.h>
153
154/*
155 * Basic routines to read and write from a vdev label.
156 * Used throughout the rest of this file.
157 */
158uint64_t
159vdev_label_offset(uint64_t psize, int l, uint64_t offset)
160{
161	ASSERT(offset < sizeof (vdev_label_t));
162	ASSERT(P2PHASE_TYPED(psize, sizeof (vdev_label_t), uint64_t) == 0);
163
164	return (offset + l * sizeof (vdev_label_t) + (l < VDEV_LABELS / 2 ?
165	    0 : psize - VDEV_LABELS * sizeof (vdev_label_t)));
166}
167
168/*
169 * Returns back the vdev label associated with the passed in offset.
170 */
171int
172vdev_label_number(uint64_t psize, uint64_t offset)
173{
174	int l;
175
176	if (offset >= psize - VDEV_LABEL_END_SIZE) {
177		offset -= psize - VDEV_LABEL_END_SIZE;
178		offset += (VDEV_LABELS / 2) * sizeof (vdev_label_t);
179	}
180	l = offset / sizeof (vdev_label_t);
181	return (l < VDEV_LABELS ? l : -1);
182}
183
184static void
185vdev_label_read(zio_t *zio, vdev_t *vd, int l, abd_t *buf, uint64_t offset,
186    uint64_t size, zio_done_func_t *done, void *private, int flags)
187{
188	ASSERT(
189	    spa_config_held(zio->io_spa, SCL_STATE, RW_READER) == SCL_STATE ||
190	    spa_config_held(zio->io_spa, SCL_STATE, RW_WRITER) == SCL_STATE);
191	ASSERT(flags & ZIO_FLAG_CONFIG_WRITER);
192
193	zio_nowait(zio_read_phys(zio, vd,
194	    vdev_label_offset(vd->vdev_psize, l, offset),
195	    size, buf, ZIO_CHECKSUM_LABEL, done, private,
196	    ZIO_PRIORITY_SYNC_READ, flags, B_TRUE));
197}
198
199void
200vdev_label_write(zio_t *zio, vdev_t *vd, int l, abd_t *buf, uint64_t offset,
201    uint64_t size, zio_done_func_t *done, void *private, int flags)
202{
203	ASSERT(
204	    spa_config_held(zio->io_spa, SCL_STATE, RW_READER) == SCL_STATE ||
205	    spa_config_held(zio->io_spa, SCL_STATE, RW_WRITER) == SCL_STATE);
206	ASSERT(flags & ZIO_FLAG_CONFIG_WRITER);
207
208	zio_nowait(zio_write_phys(zio, vd,
209	    vdev_label_offset(vd->vdev_psize, l, offset),
210	    size, buf, ZIO_CHECKSUM_LABEL, done, private,
211	    ZIO_PRIORITY_SYNC_WRITE, flags, B_TRUE));
212}
213
214static void
215root_vdev_actions_getprogress(vdev_t *vd, nvlist_t *nvl)
216{
217	spa_t *spa = vd->vdev_spa;
218
219	if (vd != spa->spa_root_vdev)
220		return;
221
222	/* provide either current or previous scan information */
223	pool_scan_stat_t ps;
224	if (spa_scan_get_stats(spa, &ps) == 0) {
225		fnvlist_add_uint64_array(nvl,
226		    ZPOOL_CONFIG_SCAN_STATS, (uint64_t *)&ps,
227		    sizeof (pool_scan_stat_t) / sizeof (uint64_t));
228	}
229
230	pool_removal_stat_t prs;
231	if (spa_removal_get_stats(spa, &prs) == 0) {
232		fnvlist_add_uint64_array(nvl,
233		    ZPOOL_CONFIG_REMOVAL_STATS, (uint64_t *)&prs,
234		    sizeof (prs) / sizeof (uint64_t));
235	}
236
237	pool_checkpoint_stat_t pcs;
238	if (spa_checkpoint_get_stats(spa, &pcs) == 0) {
239		fnvlist_add_uint64_array(nvl,
240		    ZPOOL_CONFIG_CHECKPOINT_STATS, (uint64_t *)&pcs,
241		    sizeof (pcs) / sizeof (uint64_t));
242	}
243}
244
245/*
246 * Generate the nvlist representing this vdev's config.
247 */
248nvlist_t *
249vdev_config_generate(spa_t *spa, vdev_t *vd, boolean_t getstats,
250    vdev_config_flag_t flags)
251{
252	nvlist_t *nv = NULL;
253	vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
254
255	nv = fnvlist_alloc();
256
257	fnvlist_add_string(nv, ZPOOL_CONFIG_TYPE, vd->vdev_ops->vdev_op_type);
258	if (!(flags & (VDEV_CONFIG_SPARE | VDEV_CONFIG_L2CACHE)))
259		fnvlist_add_uint64(nv, ZPOOL_CONFIG_ID, vd->vdev_id);
260	fnvlist_add_uint64(nv, ZPOOL_CONFIG_GUID, vd->vdev_guid);
261
262	if (vd->vdev_path != NULL)
263		fnvlist_add_string(nv, ZPOOL_CONFIG_PATH, vd->vdev_path);
264
265	if (vd->vdev_devid != NULL)
266		fnvlist_add_string(nv, ZPOOL_CONFIG_DEVID, vd->vdev_devid);
267
268	if (vd->vdev_physpath != NULL)
269		fnvlist_add_string(nv, ZPOOL_CONFIG_PHYS_PATH,
270		    vd->vdev_physpath);
271
272	if (vd->vdev_fru != NULL)
273		fnvlist_add_string(nv, ZPOOL_CONFIG_FRU, vd->vdev_fru);
274
275	if (vd->vdev_nparity != 0) {
276		ASSERT(strcmp(vd->vdev_ops->vdev_op_type,
277		    VDEV_TYPE_RAIDZ) == 0);
278
279		/*
280		 * Make sure someone hasn't managed to sneak a fancy new vdev
281		 * into a crufty old storage pool.
282		 */
283		ASSERT(vd->vdev_nparity == 1 ||
284		    (vd->vdev_nparity <= 2 &&
285		    spa_version(spa) >= SPA_VERSION_RAIDZ2) ||
286		    (vd->vdev_nparity <= 3 &&
287		    spa_version(spa) >= SPA_VERSION_RAIDZ3));
288
289		/*
290		 * Note that we'll add the nparity tag even on storage pools
291		 * that only support a single parity device -- older software
292		 * will just ignore it.
