xref: /netbsd-src/external/cddl/osnet/dist/uts/common/fs/zfs/vdev_label.c (revision 4122fc0e661b34d3b0d1c8c49b7ed9f8e1eabca2)
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, 2015 by Delphix. All rights reserved.
25  */
26 
27 /*
28  * Virtual Device Labels
29  * ---------------------
30  *
31  * The vdev label serves several distinct purposes:
32  *
33  *	1. Uniquely identify this device as part of a ZFS pool and confirm its
34  *	   identity within the pool.
35  *
36  * 	2. Verify that all the devices given in a configuration are present
37  *         within the pool.
38  *
39  * 	3. Determine the uberblock for the pool.
40  *
41  * 	4. In case of an import operation, determine the configuration of the
42  *         toplevel vdev of which it is a part.
43  *
44  * 	5. If an import operation cannot find all the devices in the pool,
45  *         provide enough information to the administrator to determine which
46  *         devices are missing.
47  *
48  * It is important to note that while the kernel is responsible for writing the
49  * label, it only consumes the information in the first three cases.  The
50  * latter information is only consumed in userland when determining the
51  * configuration to import a pool.
52  *
53  *
54  * Label Organization
55  * ------------------
56  *
57  * Before describing the contents of the label, it's important to understand how
58  * the labels are written and updated with respect to the uberblock.
59  *
60  * When the pool configuration is altered, either because it was newly created
61  * or a device was added, we want to update all the labels such that we can deal
62  * with fatal failure at any point.  To this end, each disk has two labels which
63  * are updated before and after the uberblock is synced.  Assuming we have
64  * labels and an uberblock with the following transaction groups:
65  *
66  *              L1          UB          L2
67  *           +------+    +------+    +------+
68  *           |      |    |      |    |      |
69  *           | t10  |    | t10  |    | t10  |
70  *           |      |    |      |    |      |
71  *           +------+    +------+    +------+
72  *
73  * In this stable state, the labels and the uberblock were all updated within
74  * the same transaction group (10).  Each label is mirrored and checksummed, so
75  * that we can detect when we fail partway through writing the label.
76  *
77  * In order to identify which labels are valid, the labels are written in the
78  * following manner:
79  *
80  * 	1. For each vdev, update 'L1' to the new label
81  * 	2. Update the uberblock
82  * 	3. For each vdev, update 'L2' to the new label
83  *
84  * Given arbitrary failure, we can determine the correct label to use based on
85  * the transaction group.  If we fail after updating L1 but before updating the
86  * UB, we will notice that L1's transaction group is greater than the uberblock,
87  * so L2 must be valid.  If we fail after writing the uberblock but before
88  * writing L2, we will notice that L2's transaction group is less than L1, and
89  * therefore L1 is valid.
90  *
91  * Another added complexity is that not every label is updated when the config
92  * is synced.  If we add a single device, we do not want to have to re-write
93  * every label for every device in the pool.  This means that both L1 and L2 may
94  * be older than the pool uberblock, because the necessary information is stored
95  * on another vdev.
96  *
97  *
98  * On-disk Format
99  * --------------
100  *
101  * The vdev label consists of two distinct parts, and is wrapped within the
102  * vdev_label_t structure.  The label includes 8k of padding to permit legacy
103  * VTOC disk labels, but is otherwise ignored.
104  *
105  * The first half of the label is a packed nvlist which contains pool wide
106  * properties, per-vdev properties, and configuration information.  It is
107  * described in more detail below.
108  *
109  * The latter half of the label consists of a redundant array of uberblocks.
110  * These uberblocks are updated whenever a transaction group is committed,
111  * or when the configuration is updated.  When a pool is loaded, we scan each
112  * vdev for the 'best' uberblock.
113  *
114  *
115  * Configuration Information
116  * -------------------------
117  *
118  * The nvlist describing the pool and vdev contains the following elements:
119  *
120  * 	version		ZFS on-disk version
121  * 	name		Pool name
122  * 	state		Pool state
123  * 	txg		Transaction group in which this label was written
124  * 	pool_guid	Unique identifier for this pool
125  * 	vdev_tree	An nvlist describing vdev tree.
126  *	features_for_read
127  *			An nvlist of the features necessary for reading the MOS.
128  *
129  * Each leaf device label also contains the following:
130  *
131  * 	top_guid	Unique ID for top-level vdev in which this is contained
132  * 	guid		Unique ID for the leaf vdev
133  *
134  * The 'vs' configuration follows the format described in 'spa_config.c'.
135  */
136 
137 #include <sys/zfs_context.h>
138 #include <sys/spa.h>
139 #include <sys/spa_impl.h>
140 #include <sys/dmu.h>
141 #include <sys/zap.h>
142 #include <sys/vdev.h>
143 #include <sys/vdev_impl.h>
144 #include <sys/uberblock_impl.h>
145 #include <sys/metaslab.h>
146 #include <sys/zio.h>
147 #include <sys/dsl_scan.h>
148 #include <sys/trim_map.h>
149 #include <sys/fs/zfs.h>
150 
151 static boolean_t vdev_trim_on_init = B_TRUE;
152 SYSCTL_DECL(_vfs_zfs_vdev);
153 SYSCTL_INT(_vfs_zfs_vdev, OID_AUTO, trim_on_init, CTLFLAG_RW,
154     &vdev_trim_on_init, 0, "Enable/disable full vdev trim on initialisation");
155 
156 /*
157  * Basic routines to read and write from a vdev label.
158  * Used throughout the rest of this file.