293		 */
294		fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vd->vdev_nparity);
295	}
296
297	if (vd->vdev_wholedisk != -1ULL)
298		fnvlist_add_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
299		    vd->vdev_wholedisk);
300
301	if (vd->vdev_not_present && !(flags & VDEV_CONFIG_MISSING))
302		fnvlist_add_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT, 1);
303
304	if (vd->vdev_isspare)
305		fnvlist_add_uint64(nv, ZPOOL_CONFIG_IS_SPARE, 1);
306
307	if (!(flags & (VDEV_CONFIG_SPARE | VDEV_CONFIG_L2CACHE)) &&
308	    vd == vd->vdev_top) {
309		fnvlist_add_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
310		    vd->vdev_ms_array);
311		fnvlist_add_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
312		    vd->vdev_ms_shift);
313		fnvlist_add_uint64(nv, ZPOOL_CONFIG_ASHIFT, vd->vdev_ashift);
314		fnvlist_add_uint64(nv, ZPOOL_CONFIG_ASIZE,
315		    vd->vdev_asize);
316		fnvlist_add_uint64(nv, ZPOOL_CONFIG_IS_LOG, vd->vdev_islog);
317		if (vd->vdev_removing) {
318			fnvlist_add_uint64(nv, ZPOOL_CONFIG_REMOVING,
319			    vd->vdev_removing);
320		}
321
322		/* zpool command expects alloc class data */
323		if (getstats && vd->vdev_alloc_bias != VDEV_BIAS_NONE) {
324			const char *bias = NULL;
325
326			switch (vd->vdev_alloc_bias) {
327			case VDEV_BIAS_LOG:
328				bias = VDEV_ALLOC_BIAS_LOG;
329				break;
330			case VDEV_BIAS_SPECIAL:
331				bias = VDEV_ALLOC_BIAS_SPECIAL;
332				break;
333			case VDEV_BIAS_DEDUP:
334				bias = VDEV_ALLOC_BIAS_DEDUP;
335				break;
336			default:
337				ASSERT3U(vd->vdev_alloc_bias, ==,
338				    VDEV_BIAS_NONE);
339			}
340			fnvlist_add_string(nv, ZPOOL_CONFIG_ALLOCATION_BIAS,
341			    bias);
342		}
343	}
344
345	if (vd->vdev_dtl_sm != NULL) {
346		fnvlist_add_uint64(nv, ZPOOL_CONFIG_DTL,
347		    space_map_object(vd->vdev_dtl_sm));
348	}
349
350	if (vic->vic_mapping_object != 0) {
351		fnvlist_add_uint64(nv, ZPOOL_CONFIG_INDIRECT_OBJECT,
352		    vic->vic_mapping_object);
353	}
354
355	if (vic->vic_births_object != 0) {
356		fnvlist_add_uint64(nv, ZPOOL_CONFIG_INDIRECT_BIRTHS,
357		    vic->vic_births_object);
358	}
359
360	if (vic->vic_prev_indirect_vdev != UINT64_MAX) {
361		fnvlist_add_uint64(nv, ZPOOL_CONFIG_PREV_INDIRECT_VDEV,
362		    vic->vic_prev_indirect_vdev);
363	}
364
365	if (vd->vdev_crtxg)
366		fnvlist_add_uint64(nv, ZPOOL_CONFIG_CREATE_TXG, vd->vdev_crtxg);
367
368	if (flags & VDEV_CONFIG_MOS) {
369		if (vd->vdev_leaf_zap != 0) {
370			ASSERT(vd->vdev_ops->vdev_op_leaf);
371			fnvlist_add_uint64(nv, ZPOOL_CONFIG_VDEV_LEAF_ZAP,
372			    vd->vdev_leaf_zap);
373		}
374
375		if (vd->vdev_top_zap != 0) {
376			ASSERT(vd == vd->vdev_top);
377			fnvlist_add_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
378			    vd->vdev_top_zap);
379		}
380
381		if (vd->vdev_resilver_deferred) {
382			ASSERT(vd->vdev_ops->vdev_op_leaf);
383			ASSERT(spa->spa_resilver_deferred);
384			fnvlist_add_boolean(nv, ZPOOL_CONFIG_RESILVER_DEFER);
385		}
386	}
387
388	if (getstats) {
389		vdev_stat_t vs;
390
391		vdev_get_stats(vd, &vs);
392		fnvlist_add_uint64_array(nv, ZPOOL_CONFIG_VDEV_STATS,
393		    (uint64_t *)&vs, sizeof (vs) / sizeof (uint64_t));
394
395		root_vdev_actions_getprogress(vd, nv);
396
397		/*
398		 * Note: this can be called from open context
399		 * (spa_get_stats()), so we need the rwlock to prevent
400		 * the mapping from being changed by condensing.
401		 */
402		rw_enter(&vd->vdev_indirect_rwlock, RW_READER);
403		if (vd->vdev_indirect_mapping != NULL) {
404			ASSERT(vd->vdev_indirect_births != NULL);
405			vdev_indirect_mapping_t *vim =
406			    vd->vdev_indirect_mapping;
407			fnvlist_add_uint64(nv, ZPOOL_CONFIG_INDIRECT_SIZE,
408			    vdev_indirect_mapping_size(vim));
409		}
410		rw_exit(&vd->vdev_indirect_rwlock);
411		if (vd->vdev_mg != NULL &&
412		    vd->vdev_mg->mg_fragmentation != ZFS_FRAG_INVALID) {
413			/*
414			 * Compute approximately how much memory would be used
415			 * for the indirect mapping if this device were to
416			 * be removed.
417			 *
418			 * Note: If the frag metric is invalid, then not
419			 * enough metaslabs have been converted to have
420			 * histograms.
421			 */
422			uint64_t seg_count = 0;
423			uint64_t to_alloc = vd->vdev_stat.vs_alloc;
424
425			/*
426			 * There are the same number of allocated segments
427			 * as free segments, so we will have at least one
428			 * entry per free segment.  However, small free
429			 * segments (smaller than vdev_removal_max_span)
430			 * will be combined with adjacent allocated segments
431			 * as a single mapping.
432			 */
433			for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
434				if (1ULL << (i + 1) < vdev_removal_max_span) {
435					to_alloc +=
436					    vd->vdev_mg->mg_histogram[i] <<
437					    i + 1;
438				} else {
439					seg_count +=
440					    vd->vdev_mg->mg_histogram[i];
441				}
442			}
443
444			/*
445			 * The maximum length of a mapping is
446			 * zfs_remove_max_segment, so we need at least one entry
447			 * per zfs_remove_max_segment of allocated data.
448			 */
449			seg_count += to_alloc / zfs_remove_max_segment;
450
451			fnvlist_add_uint64(nv, ZPOOL_CONFIG_INDIRECT_SIZE,
452			    seg_count *
453			    sizeof (vdev_indirect_mapping_entry_phys_t));
454		}
455	}
456
457	if (!vd->vdev_ops->vdev_op_leaf) {
458		nvlist_t **child;
459		int c, idx;
460
461		ASSERT(!vd->vdev_ishole);
462
463		child = kmem_alloc(vd->vdev_children * sizeof (nvlist_t *),
464		    KM_SLEEP);
465
466		for (c = 0, idx = 0; c < vd->vdev_children; c++) {
467			vdev_t *cvd = vd->vdev_child[c];
468
469			/*
470			 * If we're generating an nvlist of removing
471			 * vdevs then skip over any device which is
472			 * not being removed.