159  */
160 uint64_t
vdev_label_offset(uint64_t psize,int l,uint64_t offset)161 vdev_label_offset(uint64_t psize, int l, uint64_t offset)
162 {
163 	ASSERT(offset < sizeof (vdev_label_t));
164 	ASSERT(P2PHASE_TYPED(psize, sizeof (vdev_label_t), uint64_t) == 0);
165 
166 	return (offset + l * sizeof (vdev_label_t) + (l < VDEV_LABELS / 2 ?
167 	    0 : psize - VDEV_LABELS * sizeof (vdev_label_t)));
168 }
169 
170 /*
171  * Returns back the vdev label associated with the passed in offset.
172  */
173 int
vdev_label_number(uint64_t psize,uint64_t offset)174 vdev_label_number(uint64_t psize, uint64_t offset)
175 {
176 	int l;
177 
178 	if (offset >= psize - VDEV_LABEL_END_SIZE) {
179 		offset -= psize - VDEV_LABEL_END_SIZE;
180 		offset += (VDEV_LABELS / 2) * sizeof (vdev_label_t);
181 	}
182 	l = offset / sizeof (vdev_label_t);
183 	return (l < VDEV_LABELS ? l : -1);
184 }
185 
186 static void
vdev_label_read(zio_t * zio,vdev_t * vd,int l,void * buf,uint64_t offset,uint64_t size,zio_done_func_t * done,void * private,int flags)187 vdev_label_read(zio_t *zio, vdev_t *vd, int l, void *buf, uint64_t offset,
188     uint64_t size, zio_done_func_t *done, void *private, int flags)
189 {
190 	ASSERT(spa_config_held(zio->io_spa, SCL_STATE_ALL, RW_WRITER) ==
191 	    SCL_STATE_ALL);
192 	ASSERT(flags & ZIO_FLAG_CONFIG_WRITER);
193 
194 	zio_nowait(zio_read_phys(zio, vd,
195 	    vdev_label_offset(vd->vdev_psize, l, offset),
196 	    size, buf, ZIO_CHECKSUM_LABEL, done, private,
197 	    ZIO_PRIORITY_SYNC_READ, flags, B_TRUE));
198 }
199 
200 static void
vdev_label_write(zio_t * zio,vdev_t * vd,int l,void * buf,uint64_t offset,uint64_t size,zio_done_func_t * done,void * private,int flags)201 vdev_label_write(zio_t *zio, vdev_t *vd, int l, void *buf, uint64_t offset,
202     uint64_t size, zio_done_func_t *done, void *private, int flags)
203 {
204 	ASSERT(spa_config_held(zio->io_spa, SCL_ALL, RW_WRITER) == SCL_ALL ||
205 	    (spa_config_held(zio->io_spa, SCL_CONFIG | SCL_STATE, RW_READER) ==
206 	    (SCL_CONFIG | SCL_STATE) &&
207 	    dsl_pool_sync_context(spa_get_dsl(zio->io_spa))));
208 	ASSERT(flags & ZIO_FLAG_CONFIG_WRITER);
209 
210 	zio_nowait(zio_write_phys(zio, vd,
211 	    vdev_label_offset(vd->vdev_psize, l, offset),
212 	    size, buf, ZIO_CHECKSUM_LABEL, done, private,
213 	    ZIO_PRIORITY_SYNC_WRITE, flags, B_TRUE));
214 }
215 
216 /*
217  * Generate the nvlist representing this vdev's config.
218  */
219 nvlist_t *
vdev_config_generate(spa_t * spa,vdev_t * vd,boolean_t getstats,vdev_config_flag_t flags)220 vdev_config_generate(spa_t *spa, vdev_t *vd, boolean_t getstats,
221     vdev_config_flag_t flags)
222 {
223 	nvlist_t *nv = NULL;
224 
225 	nv = fnvlist_alloc();
226 
227 	fnvlist_add_string(nv, ZPOOL_CONFIG_TYPE, vd->vdev_ops->vdev_op_type);
228 	if (!(flags & (VDEV_CONFIG_SPARE | VDEV_CONFIG_L2CACHE)))
229 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_ID, vd->vdev_id);
230 	fnvlist_add_uint64(nv, ZPOOL_CONFIG_GUID, vd->vdev_guid);
231 
232 	if (vd->vdev_path != NULL)
233 		fnvlist_add_string(nv, ZPOOL_CONFIG_PATH, vd->vdev_path);
234 
235 	if (vd->vdev_devid != NULL)
236 		fnvlist_add_string(nv, ZPOOL_CONFIG_DEVID, vd->vdev_devid);
237 
238 	if (vd->vdev_physpath != NULL)
239 		fnvlist_add_string(nv, ZPOOL_CONFIG_PHYS_PATH,
240 		    vd->vdev_physpath);
241 
242 	if (vd->vdev_fru != NULL)
243 		fnvlist_add_string(nv, ZPOOL_CONFIG_FRU, vd->vdev_fru);
244 
245 	if (vd->vdev_nparity != 0) {
246 		ASSERT(strcmp(vd->vdev_ops->vdev_op_type,
247 		    VDEV_TYPE_RAIDZ) == 0);
248 
249 		/*
250 		 * Make sure someone hasn't managed to sneak a fancy new vdev
251 		 * into a crufty old storage pool.
252 		 */
253 		ASSERT(vd->vdev_nparity == 1 ||
254 		    (vd->vdev_nparity <= 2 &&
255 		    spa_version(spa) >= SPA_VERSION_RAIDZ2) ||
256 		    (vd->vdev_nparity <= 3 &&
257 		    spa_version(spa) >= SPA_VERSION_RAIDZ3));
258 
259 		/*
260 		 * Note that we'll add the nparity tag even on storage pools
261 		 * that only support a single parity device -- older software
262 		 * will just ignore it.