473			 */
474			if ((flags & VDEV_CONFIG_REMOVING) &&
475			    !cvd->vdev_removing)
476				continue;
477
478			child[idx++] = vdev_config_generate(spa, cvd,
479			    getstats, flags);
480		}
481
482		if (idx) {
483			fnvlist_add_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
484			    child, idx);
485		}
486
487		for (c = 0; c < idx; c++)
488			nvlist_free(child[c]);
489
490		kmem_free(child, vd->vdev_children * sizeof (nvlist_t *));
491
492	} else {
493		const char *aux = NULL;
494
495		if (vd->vdev_offline && !vd->vdev_tmpoffline)
496			fnvlist_add_uint64(nv, ZPOOL_CONFIG_OFFLINE, B_TRUE);
497		if (vd->vdev_resilver_txg != 0)
498			fnvlist_add_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
499			    vd->vdev_resilver_txg);
500		if (vd->vdev_faulted)
501			fnvlist_add_uint64(nv, ZPOOL_CONFIG_FAULTED, B_TRUE);
502		if (vd->vdev_degraded)
503			fnvlist_add_uint64(nv, ZPOOL_CONFIG_DEGRADED, B_TRUE);
504		if (vd->vdev_removed)
505			fnvlist_add_uint64(nv, ZPOOL_CONFIG_REMOVED, B_TRUE);
506		if (vd->vdev_unspare)
507			fnvlist_add_uint64(nv, ZPOOL_CONFIG_UNSPARE, B_TRUE);
508		if (vd->vdev_ishole)
509			fnvlist_add_uint64(nv, ZPOOL_CONFIG_IS_HOLE, B_TRUE);
510
511		switch (vd->vdev_stat.vs_aux) {
512		case VDEV_AUX_ERR_EXCEEDED:
513			aux = "err_exceeded";
514			break;
515
516		case VDEV_AUX_EXTERNAL:
517			aux = "external";
518			break;
519		}
520
521		if (aux != NULL)
522			fnvlist_add_string(nv, ZPOOL_CONFIG_AUX_STATE, aux);
523
524		if (vd->vdev_splitting && vd->vdev_orig_guid != 0LL) {
525			fnvlist_add_uint64(nv, ZPOOL_CONFIG_ORIG_GUID,
526			    vd->vdev_orig_guid);
527		}
528	}
529
530	return (nv);
531}
532
533/*
534 * Generate a view of the top-level vdevs.  If we currently have holes
535 * in the namespace, then generate an array which contains a list of holey
536 * vdevs.  Additionally, add the number of top-level children that currently
537 * exist.
538 */
539void
540vdev_top_config_generate(spa_t *spa, nvlist_t *config)
541{
542	vdev_t *rvd = spa->spa_root_vdev;
543	uint64_t *array;
544	uint_t c, idx;
545
546	array = kmem_alloc(rvd->vdev_children * sizeof (uint64_t), KM_SLEEP);
547
548	for (c = 0, idx = 0; c < rvd->vdev_children; c++) {
549		vdev_t *tvd = rvd->vdev_child[c];
550
551		if (tvd->vdev_ishole) {
552			array[idx++] = c;
553		}
554	}
555
556	if (idx) {
557		VERIFY(nvlist_add_uint64_array(config, ZPOOL_CONFIG_HOLE_ARRAY,
558		    array, idx) == 0);
559	}
560
561	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_VDEV_CHILDREN,
562	    rvd->vdev_children) == 0);
563
564	kmem_free(array, rvd->vdev_children * sizeof (uint64_t));
565}
566
567/*
568 * Returns the configuration from the label of the given vdev. For vdevs
569 * which don't have a txg value stored on their label (i.e. spares/cache)
570 * or have not been completely initialized (txg = 0) just return
571 * the configuration from the first valid label we find. Otherwise,
572 * find the most up-to-date label that does not exceed the specified
573 * 'txg' value.
574 */
575nvlist_t *
576vdev_label_read_config(vdev_t *vd, uint64_t txg)
577{
578	spa_t *spa = vd->vdev_spa;
579	nvlist_t *config = NULL;
580	vdev_phys_t *vp;
581	abd_t *vp_abd;
582	zio_t *zio;
583	uint64_t best_txg = 0;
584	uint64_t label_txg = 0;
585	int error = 0;
586	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL |
587	    ZIO_FLAG_SPECULATIVE;
588
589	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
590
591	if (!vdev_readable(vd))
592		return (NULL);
593
594	vp_abd = abd_alloc_linear(sizeof (vdev_phys_t), B_TRUE);
595	vp = abd_to_buf(vp_abd);
596
597retry:
598	for (int l = 0; l < VDEV_LABELS; l++) {
599		nvlist_t *label = NULL;
600
601		zio = zio_root(spa, NULL, NULL, flags);
602
603		vdev_label_read(zio, vd, l, vp_abd,
604		    offsetof(vdev_label_t, vl_vdev_phys),
605		    sizeof (vdev_phys_t), NULL, NULL, flags);
606
607		if (zio_wait(zio) == 0 &&
608		    nvlist_unpack(vp->vp_nvlist, sizeof (vp->vp_nvlist),
609		    &label, 0) == 0) {
610			/*
611			 * Auxiliary vdevs won't have txg values in their
612			 * labels and newly added vdevs may not have been
613			 * completely initialized so just return the
614			 * configuration from the first valid label we
615			 * encounter.
616			 */
617			error = nvlist_lookup_uint64(label,
618			    ZPOOL_CONFIG_POOL_TXG, &label_txg);
619			if ((error || label_txg == 0) && !config) {
620				config = label;
621				break;
622			} else if (label_txg <= txg && label_txg > best_txg) {
623				best_txg = label_txg;
624				nvlist_free(config);
625				config = fnvlist_dup(label);
626			}
627		}
628
629		if (label != NULL) {
630			nvlist_free(label);
631			label = NULL;
632		}
633	}
634
635	if (config == NULL && !(flags & ZIO_FLAG_TRYHARD)) {
636		flags |= ZIO_FLAG_TRYHARD;
637		goto retry;
638	}
639
640	/*
641	 * We found a valid label but it didn't pass txg restrictions.