263 		 */
264 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_NPARITY, vd->vdev_nparity);
265 	}
266 
267 	if (vd->vdev_wholedisk != -1ULL)
268 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
269 		    vd->vdev_wholedisk);
270 
271 	if (vd->vdev_not_present)
272 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT, 1);
273 
274 	if (vd->vdev_isspare)
275 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_IS_SPARE, 1);
276 
277 	if (!(flags & (VDEV_CONFIG_SPARE | VDEV_CONFIG_L2CACHE)) &&
278 	    vd == vd->vdev_top) {
279 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
280 		    vd->vdev_ms_array);
281 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
282 		    vd->vdev_ms_shift);
283 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_ASHIFT, vd->vdev_ashift);
284 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_ASIZE,
285 		    vd->vdev_asize);
286 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_IS_LOG, vd->vdev_islog);
287 		if (vd->vdev_removing)
288 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_REMOVING,
289 			    vd->vdev_removing);
290 	}
291 
292 	if (vd->vdev_dtl_sm != NULL) {
293 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_DTL,
294 		    space_map_object(vd->vdev_dtl_sm));
295 	}
296 
297 	if (vd->vdev_crtxg)
298 		fnvlist_add_uint64(nv, ZPOOL_CONFIG_CREATE_TXG, vd->vdev_crtxg);
299 
300 	if (flags & VDEV_CONFIG_MOS) {
301 		if (vd->vdev_leaf_zap != 0) {
302 			ASSERT(vd->vdev_ops->vdev_op_leaf);
303 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_VDEV_LEAF_ZAP,
304 			    vd->vdev_leaf_zap);
305 		}
306 
307 		if (vd->vdev_top_zap != 0) {
308 			ASSERT(vd == vd->vdev_top);
309 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
310 			    vd->vdev_top_zap);
311 		}
312 	}
313 
314 	if (getstats) {
315 		vdev_stat_t vs;
316 		pool_scan_stat_t ps;
317 
318 		vdev_get_stats(vd, &vs);
319 		fnvlist_add_uint64_array(nv, ZPOOL_CONFIG_VDEV_STATS,
320 		    (uint64_t *)&vs, sizeof (vs) / sizeof (uint64_t));
321 
322 		/* provide either current or previous scan information */
323 		if (spa_scan_get_stats(spa, &ps) == 0) {
324 			fnvlist_add_uint64_array(nv,
325 			    ZPOOL_CONFIG_SCAN_STATS, (uint64_t *)&ps,
326 			    sizeof (pool_scan_stat_t) / sizeof (uint64_t));
327 		}
328 	}
329 
330 	if (!vd->vdev_ops->vdev_op_leaf) {
331 		nvlist_t **child;
332 		int c, idx;
333 
334 		ASSERT(!vd->vdev_ishole);
335 
336 		child = kmem_alloc(vd->vdev_children * sizeof (nvlist_t *),
337 		    KM_SLEEP);
338 
339 		for (c = 0, idx = 0; c < vd->vdev_children; c++) {
340 			vdev_t *cvd = vd->vdev_child[c];
341 
342 			/*
343 			 * If we're generating an nvlist of removing
344 			 * vdevs then skip over any device which is
345 			 * not being removed.
346 			 */
347 			if ((flags & VDEV_CONFIG_REMOVING) &&
348 			    !cvd->vdev_removing)
349 				continue;
350 
351 			child[idx++] = vdev_config_generate(spa, cvd,
352 			    getstats, flags);
353 		}
354 
355 		if (idx) {
356 			fnvlist_add_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
357 			    child, idx);
358 		}
359 
360 		for (c = 0; c < idx; c++)
361 			nvlist_free(child[c]);
362 
363 		kmem_free(child, vd->vdev_children * sizeof (nvlist_t *));
364 
365 	} else {
366 		const char *aux = NULL;
367 
368 		if (vd->vdev_offline && !vd->vdev_tmpoffline)
369 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_OFFLINE, B_TRUE);
370 		if (vd->vdev_resilver_txg != 0)
371 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
372 			    vd->vdev_resilver_txg);
373 		if (vd->vdev_faulted)
374 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_FAULTED, B_TRUE);
375 		if (vd->vdev_degraded)
376 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_DEGRADED, B_TRUE);
377 		if (vd->vdev_removed)
378 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_REMOVED, B_TRUE);
379 		if (vd->vdev_unspare)
380 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_UNSPARE, B_TRUE);
381 		if (vd->vdev_ishole)
382 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_IS_HOLE, B_TRUE);
383 
384 		switch (vd->vdev_stat.vs_aux) {
385 		case VDEV_AUX_ERR_EXCEEDED:
386 			aux = "err_exceeded";
387 			break;
388 
389 		case VDEV_AUX_EXTERNAL:
390 			aux = "external";
391 			break;
392 		}
393 
394 		if (aux != NULL)
395 			fnvlist_add_string(nv, ZPOOL_CONFIG_AUX_STATE, aux);
396 
397 		if (vd->vdev_splitting && vd->vdev_orig_guid != 0LL) {
398 			fnvlist_add_uint64(nv, ZPOOL_CONFIG_ORIG_GUID,
399 			    vd->vdev_orig_guid);
400 		}
401 	}
402 
403 	return (nv);
404 }
405 
406 /*
407  * Generate a view of the top-level vdevs.  If we currently have holes
408  * in the namespace, then generate an array which contains a list of holey
409  * vdevs.  Additionally, add the number of top-level children that currently
410  * exist.
411  */
412 void
vdev_top_config_generate(spa_t * spa,nvlist_t * config)413 vdev_top_config_generate(spa_t *spa, nvlist_t *config)
414 {
415 	vdev_t *rvd = spa->spa_root_vdev;
416 	uint64_t *array;
417 	uint_t c, idx;
418 
419 	array = kmem_alloc(rvd->vdev_children * sizeof (uint64_t), KM_SLEEP);
420 
421 	for (c = 0, idx = 0; c < rvd->vdev_children; c++) {
422 		vdev_t *tvd = rvd->vdev_child[c];
423 
424 		if (tvd->vdev_ishole)
425 			array[idx++] = c;
426 	}
427 
428 	if (idx) {
429 		VERIFY(nvlist_add_uint64_array(config, ZPOOL_CONFIG_HOLE_ARRAY,
430 		    array, idx) == 0);
431 	}
432 
433 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_VDEV_CHILDREN,
434 	    rvd->vdev_children) == 0);
435 
436 	kmem_free(array, rvd->vdev_children * sizeof (uint64_t));
437 }
438 
439 /*
440  * Returns the configuration from the label of the given vdev. For vdevs
441  * which don't have a txg value stored on their label (i.e. spares/cache)
442  * or have not been completely initialized (txg = 0) just return
443  * the configuration from the first valid label we find. Otherwise,
444  * find the most up-to-date label that does not exceed the specified
445  * 'txg' value.
446  */
447 nvlist_t *
vdev_label_read_config(vdev_t * vd,uint64_t txg)448 vdev_label_read_config(vdev_t *vd, uint64_t txg)
449 {
450 	spa_t *spa = vd->vdev_spa;
451 	nvlist_t *config = NULL;
452 	vdev_phys_t *vp;
453 	zio_t *zio;
454 	uint64_t best_txg = 0;
455 	int error = 0;
456 	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL |
457 	    ZIO_FLAG_SPECULATIVE;
458 
459 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
460 
461 	if (!vdev_readable(vd))
462 		return (NULL);
463 
464 	vp = zio_buf_alloc(sizeof (vdev_phys_t));
465 
466 retry:
467 	for (int l = 0; l < VDEV_LABELS; l++) {
468 		nvlist_t *label = NULL;
469 
470 		zio = zio_root(spa, NULL, NULL, flags);
471 
472 		vdev_label_read(zio, vd, l, vp,
473 		    offsetof(vdev_label_t, vl_vdev_phys),
474 		    sizeof (vdev_phys_t), NULL, NULL, flags);
475 
476 		if (zio_wait(zio) == 0 &&
477 		    nvlist_unpack(vp->vp_nvlist, sizeof (vp->vp_nvlist),
478 		    &label, 0) == 0) {
479 			uint64_t label_txg = 0;
480 
481 			/*
482 			 * Auxiliary vdevs won't have txg values in their
483 			 * labels and newly added vdevs may not have been
484 			 * completely initialized so just return the
485 			 * configuration from the first valid label we
486 			 * encounter.
487 			 */
488 			error = nvlist_lookup_uint64(label,
489 			    ZPOOL_CONFIG_POOL_TXG, &label_txg);
490 			if ((error || label_txg == 0) && !config) {
491 				config = label;
492 				break;
493 			} else if (label_txg <= txg && label_txg > best_txg) {
494 				best_txg = label_txg;
495 				nvlist_free(config);
496 				config = fnvlist_dup(label);
497 			}
498 		}
499 
500 		if (label != NULL) {
501 			nvlist_free(label);
502 			label = NULL;
503 		}
504 	}
505 
506 	if (config == NULL && !(flags & ZIO_FLAG_TRYHARD)) {
507 		flags |= ZIO_FLAG_TRYHARD;
508 		goto retry;
509 	}
510 
511 	zio_buf_free(vp, sizeof (vdev_phys_t));
512 
513 	return (config);
514 }
515 
516 /*
517  * Determine if a device is in use.  The 'spare_guid' parameter will be filled
518  * in with the device guid if this spare is active elsewhere on the system.
519  */
520 static boolean_t
vdev_inuse(vdev_t * vd,uint64_t crtxg,vdev_labeltype_t reason,uint64_t * spare_guid,uint64_t * l2cache_guid)521 vdev_inuse(vdev_t *vd, uint64_t crtxg, vdev_labeltype_t reason,
522     uint64_t *spare_guid, uint64_t *l2cache_guid)
523 {
524 	spa_t *spa = vd->vdev_spa;
525 	uint64_t state, pool_guid, device_guid, txg, spare_pool;
526 	uint64_t vdtxg = 0;
527 	nvlist_t *label;
528 
529 	if (spare_guid)
530 		*spare_guid = 0ULL;
531 	if (l2cache_guid)
532 		*l2cache_guid = 0ULL;
533 
534 	/*
535 	 * Read the label, if any, and perform some basic sanity checks.
536 	 */
537 	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL)
538 		return (B_FALSE);
539 
540 	(void) nvlist_lookup_uint64(label, ZPOOL_CONFIG_CREATE_TXG,
541 	    &vdtxg);
542 
543 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
544 	    &state) != 0 ||
545 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
546 	    &device_guid) != 0) {
547 		nvlist_free(label);
548 		return (B_FALSE);
549 	}
550 
551 	if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
552 	    (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID,
553 	    &pool_guid) != 0 ||
554 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_TXG,
555 	    &txg) != 0)) {
556 		nvlist_free(label);
557 		return (B_FALSE);
558 	}
559 
560 	nvlist_free(label);
561 
562 	/*
563 	 * Check to see if this device indeed belongs to the pool it claims to
564 	 * be a part of.  The only way this is allowed is if the device is a hot
565 	 * spare (which we check for later on).