642	 */
643	if (config == NULL && label_txg != 0) {
644		vdev_dbgmsg(vd, "label discarded as txg is too large "
645		    "(%llu > %llu)", (u_longlong_t)label_txg,
646		    (u_longlong_t)txg);
647	}
648
649	abd_free(vp_abd);
650
651	return (config);
652}
653
654/*
655 * Determine if a device is in use.  The 'spare_guid' parameter will be filled
656 * in with the device guid if this spare is active elsewhere on the system.
657 */
658static boolean_t
659vdev_inuse(vdev_t *vd, uint64_t crtxg, vdev_labeltype_t reason,
660    uint64_t *spare_guid, uint64_t *l2cache_guid)
661{
662	spa_t *spa = vd->vdev_spa;
663	uint64_t state, pool_guid, device_guid, txg, spare_pool;
664	uint64_t vdtxg = 0;
665	nvlist_t *label;
666
667	if (spare_guid)
668		*spare_guid = 0ULL;
669	if (l2cache_guid)
670		*l2cache_guid = 0ULL;
671
672	/*
673	 * Read the label, if any, and perform some basic sanity checks.
674	 */
675	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL)
676		return (B_FALSE);
677
678	(void) nvlist_lookup_uint64(label, ZPOOL_CONFIG_CREATE_TXG,
679	    &vdtxg);
680
681	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
682	    &state) != 0 ||
683	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
684	    &device_guid) != 0) {
685		nvlist_free(label);
686		return (B_FALSE);
687	}
688
689	if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
690	    (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
691	    &pool_guid) != 0 ||
692	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_TXG,
693	    &txg) != 0)) {
694		nvlist_free(label);
695		return (B_FALSE);
696	}
697
698	nvlist_free(label);
699
700	/*
701	 * Check to see if this device indeed belongs to the pool it claims to
702	 * be a part of.  The only way this is allowed is if the device is a hot
703	 * spare (which we check for later on).
704	 */
705	if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
706	    !spa_guid_exists(pool_guid, device_guid) &&
707	    !spa_spare_exists(device_guid, NULL, NULL) &&
708	    !spa_l2cache_exists(device_guid, NULL))
709		return (B_FALSE);
710
711	/*
712	 * If the transaction group is zero, then this an initialized (but
713	 * unused) label.  This is only an error if the create transaction
714	 * on-disk is the same as the one we're using now, in which case the
715	 * user has attempted to add the same vdev multiple times in the same
716	 * transaction.
717	 */
718	if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
719	    txg == 0 && vdtxg == crtxg)
720		return (B_TRUE);
721
722	/*
723	 * Check to see if this is a spare device.  We do an explicit check for
724	 * spa_has_spare() here because it may be on our pending list of spares
725	 * to add.  We also check if it is an l2cache device.
726	 */
727	if (spa_spare_exists(device_guid, &spare_pool, NULL) ||
728	    spa_has_spare(spa, device_guid)) {
729		if (spare_guid)
730			*spare_guid = device_guid;
731
732		switch (reason) {
733		case VDEV_LABEL_CREATE:
734		case VDEV_LABEL_L2CACHE:
735			return (B_TRUE);
736
737		case VDEV_LABEL_REPLACE:
738			return (!spa_has_spare(spa, device_guid) ||
739			    spare_pool != 0ULL);
740
741		case VDEV_LABEL_SPARE:
742			return (spa_has_spare(spa, device_guid));
743		}
744	}
745
746	/*
747	 * Check to see if this is an l2cache device.
748	 */
749	if (spa_l2cache_exists(device_guid, NULL))
750		return (B_TRUE);
751
752	/*
753	 * We can't rely on a pool's state if it's been imported
754	 * read-only.  Instead we look to see if the pools is marked
755	 * read-only in the namespace and set the state to active.
756	 */
757	if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
758	    (spa = spa_by_guid(pool_guid, device_guid)) != NULL &&
759	    spa_mode(spa) == FREAD)
760		state = POOL_STATE_ACTIVE;
761
762	/*
763	 * If the device is marked ACTIVE, then this device is in use by another
764	 * pool on the system.
765	 */
766	return (state == POOL_STATE_ACTIVE);
767}
768
769/*
770 * Initialize a vdev label.  We check to make sure each leaf device is not in
771 * use, and writable.  We put down an initial label which we will later
772 * overwrite with a complete label.  Note that it's important to do this
773 * sequentially, not in parallel, so that we catch cases of multiple use of the
774 * same leaf vdev in the vdev we're creating -- e.g. mirroring a disk with
775 * itself.
776 */
777int
778vdev_label_init(vdev_t *vd, uint64_t crtxg, vdev_labeltype_t reason)
779{
780	spa_t *spa = vd->vdev_spa;
781	nvlist_t *label;
782	vdev_phys_t *vp;
783	abd_t *vp_abd;
784	abd_t *pad2;
785	uberblock_t *ub;
786	abd_t *ub_abd;
787	zio_t *zio;
788	char *buf;
789	size_t buflen;
790	int error;
791	uint64_t spare_guid, l2cache_guid;
792	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL;
793
794	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
795
796	for (int c = 0; c < vd->vdev_children; c++)
797		if ((error = vdev_label_init(vd->vdev_child[c],
798		    crtxg, reason)) != 0)
799			return (error);
800
801	/* Track the creation time for this vdev */
802	vd->vdev_crtxg = crtxg;
803
804	if (!vd->vdev_ops->vdev_op_leaf || !spa_writeable(spa))
805		return (0);
806
807	/*
808	 * Dead vdevs cannot be initialized.
809	 */
810	if (vdev_is_dead(vd))
811		return (SET_ERROR(EIO));
812
813	/*
814	 * Determine if the vdev is in use.
815	 */
816	if (reason != VDEV_LABEL_REMOVE && reason != VDEV_LABEL_SPLIT &&
817	    vdev_inuse(vd, crtxg, reason, &spare_guid, &l2cache_guid))
818		return (SET_ERROR(EBUSY));
819
820	/*
821	 * If this is a request to add or replace a spare or l2cache device
822	 * that is in use elsewhere on the system, then we must update the
823	 * guid (which was initialized to a random value) to reflect the
824	 * actual GUID (which is shared between multiple pools).
825	 */
826	if (reason != VDEV_LABEL_REMOVE && reason != VDEV_LABEL_L2CACHE &&
827	    spare_guid != 0ULL) {
828		uint64_t guid_delta = spare_guid - vd->vdev_guid;
829
830		vd->vdev_guid += guid_delta;
831
832		for (vdev_t *pvd = vd; pvd != NULL; pvd = pvd->vdev_parent)
833			pvd->vdev_guid_sum += guid_delta;
834
835		/*
836		 * If this is a replacement, then we want to fallthrough to the
837		 * rest of the code.  If we're adding a spare, then it's already
838		 * labeled appropriately and we can just return.