566 	 */
567 	if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
568 	    !spa_guid_exists(pool_guid, device_guid) &&
569 	    !spa_spare_exists(device_guid, NULL, NULL) &&
570 	    !spa_l2cache_exists(device_guid, NULL))
571 		return (B_FALSE);
572 
573 	/*
574 	 * If the transaction group is zero, then this an initialized (but
575 	 * unused) label.  This is only an error if the create transaction
576 	 * on-disk is the same as the one we're using now, in which case the
577 	 * user has attempted to add the same vdev multiple times in the same
578 	 * transaction.
579 	 */
580 	if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
581 	    txg == 0 && vdtxg == crtxg)
582 		return (B_TRUE);
583 
584 	/*
585 	 * Check to see if this is a spare device.  We do an explicit check for
586 	 * spa_has_spare() here because it may be on our pending list of spares
587 	 * to add.  We also check if it is an l2cache device.
588 	 */
589 	if (spa_spare_exists(device_guid, &spare_pool, NULL) ||
590 	    spa_has_spare(spa, device_guid)) {
591 		if (spare_guid)
592 			*spare_guid = device_guid;
593 
594 		switch (reason) {
595 		case VDEV_LABEL_CREATE:
596 		case VDEV_LABEL_L2CACHE:
597 			return (B_TRUE);
598 
599 		case VDEV_LABEL_REPLACE:
600 			return (!spa_has_spare(spa, device_guid) ||
601 			    spare_pool != 0ULL);
602 
603 		case VDEV_LABEL_SPARE:
604 			return (spa_has_spare(spa, device_guid));
605 		}
606 	}
607 
608 	/*
609 	 * Check to see if this is an l2cache device.
610 	 */
611 	if (spa_l2cache_exists(device_guid, NULL))
612 		return (B_TRUE);
613 
614 	/*
615 	 * We can't rely on a pool's state if it's been imported
616 	 * read-only.  Instead we look to see if the pools is marked
617 	 * read-only in the namespace and set the state to active.
618 	 */
619 	if (state != POOL_STATE_SPARE && state != POOL_STATE_L2CACHE &&
620 	    (spa = spa_by_guid(pool_guid, device_guid)) != NULL &&
621 	    spa_mode(spa) == FREAD)
622 		state = POOL_STATE_ACTIVE;
623 
624 	/*
625 	 * If the device is marked ACTIVE, then this device is in use by another
626 	 * pool on the system.
627 	 */
628 	return (state == POOL_STATE_ACTIVE);
629 }
630 
631 /*
632  * Initialize a vdev label.  We check to make sure each leaf device is not in
633  * use, and writable.  We put down an initial label which we will later
634  * overwrite with a complete label.  Note that it's important to do this
635  * sequentially, not in parallel, so that we catch cases of multiple use of the
636  * same leaf vdev in the vdev we're creating -- e.g. mirroring a disk with
637  * itself.
638  */
639 int
vdev_label_init(vdev_t * vd,uint64_t crtxg,vdev_labeltype_t reason)640 vdev_label_init(vdev_t *vd, uint64_t crtxg, vdev_labeltype_t reason)
641 {
642 	spa_t *spa = vd->vdev_spa;
643 	nvlist_t *label;
644 	vdev_phys_t *vp;
645 	char *pad2;
646 	uberblock_t *ub;
647 	zio_t *zio;
648 	char *buf;
649 	size_t buflen;
650 	int error;
651 	uint64_t spare_guid = 0, l2cache_guid;
652 	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL;
653 
654 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
655 
656 	for (int c = 0; c < vd->vdev_children; c++)
657 		if ((error = vdev_label_init(vd->vdev_child[c],
658 		    crtxg, reason)) != 0)
659 			return (error);
660 
661 	/* Track the creation time for this vdev */
662 	vd->vdev_crtxg = crtxg;
663 
664 	if (!vd->vdev_ops->vdev_op_leaf || !spa_writeable(spa))
665 		return (0);
666 
667 	/*
668 	 * Dead vdevs cannot be initialized.
669 	 */
670 	if (vdev_is_dead(vd))
671 		return (SET_ERROR(EIO));
672 
673 	/*
674 	 * Determine if the vdev is in use.
675 	 */
676 	if (reason != VDEV_LABEL_REMOVE && reason != VDEV_LABEL_SPLIT &&
677 	    vdev_inuse(vd, crtxg, reason, &spare_guid, &l2cache_guid))
678 		return (SET_ERROR(EBUSY));
679 
680 	/*
681 	 * If this is a request to add or replace a spare or l2cache device
682 	 * that is in use elsewhere on the system, then we must update the
683 	 * guid (which was initialized to a random value) to reflect the
684 	 * actual GUID (which is shared between multiple pools).
685 	 */
686 	if (reason != VDEV_LABEL_REMOVE && reason != VDEV_LABEL_L2CACHE &&
687 	    spare_guid != 0ULL) {
688 		uint64_t guid_delta = spare_guid - vd->vdev_guid;
689 
690 		vd->vdev_guid += guid_delta;
691 
692 		for (vdev_t *pvd = vd; pvd != NULL; pvd = pvd->vdev_parent)
693 			pvd->vdev_guid_sum += guid_delta;
694 
695 		/*
696 		 * If this is a replacement, then we want to fallthrough to the
697 		 * rest of the code.  If we're adding a spare, then it's already
698 		 * labeled appropriately and we can just return.
699 		 */
700 		if (reason == VDEV_LABEL_SPARE)
701 			return (0);
702 		ASSERT(reason == VDEV_LABEL_REPLACE ||
703 		    reason == VDEV_LABEL_SPLIT);
704 	}
705 
706 	if (reason != VDEV_LABEL_REMOVE && reason != VDEV_LABEL_SPARE &&
707 	    l2cache_guid != 0ULL) {
708 		uint64_t guid_delta = l2cache_guid - vd->vdev_guid;
709 
710 		vd->vdev_guid += guid_delta;
711 
712 		for (vdev_t *pvd = vd; pvd != NULL; pvd = pvd->vdev_parent)
713 			pvd->vdev_guid_sum += guid_delta;
714 
715 		/*
716 		 * If this is a replacement, then we want to fallthrough to the
717 		 * rest of the code.  If we're adding an l2cache, then it's
718 		 * already labeled appropriately and we can just return.
719 		 */
720 		if (reason == VDEV_LABEL_L2CACHE)
721 			return (0);
722 		ASSERT(reason == VDEV_LABEL_REPLACE);
723 	}
724 
725 	/*
726 	 * TRIM the whole thing so that we start with a clean slate.
727 	 * It's just an optimization, so we don't care if it fails.
728 	 * Don't TRIM if removing so that we don't interfere with zpool
729 	 * disaster recovery.
730 	 */
731 	if (zfs_trim_enabled && vdev_trim_on_init && !vd->vdev_notrim &&
732 	    (reason == VDEV_LABEL_CREATE || reason == VDEV_LABEL_SPARE ||
733 	    reason == VDEV_LABEL_L2CACHE))
734 		zio_wait(zio_trim(NULL, spa, vd, 0, vd->vdev_psize));
735 
736 	/*
737 	 * Initialize its label.
738 	 */
739 	vp = zio_buf_alloc(sizeof (vdev_phys_t));
740 	bzero(vp, sizeof (vdev_phys_t));
741 
742 	/*
743 	 * Generate a label describing the pool and our top-level vdev.
744 	 * We mark it as being from txg 0 to indicate that it's not
745 	 * really part of an active pool just yet.  The labels will
746 	 * be written again with a meaningful txg by spa_sync().
747 	 */
748 	if (reason == VDEV_LABEL_SPARE ||
749 	    (reason == VDEV_LABEL_REMOVE && vd->vdev_isspare)) {
750 		/*
751 		 * For inactive hot spares, we generate a special label that
752 		 * identifies as a mutually shared hot spare.  We write the
753 		 * label if we are adding a hot spare, or if we are removing an
754 		 * active hot spare (in which case we want to revert the
755 		 * labels).
756 		 */
757 		VERIFY(nvlist_alloc(&label, NV_UNIQUE_NAME, KM_SLEEP) == 0);
758 
759 		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_VERSION,
760 		    spa_version(spa)) == 0);
761 		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_POOL_STATE,
762 		    POOL_STATE_SPARE) == 0);
763 		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_GUID,
764 		    vd->vdev_guid) == 0);
765 	} else if (reason == VDEV_LABEL_L2CACHE ||
766 	    (reason == VDEV_LABEL_REMOVE && vd->vdev_isl2cache)) {
767 		/*
768 		 * For level 2 ARC devices, add a special label.
769 		 */
770 		VERIFY(nvlist_alloc(&label, NV_UNIQUE_NAME, KM_SLEEP) == 0);
771 
772 		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_VERSION,
773 		    spa_version(spa)) == 0);
774 		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_POOL_STATE,
775 		    POOL_STATE_L2CACHE) == 0);
776 		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_GUID,
777 		    vd->vdev_guid) == 0);
778 	} else {
779 		uint64_t txg = 0ULL;
780 
781 		if (reason == VDEV_LABEL_SPLIT)
782 			txg = spa->spa_uberblock.ub_txg;
783 		label = spa_config_generate(spa, vd, txg, B_FALSE);
784 
785 		/*
786 		 * Add our creation time.  This allows us to detect multiple
787 		 * vdev uses as described above, and automatically expires if we
788 		 * fail.
789 		 */
790 		VERIFY(nvlist_add_uint64(label, ZPOOL_CONFIG_CREATE_TXG,
791 		    crtxg) == 0);
792 	}
793 
794 	buf = vp->vp_nvlist;
795 	buflen = sizeof (vp->vp_nvlist);
796 
797 	error = nvlist_pack(label, &buf, &buflen, NV_ENCODE_XDR, KM_SLEEP);
798 	if (error != 0) {
799 		nvlist_free(label);
800 		zio_buf_free(vp, sizeof (vdev_phys_t));
801 		/* EFAULT means nvlist_pack ran out of room */
802 		return (error == EFAULT ? ENAMETOOLONG : EINVAL);
803 	}
804 
805 	/*
806 	 * Initialize uberblock template.
807 	 */
808 	ub = zio_buf_alloc(VDEV_UBERBLOCK_RING);
809 	bzero(ub, VDEV_UBERBLOCK_RING);
810 	*ub = spa->spa_uberblock;
811 	ub->ub_txg = 0;
812 
813 	/* Initialize the 2nd padding area. */
814 	pad2 = zio_buf_alloc(VDEV_PAD_SIZE);
815 	bzero(pad2, VDEV_PAD_SIZE);
816 
817 	/*
818 	 * Write everything in parallel.