839		 */
840		if (reason == VDEV_LABEL_SPARE)
841			return (0);
842		ASSERT(reason == VDEV_LABEL_REPLACE ||
843		    reason == VDEV_LABEL_SPLIT);
844	}
845
846	if (reason != VDEV_LABEL_REMOVE && reason != VDEV_LABEL_SPARE &&
847	    l2cache_guid != 0ULL) {
848		uint64_t guid_delta = l2cache_guid - vd->vdev_guid;
849
850		vd->vdev_guid += guid_delta;
851
852		for (vdev_t *pvd = vd; pvd != NULL; pvd = pvd->vdev_parent)
853			pvd->vdev_guid_sum += guid_delta;
854
855		/*
856		 * If this is a replacement, then we want to fallthrough to the
857		 * rest of the code.  If we're adding an l2cache, then it's
858		 * already labeled appropriately and we can just return.
859		 */
860		if (reason == VDEV_LABEL_L2CACHE)
861			return (0);
862		ASSERT(reason == VDEV_LABEL_REPLACE);
863	}
864
865	/*
866	 * Initialize its label.
867	 */
868	vp_abd = abd_alloc_linear(sizeof (vdev_phys_t), B_TRUE);
869	abd_zero(vp_abd, sizeof (vdev_phys_t));
870	vp = abd_to_buf(vp_abd);
871
872	/*
873	 * Generate a label describing the pool and our top-level vdev.
874	 * We mark it as being from txg 0 to indicate that it's not
875	 * really part of an active pool just yet.  The labels will
876	 * be written again with a meaningful txg by spa_sync().
877	 */
878	if (reason == VDEV_LABEL_SPARE ||
879	    (reason == VDEV_LABEL_REMOVE && vd->vdev_isspare)) {
880		/*
881		 * For inactive hot spares, we generate a special label that
882		 * identifies as a mutually shared hot spare.  We write the
883		 * label if we are adding a hot spare, or if we are removing an
884		 * active hot spare (in which case we want to revert the
885		 * labels).
886		 */
887		VERIFY(nvlist_alloc(&label, NV_UNIQUE_NAME, KM_SLEEP) == 0);
888
889		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_VERSION,
890		    spa_version(spa)) == 0);
891		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_POOL_STATE,
892		    POOL_STATE_SPARE) == 0);
893		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_GUID,
894		    vd->vdev_guid) == 0);
895	} else if (reason == VDEV_LABEL_L2CACHE ||
896	    (reason == VDEV_LABEL_REMOVE && vd->vdev_isl2cache)) {
897		/*
898		 * For level 2 ARC devices, add a special label.
899		 */
900		VERIFY(nvlist_alloc(&label, NV_UNIQUE_NAME, KM_SLEEP) == 0);
901
902		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_VERSION,
903		    spa_version(spa)) == 0);
904		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_POOL_STATE,
905		    POOL_STATE_L2CACHE) == 0);
906		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_GUID,
907		    vd->vdev_guid) == 0);
908	} else {
909		uint64_t txg = 0ULL;
910
911		if (reason == VDEV_LABEL_SPLIT)
912			txg = spa->spa_uberblock.ub_txg;
913		label = spa_config_generate(spa, vd, txg, B_FALSE);
914
915		/*
916		 * Add our creation time.  This allows us to detect multiple
917		 * vdev uses as described above, and automatically expires if we
918		 * fail.
919		 */
920		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_CREATE_TXG,
921		    crtxg) == 0);
922	}
923
924	buf = vp->vp_nvlist;
925	buflen = sizeof (vp->vp_nvlist);
926
927	error = nvlist_pack(label, &buf, &buflen, NV_ENCODE_XDR, KM_SLEEP);
928	if (error != 0) {
929		nvlist_free(label);
930		abd_free(vp_abd);
931		/* EFAULT means nvlist_pack ran out of room */
932		return (error == EFAULT ? ENAMETOOLONG : EINVAL);
933	}
934
935	/*
936	 * Initialize uberblock template.
937	 */
938	ub_abd = abd_alloc_linear(VDEV_UBERBLOCK_RING, B_TRUE);
939	abd_zero(ub_abd, VDEV_UBERBLOCK_RING);
940	abd_copy_from_buf(ub_abd, &spa->spa_uberblock, sizeof (uberblock_t));
941	ub = abd_to_buf(ub_abd);
942	ub->ub_txg = 0;
943
944	/* Initialize the 2nd padding area. */
945	pad2 = abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE);
946	abd_zero(pad2, VDEV_PAD_SIZE);
947
948	/*
949	 * Write everything in parallel.
950	 */
951retry:
952	zio = zio_root(spa, NULL, NULL, flags);
953
954	for (int l = 0; l < VDEV_LABELS; l++) {
955
956		vdev_label_write(zio, vd, l, vp_abd,
957		    offsetof(vdev_label_t, vl_vdev_phys),
958		    sizeof (vdev_phys_t), NULL, NULL, flags);
959
960		/*
961		 * Skip the 1st padding area.
962		 * Zero out the 2nd padding area where it might have
963		 * left over data from previous filesystem format.
964		 */
965		vdev_label_write(zio, vd, l, pad2,
966		    offsetof(vdev_label_t, vl_pad2),
967		    VDEV_PAD_SIZE, NULL, NULL, flags);
968
969		vdev_label_write(zio, vd, l, ub_abd,
970		    offsetof(vdev_label_t, vl_uberblock),
971		    VDEV_UBERBLOCK_RING, NULL, NULL, flags);
972	}
973
974	error = zio_wait(zio);
975
976	if (error != 0 && !(flags & ZIO_FLAG_TRYHARD)) {
977		flags |= ZIO_FLAG_TRYHARD;
978		goto retry;
979	}
980
981	nvlist_free(label);
982	abd_free(pad2);
983	abd_free(ub_abd);
984	abd_free(vp_abd);
985
986	/*
987	 * If this vdev hasn't been previously identified as a spare, then we
988	 * mark it as such only if a) we are labeling it as a spare, or b) it
989	 * exists as a spare elsewhere in the system.  Do the same for
990	 * level 2 ARC devices.
991	 */
992	if (error == 0 && !vd->vdev_isspare &&
993	    (reason == VDEV_LABEL_SPARE ||
994	    spa_spare_exists(vd->vdev_guid, NULL, NULL)))
995		spa_spare_add(vd);
996
997	if (error == 0 && !vd->vdev_isl2cache &&
998	    (reason == VDEV_LABEL_L2CACHE ||
999	    spa_l2cache_exists(vd->vdev_guid, NULL)))
1000		spa_l2cache_add(vd);
1001
1002	return (error);
1003}
1004
1005/*
1006 * ==========================================================================
1007 * uberblock load/sync
1008 * ==========================================================================
1009 */
1010
1011/*
1012 * Consider the following situation: txg is safely synced to disk.  We've
1013 * written the first uberblock for txg + 1, and then we lose power.  When we
1014 * come back up, we fail to see the uberblock for txg + 1 because, say,
1015 * it was on a mirrored device and the replica to which we wrote txg + 1
1016 * is now offline.  If we then make some changes and sync txg + 1, and then
1017 * the missing replica comes back, then for a few seconds we'll have two
1018 * conflicting uberblocks on disk with the same txg.  The solution is simple:
1019 * among uberblocks with equal txg, choose the one with the latest timestamp.