819 	 */
820 retry:
821 	zio = zio_root(spa, NULL, NULL, flags);
822 
823 	for (int l = 0; l < VDEV_LABELS; l++) {
824 
825 		vdev_label_write(zio, vd, l, vp,
826 		    offsetof(vdev_label_t, vl_vdev_phys),
827 		    sizeof (vdev_phys_t), NULL, NULL, flags);
828 
829 		/*
830 		 * Skip the 1st padding area.
831 		 * Zero out the 2nd padding area where it might have
832 		 * left over data from previous filesystem format.
833 		 */
834 		vdev_label_write(zio, vd, l, pad2,
835 		    offsetof(vdev_label_t, vl_pad2),
836 		    VDEV_PAD_SIZE, NULL, NULL, flags);
837 
838 		vdev_label_write(zio, vd, l, ub,
839 		    offsetof(vdev_label_t, vl_uberblock),
840 		    VDEV_UBERBLOCK_RING, NULL, NULL, flags);
841 	}
842 
843 	error = zio_wait(zio);
844 
845 	if (error != 0 && !(flags & ZIO_FLAG_TRYHARD)) {
846 		flags |= ZIO_FLAG_TRYHARD;
847 		goto retry;
848 	}
849 
850 	nvlist_free(label);
851 	zio_buf_free(pad2, VDEV_PAD_SIZE);
852 	zio_buf_free(ub, VDEV_UBERBLOCK_RING);
853 	zio_buf_free(vp, sizeof (vdev_phys_t));
854 
855 	/*
856 	 * If this vdev hasn't been previously identified as a spare, then we
857 	 * mark it as such only if a) we are labeling it as a spare, or b) it
858 	 * exists as a spare elsewhere in the system.  Do the same for
859 	 * level 2 ARC devices.
860 	 */
861 	if (error == 0 && !vd->vdev_isspare &&
862 	    (reason == VDEV_LABEL_SPARE ||
863 	    spa_spare_exists(vd->vdev_guid, NULL, NULL)))
864 		spa_spare_add(vd);
865 
866 	if (error == 0 && !vd->vdev_isl2cache &&
867 	    (reason == VDEV_LABEL_L2CACHE ||
868 	    spa_l2cache_exists(vd->vdev_guid, NULL)))
869 		spa_l2cache_add(vd);
870 
871 	return (error);
872 }
873 
874 int
vdev_label_write_pad2(vdev_t * vd,const char * buf,size_t size)875 vdev_label_write_pad2(vdev_t *vd, const char *buf, size_t size)
876 {
877 	spa_t *spa = vd->vdev_spa;
878 	zio_t *zio;
879 	char *pad2;
880 	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL;
881 	int error;
882 
883 	if (size > VDEV_PAD_SIZE)
884 		return (EINVAL);
885 
886 	if (!vd->vdev_ops->vdev_op_leaf)
887 		return (ENODEV);
888 	if (vdev_is_dead(vd))
889 		return (ENXIO);
890 
891 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
892 
893 	pad2 = zio_buf_alloc(VDEV_PAD_SIZE);
894 	bzero(pad2, VDEV_PAD_SIZE);
895 	memcpy(pad2, buf, size);
896 
897 retry:
898 	zio = zio_root(spa, NULL, NULL, flags);
899 	vdev_label_write(zio, vd, 0, pad2,
900 	    offsetof(vdev_label_t, vl_pad2),
901 	    VDEV_PAD_SIZE, NULL, NULL, flags);
902 	error = zio_wait(zio);
903 	if (error != 0 && !(flags & ZIO_FLAG_TRYHARD)) {
904 		flags |= ZIO_FLAG_TRYHARD;
905 		goto retry;
906 	}
907 
908 	zio_buf_free(pad2, VDEV_PAD_SIZE);
909 	return (error);
910 }
911 
912 /*
913  * ==========================================================================
914  * uberblock load/sync
915  * ==========================================================================
916  */
917 
918 /*
919  * Consider the following situation: txg is safely synced to disk.  We've
920  * written the first uberblock for txg + 1, and then we lose power.  When we
921  * come back up, we fail to see the uberblock for txg + 1 because, say,
922  * it was on a mirrored device and the replica to which we wrote txg + 1
923  * is now offline.  If we then make some changes and sync txg + 1, and then
924  * the missing replica comes back, then for a few seconds we'll have two
925  * conflicting uberblocks on disk with the same txg.  The solution is simple:
926  * among uberblocks with equal txg, choose the one with the latest timestamp.
927  */
928 static int
vdev_uberblock_compare(uberblock_t * ub1,uberblock_t * ub2)929 vdev_uberblock_compare(uberblock_t *ub1, uberblock_t *ub2)
930 {
931 	if (ub1->ub_txg < ub2->ub_txg)
932 		return (-1);
933 	if (ub1->ub_txg > ub2->ub_txg)
934 		return (1);
935 
936 	if (ub1->ub_timestamp < ub2->ub_timestamp)
937 		return (-1);
938 	if (ub1->ub_timestamp > ub2->ub_timestamp)
939 		return (1);
940 
941 	return (0);
942 }
943 
944 struct ubl_cbdata {
945 	uberblock_t	*ubl_ubbest;	/* Best uberblock */
946 	vdev_t		*ubl_vd;	/* vdev associated with the above */
947 };
948 
949 static void
vdev_uberblock_load_done(zio_t * zio)950 vdev_uberblock_load_done(zio_t *zio)
951 {
952 	vdev_t *vd = zio->io_vd;
953 	spa_t *spa = zio->io_spa;
954 	zio_t *rio = zio->io_private;
955 	uberblock_t *ub = zio->io_data;
956 	struct ubl_cbdata *cbp = rio->io_private;
957 
958 	ASSERT3U(zio->io_size, ==, VDEV_UBERBLOCK_SIZE(vd));
959 
960 	if (zio->io_error == 0 && uberblock_verify(ub) == 0) {
961 		mutex_enter(&rio->io_lock);
962 		if (ub->ub_txg <= spa->spa_load_max_txg &&
963 		    vdev_uberblock_compare(ub, cbp->ubl_ubbest) > 0) {
964 			/*
965 			 * Keep track of the vdev in which this uberblock
966 			 * was found. We will use this information later
967 			 * to obtain the config nvlist associated with
968 			 * this uberblock.