1020 */
1021static int
1022vdev_uberblock_compare(const uberblock_t *ub1, const uberblock_t *ub2)
1023{
1024	int cmp = AVL_CMP(ub1->ub_txg, ub2->ub_txg);
1025
1026	if (likely(cmp))
1027		return (cmp);
1028
1029	cmp = AVL_CMP(ub1->ub_timestamp, ub2->ub_timestamp);
1030	if (likely(cmp))
1031		return (cmp);
1032
1033	/*
1034	 * If MMP_VALID(ub) && MMP_SEQ_VALID(ub) then the host has an MMP-aware
1035	 * ZFS, e.g. zfsonlinux >= 0.7.
1036	 *
1037	 * If one ub has MMP and the other does not, they were written by
1038	 * different hosts, which matters for MMP.  So we treat no MMP/no SEQ as
1039	 * a 0 value.
1040	 *
1041	 * Since timestamp and txg are the same if we get this far, either is
1042	 * acceptable for importing the pool.
1043	 */
1044	unsigned int seq1 = 0;
1045	unsigned int seq2 = 0;
1046
1047	if (MMP_VALID(ub1) && MMP_SEQ_VALID(ub1))
1048		seq1 = MMP_SEQ(ub1);
1049
1050	if (MMP_VALID(ub2) && MMP_SEQ_VALID(ub2))
1051		seq2 = MMP_SEQ(ub2);
1052
1053	return (AVL_CMP(seq1, seq2));
1054}
1055
1056struct ubl_cbdata {
1057	uberblock_t	*ubl_ubbest;	/* Best uberblock */
1058	vdev_t		*ubl_vd;	/* vdev associated with the above */
1059};
1060
1061static void
1062vdev_uberblock_load_done(zio_t *zio)
1063{
1064	vdev_t *vd = zio->io_vd;
1065	spa_t *spa = zio->io_spa;
1066	zio_t *rio = zio->io_private;
1067	uberblock_t *ub = abd_to_buf(zio->io_abd);
1068	struct ubl_cbdata *cbp = rio->io_private;
1069
1070	ASSERT3U(zio->io_size, ==, VDEV_UBERBLOCK_SIZE(vd));
1071
1072	if (zio->io_error == 0 && uberblock_verify(ub) == 0) {
1073		mutex_enter(&rio->io_lock);
1074		if (ub->ub_txg <= spa->spa_load_max_txg &&
1075		    vdev_uberblock_compare(ub, cbp->ubl_ubbest) > 0) {
1076			/*
1077			 * Keep track of the vdev in which this uberblock
1078			 * was found. We will use this information later
1079			 * to obtain the config nvlist associated with
1080			 * this uberblock.
1081			 */
1082			*cbp->ubl_ubbest = *ub;
1083			cbp->ubl_vd = vd;
1084		}
1085		mutex_exit(&rio->io_lock);
1086	}
1087
1088	abd_free(zio->io_abd);
1089}
1090
1091static void
1092vdev_uberblock_load_impl(zio_t *zio, vdev_t *vd, int flags,
1093    struct ubl_cbdata *cbp)
1094{
1095	for (int c = 0; c < vd->vdev_children; c++)
1096		vdev_uberblock_load_impl(zio, vd->vdev_child[c], flags, cbp);
1097
1098	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1099		for (int l = 0; l < VDEV_LABELS; l++) {
1100			for (int n = 0; n < VDEV_UBERBLOCK_COUNT(vd); n++) {
1101				vdev_label_read(zio, vd, l,
1102				    abd_alloc_linear(VDEV_UBERBLOCK_SIZE(vd),
1103				    B_TRUE), VDEV_UBERBLOCK_OFFSET(vd, n),
1104				    VDEV_UBERBLOCK_SIZE(vd),
1105				    vdev_uberblock_load_done, zio, flags);
1106			}
1107		}
1108	}
1109}
1110
1111/*
1112 * Reads the 'best' uberblock from disk along with its associated
1113 * configuration. First, we read the uberblock array of each label of each
1114 * vdev, keeping track of the uberblock with the highest txg in each array.
1115 * Then, we read the configuration from the same vdev as the best uberblock.
1116 */
1117void
1118vdev_uberblock_load(vdev_t *rvd, uberblock_t *ub, nvlist_t **config)
1119{
1120	zio_t *zio;
1121	spa_t *spa = rvd->vdev_spa;
1122	struct ubl_cbdata cb;
1123	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL |
1124	    ZIO_FLAG_SPECULATIVE | ZIO_FLAG_TRYHARD;
1125
1126	ASSERT(ub);
1127	ASSERT(config);
1128
1129	bzero(ub, sizeof (uberblock_t));
1130	*config = NULL;
1131
1132	cb.ubl_ubbest = ub;
1133	cb.ubl_vd = NULL;
1134
1135	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
1136	zio = zio_root(spa, NULL, &cb, flags);
1137	vdev_uberblock_load_impl(zio, rvd, flags, &cb);
1138	(void) zio_wait(zio);
1139
1140	/*
1141	 * It's possible that the best uberblock was discovered on a label
1142	 * that has a configuration which was written in a future txg.
1143	 * Search all labels on this vdev to find the configuration that
1144	 * matches the txg for our uberblock.
1145	 */
1146	if (cb.ubl_vd != NULL) {
1147		vdev_dbgmsg(cb.ubl_vd, "best uberblock found for spa %s. "
1148		    "txg %llu", spa->spa_name, (u_longlong_t)ub->ub_txg);
1149
1150		*config = vdev_label_read_config(cb.ubl_vd, ub->ub_txg);
1151		if (*config == NULL && spa->spa_extreme_rewind) {
1152			vdev_dbgmsg(cb.ubl_vd, "failed to read label config. "
1153			    "Trying again without txg restrictions.");
1154			*config = vdev_label_read_config(cb.ubl_vd, UINT64_MAX);
1155		}
1156		if (*config == NULL) {
1157			vdev_dbgmsg(cb.ubl_vd, "failed to read label config");
1158		}
1159	}
1160	spa_config_exit(spa, SCL_ALL, FTAG);
1161}
1162
1163/*
1164 * On success, increment root zio's count of good writes.
1165 * We only get credit for writes to known-visible vdevs; see spa_vdev_add().
1166 */
1167static void
1168vdev_uberblock_sync_done(zio_t *zio)
1169{
1170	uint64_t *good_writes = zio->io_private;
1171
1172	if (zio->io_error == 0 && zio->io_vd->vdev_top->vdev_ms_array != 0)
1173		atomic_inc_64(good_writes);
1174}
1175
1176/*
1177 * Write the uberblock to all labels of all leaves of the specified vdev.
1178 */
1179static void
1180vdev_uberblock_sync(zio_t *zio, uint64_t *good_writes,
1181    uberblock_t *ub, vdev_t *vd, int flags)
1182{
1183	for (uint64_t c = 0; c < vd->vdev_children; c++) {
1184		vdev_uberblock_sync(zio, good_writes,
1185		    ub, vd->vdev_child[c], flags);
1186	}
1187
1188	if (!vd->vdev_ops->vdev_op_leaf)
1189		return;
1190
1191	if (!vdev_writeable(vd))
1192		return;
1193
1194	int m = spa_multihost(vd->vdev_spa) ? MMP_BLOCKS_PER_LABEL : 0;
1195	int n = ub->ub_txg % (VDEV_UBERBLOCK_COUNT(vd) - m);
1196
1197	/* Copy the uberblock_t into the ABD */
1198	abd_t *ub_abd = abd_alloc_for_io(VDEV_UBERBLOCK_SIZE(vd), B_TRUE);
1199	abd_zero(ub_abd, VDEV_UBERBLOCK_SIZE(vd));
1200	abd_copy_from_buf(ub_abd, ub, sizeof (uberblock_t));
1201
1202	for (int l = 0; l < VDEV_LABELS; l++)
1203		vdev_label_write(zio, vd, l, ub_abd,
1204		    VDEV_UBERBLOCK_OFFSET(vd, n), VDEV_UBERBLOCK_SIZE(vd),
1205		    vdev_uberblock_sync_done, good_writes,
1206		    flags | ZIO_FLAG_DONT_PROPAGATE);
1207
1208	abd_free(ub_abd);
1209}
1210
1211/* Sync the uberblocks to all vdevs in svd[] */
1212int
1213vdev_uberblock_sync_list(vdev_t **svd, int svdcount, uberblock_t *ub, int flags)
1214{
1215	spa_t *spa = svd[0]->vdev_spa;
1216	zio_t *zio;
1217	uint64_t good_writes = 0;
1218
1219	zio = zio_root(spa, NULL, NULL, flags);
1220
1221	for (int v = 0; v < svdcount; v++)
1222		vdev_uberblock_sync(zio, &good_writes, ub, svd[v], flags);
1223
1224	(void) zio_wait(zio);
1225
1226	/*
1227	 * Flush the uberblocks to disk.  This ensures that the odd labels
1228	 * are no longer needed (because the new uberblocks and the even
1229	 * labels are safely on disk), so it is safe to overwrite them.
1230	 */
1231	zio = zio_root(spa, NULL, NULL, flags);
1232
1233	for (int v = 0; v < svdcount; v++) {
1234		if (vdev_writeable(svd[v])) {
1235			zio_flush(zio, svd[v]);
1236		}
1237	}
1238
1239	(void) zio_wait(zio);
1240
1241	return (good_writes >= 1 ? 0 : EIO);
1242}
1243
1244/*
1245 * On success, increment the count of good writes for our top-level vdev.
1246 */
1247static void
1248vdev_label_sync_done(zio_t *zio)
1249{
1250	uint64_t *good_writes = zio->io_private;
1251
1252	if (zio->io_error == 0)
1253		atomic_inc_64(good_writes);
1254}
1255
1256/*
1257 * If there weren't enough good writes, indicate failure to the parent.
1258 */
1259static void
1260vdev_label_sync_top_done(zio_t *zio)
1261{
1262	uint64_t *good_writes = zio->io_private;
1263
1264	if (*good_writes == 0)
1265		zio->io_error = SET_ERROR(EIO);
1266
1267	kmem_free(good_writes, sizeof (uint64_t));
1268}
1269
1270/*
1271 * We ignore errors for log and cache devices, simply free the private data.
1272 */
1273static void
1274vdev_label_sync_ignore_done(zio_t *zio)
1275{
1276	kmem_free(zio->io_private, sizeof (uint64_t));
1277}
1278
1279/*
1280 * Write all even or odd labels to all leaves of the specified vdev.
1281 */
1282static void
1283vdev_label_sync(zio_t *zio, uint64_t *good_writes,
1284    vdev_t *vd, int l, uint64_t txg, int flags)
1285{
1286	nvlist_t *label;
1287	vdev_phys_t *vp;
1288	abd_t *vp_abd;
1289	char *buf;
1290	size_t buflen;
1291
1292	for (int c = 0; c < vd->vdev_children; c++) {
1293		vdev_label_sync(zio, good_writes,
1294		    vd->vdev_child[c], l, txg, flags);
1295	}
1296
1297	if (!vd->vdev_ops->vdev_op_leaf)
1298		return;
1299
1300	if (!vdev_writeable(vd))
1301		return;
1302
1303	/*
1304	 * Generate a label describing the top-level config to which we belong.
1305	 */
1306	label = spa_config_generate(vd->vdev_spa, vd, txg, B_FALSE);
1307
1308	vp_abd = abd_alloc_linear(sizeof (vdev_phys_t), B_TRUE);
1309	abd_zero(vp_abd, sizeof (vdev_phys_t));
1310	vp = abd_to_buf(vp_abd);
1311
1312	buf = vp->vp_nvlist;
1313	buflen = sizeof (vp->vp_nvlist);
1314
1315	if (nvlist_pack(label, &buf, &buflen, NV_ENCODE_XDR, KM_SLEEP) == 0) {
1316		for (; l < VDEV_LABELS; l += 2) {
1317			vdev_label_write(zio, vd, l, vp_abd,
1318			    offsetof(vdev_label_t, vl_vdev_phys),
1319			    sizeof (vdev_phys_t),
1320			    vdev_label_sync_done, good_writes,
1321			    flags | ZIO_FLAG_DONT_PROPAGATE);
1322		}
1323	}
1324
1325	abd_free(vp_abd);
1326	nvlist_free(label);
1327}
1328
1329int
1330vdev_label_sync_list(spa_t *spa, int l, uint64_t txg, int flags)
1331{
1332	list_t *dl = &spa->spa_config_dirty_list;
1333	vdev_t *vd;
1334	zio_t *zio;
1335	int error;
1336
1337	/*
1338	 * Write the new labels to disk.
1339	 */
1340	zio = zio_root(spa, NULL, NULL, flags);
1341
1342	for (vd = list_head(dl); vd != NULL; vd = list_next(dl, vd)) {
1343		uint64_t *good_writes = kmem_zalloc(sizeof (uint64_t),
1344		    KM_SLEEP);
1345
1346		ASSERT(!vd->vdev_ishole);
1347
1348		zio_t *vio = zio_null(zio, spa, NULL,
1349		    (vd->vdev_islog || vd->vdev_aux != NULL) ?
1350		    vdev_label_sync_ignore_done : vdev_label_sync_top_done,
1351		    good_writes, flags);
1352		vdev_label_sync(vio, good_writes, vd, l, txg, flags);
1353		zio_nowait(vio);
1354	}
1355
1356	error = zio_wait(zio);
1357
1358	/*
1359	 * Flush the new labels to disk.
1360	 */
1361	zio = zio_root(spa, NULL, NULL, flags);
1362
1363	for (vd = list_head(dl); vd != NULL; vd = list_next(dl, vd))
1364		zio_flush(zio, vd);
1365
1366	(void) zio_wait(zio);
1367
1368	return (error);
1369}
1370
1371/*
1372 * Sync the uberblock and any changes to the vdev configuration.
1373 *
1374 * The order of operations is carefully crafted to ensure that
1375 * if the system panics or loses power at any time, the state on disk
1376 * is still transactionally consistent.  The in-line comments below
1377 * describe the failure semantics at each stage.
1378 *
1379 * Moreover, vdev_config_sync() is designed to be idempotent: if it fails
1380 * at any time, you can just call it again, and it will resume its work.
1381 */
1382int
1383vdev_config_sync(vdev_t **svd, int svdcount, uint64_t txg)
1384{
1385	spa_t *spa = svd[0]->vdev_spa;
1386	uberblock_t *ub = &spa->spa_uberblock;
1387	int error = 0;
1388	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL;
1389
1390	ASSERT(svdcount != 0);
1391retry:
1392	/*
1393	 * Normally, we don't want to try too hard to write every label and
1394	 * uberblock.  If there is a flaky disk, we don't want the rest of the
1395	 * sync process to block while we retry.  But if we can't write a
1396	 * single label out, we should retry with ZIO_FLAG_TRYHARD before
1397	 * bailing out and declaring the pool faulted.
1398	 */
1399	if (error != 0) {
1400		if ((flags & ZIO_FLAG_TRYHARD) != 0)
1401			return (error);
1402		flags |= ZIO_FLAG_TRYHARD;
1403	}
1404
1405	ASSERT(ub->ub_txg <= txg);
1406
1407	/*
1408	 * If this isn't a resync due to I/O errors,
1409	 * and nothing changed in this transaction group,
1410	 * and the vdev configuration hasn't changed,
1411	 * then there's nothing to do.
1412	 */
1413	if (ub->ub_txg < txg) {
1414		boolean_t changed = uberblock_update(ub, spa->spa_root_vdev,
1415		    txg, spa->spa_mmp.mmp_delay);
1416
1417		if (!changed && list_is_empty(&spa->spa_config_dirty_list))
1418			return (0);
1419	}
1420
1421	if (txg > spa_freeze_txg(spa))
1422		return (0);
1423
1424	ASSERT(txg <= spa->spa_final_txg);
1425
1426	/*
1427	 * Flush the write cache of every disk that's been written to
1428	 * in this transaction group.  This ensures that all blocks
1429	 * written in this txg will be committed to stable storage
1430	 * before any uberblock that references them.
1431	 */
1432	zio_t *zio = zio_root(spa, NULL, NULL, flags);
1433
1434	for (vdev_t *vd =
1435	    txg_list_head(&spa->spa_vdev_txg_list, TXG_CLEAN(txg)); vd != NULL;
1436	    vd = txg_list_next(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg)))
1437		zio_flush(zio, vd);
1438
1439	(void) zio_wait(zio);
1440
1441	/*
1442	 * Sync out the even labels (L0, L2) for every dirty vdev.  If the
1443	 * system dies in the middle of this process, that's OK: all of the
1444	 * even labels that made it to disk will be newer than any uberblock,
1445	 * and will therefore be considered invalid.  The odd labels (L1, L3),
1446	 * which have not yet been touched, will still be valid.  We flush
1447	 * the new labels to disk to ensure that all even-label updates
1448	 * are committed to stable storage before the uberblock update.
1449	 */
1450	if ((error = vdev_label_sync_list(spa, 0, txg, flags)) != 0) {
1451		if ((flags & ZIO_FLAG_TRYHARD) != 0) {
1452			zfs_dbgmsg("vdev_label_sync_list() returned error %d "
1453			    "for pool '%s' when syncing out the even labels "
1454			    "of dirty vdevs", error, spa_name(spa));
1455		}
1456		goto retry;
1457	}
1458
1459	/*
1460	 * Sync the uberblocks to all vdevs in svd[].
1461	 * If the system dies in the middle of this step, there are two cases
1462	 * to consider, and the on-disk state is consistent either way:
1463	 *
1464	 * (1)	If none of the new uberblocks made it to disk, then the
1465	 *	previous uberblock will be the newest, and the odd labels
1466	 *	(which had not yet been touched) will be valid with respect
1467	 *	to that uberblock.
1468	 *
1469	 * (2)	If one or more new uberblocks made it to disk, then they
1470	 *	will be the newest, and the even labels (which had all
1471	 *	been successfully committed) will be valid with respect
1472	 *	to the new uberblocks.
1473	 */
1474	if ((error = vdev_uberblock_sync_list(svd, svdcount, ub, flags)) != 0) {
1475		if ((flags & ZIO_FLAG_TRYHARD) != 0) {
1476			zfs_dbgmsg("vdev_uberblock_sync_list() returned error "
1477			    "%d for pool '%s'", error, spa_name(spa));
1478		}
1479		goto retry;
1480	}
1481
1482	if (spa_multihost(spa))
1483		mmp_update_uberblock(spa, ub);
1484
1485	/*
1486	 * Sync out odd labels for every dirty vdev.  If the system dies
1487	 * in the middle of this process, the even labels and the new
1488	 * uberblocks will suffice to open the pool.  The next time
1489	 * the pool is opened, the first thing we'll do -- before any
1490	 * user data is modified -- is mark every vdev dirty so that
1491	 * all labels will be brought up to date.  We flush the new labels
1492	 * to disk to ensure that all odd-label updates are committed to
1493	 * stable storage before the next transaction group begins.
1494	 */
1495	if ((error = vdev_label_sync_list(spa, 1, txg, flags)) != 0) {
1496		if ((flags & ZIO_FLAG_TRYHARD) != 0) {
1497			zfs_dbgmsg("vdev_label_sync_list() returned error %d "
1498			    "for pool '%s' when syncing out the odd labels of "
1499			    "dirty vdevs", error, spa_name(spa));
1500		}
1501		goto retry;
1502	}
1503
1504	return (0);
1505}
1506