969 			 */
970 			*cbp->ubl_ubbest = *ub;
971 			cbp->ubl_vd = vd;
972 		}
973 		mutex_exit(&rio->io_lock);
974 	}
975 
976 	zio_buf_free(zio->io_data, zio->io_size);
977 }
978 
979 static void
vdev_uberblock_load_impl(zio_t * zio,vdev_t * vd,int flags,struct ubl_cbdata * cbp)980 vdev_uberblock_load_impl(zio_t *zio, vdev_t *vd, int flags,
981     struct ubl_cbdata *cbp)
982 {
983 	for (int c = 0; c < vd->vdev_children; c++)
984 		vdev_uberblock_load_impl(zio, vd->vdev_child[c], flags, cbp);
985 
986 	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
987 		for (int l = 0; l < VDEV_LABELS; l++) {
988 			for (int n = 0; n < VDEV_UBERBLOCK_COUNT(vd); n++) {
989 				vdev_label_read(zio, vd, l,
990 				    zio_buf_alloc(VDEV_UBERBLOCK_SIZE(vd)),
991 				    VDEV_UBERBLOCK_OFFSET(vd, n),
992 				    VDEV_UBERBLOCK_SIZE(vd),
993 				    vdev_uberblock_load_done, zio, flags);
994 			}
995 		}
996 	}
997 }
998 
999 /*
1000  * Reads the 'best' uberblock from disk along with its associated
1001  * configuration. First, we read the uberblock array of each label of each
1002  * vdev, keeping track of the uberblock with the highest txg in each array.
1003  * Then, we read the configuration from the same vdev as the best uberblock.
1004  */
1005 void
vdev_uberblock_load(vdev_t * rvd,uberblock_t * ub,nvlist_t ** config)1006 vdev_uberblock_load(vdev_t *rvd, uberblock_t *ub, nvlist_t **config)
1007 {
1008 	zio_t *zio;
1009 	spa_t *spa = rvd->vdev_spa;
1010 	struct ubl_cbdata cb;
1011 	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL |
1012 	    ZIO_FLAG_SPECULATIVE | ZIO_FLAG_TRYHARD;
1013 
1014 	ASSERT(ub);
1015 	ASSERT(config);
1016 
1017 	bzero(ub, sizeof (uberblock_t));
1018 	*config = NULL;
1019 
1020 	cb.ubl_ubbest = ub;
1021 	cb.ubl_vd = NULL;
1022 
1023 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
1024 	zio = zio_root(spa, NULL, &cb, flags);
1025 	vdev_uberblock_load_impl(zio, rvd, flags, &cb);
1026 	(void) zio_wait(zio);
1027 
1028 	/*
1029 	 * It's possible that the best uberblock was discovered on a label
1030 	 * that has a configuration which was written in a future txg.
1031 	 * Search all labels on this vdev to find the configuration that
1032 	 * matches the txg for our uberblock.
1033 	 */
1034 	if (cb.ubl_vd != NULL)
1035 		*config = vdev_label_read_config(cb.ubl_vd, ub->ub_txg);
1036 	spa_config_exit(spa, SCL_ALL, FTAG);
1037 }
1038 
1039 /*
1040  * On success, increment root zio's count of good writes.
1041  * We only get credit for writes to known-visible vdevs; see spa_vdev_add().
1042  */
1043 static void
vdev_uberblock_sync_done(zio_t * zio)1044 vdev_uberblock_sync_done(zio_t *zio)
1045 {
1046 	uint64_t *good_writes = zio->io_private;
1047 
1048 	if (zio->io_error == 0 && zio->io_vd->vdev_top->vdev_ms_array != 0)
1049 		atomic_inc_64(good_writes);
1050 }
1051 
1052 /*
1053  * Write the uberblock to all labels of all leaves of the specified vdev.
1054  */
1055 static void
vdev_uberblock_sync(zio_t * zio,uberblock_t * ub,vdev_t * vd,int flags)1056 vdev_uberblock_sync(zio_t *zio, uberblock_t *ub, vdev_t *vd, int flags)
1057 {
1058 	uberblock_t *ubbuf;
1059 	int n;
1060 
1061 	for (int c = 0; c < vd->vdev_children; c++)
1062 		vdev_uberblock_sync(zio, ub, vd->vdev_child[c], flags);
1063 
1064 	if (!vd->vdev_ops->vdev_op_leaf)
1065 		return;
1066 
1067 	if (!vdev_writeable(vd))
1068 		return;
1069 
1070 	n = ub->ub_txg & (VDEV_UBERBLOCK_COUNT(vd) - 1);
1071 
1072 	ubbuf = zio_buf_alloc(VDEV_UBERBLOCK_SIZE(vd));
1073 	bzero(ubbuf, VDEV_UBERBLOCK_SIZE(vd));
1074 	*ubbuf = *ub;
1075 
1076 	for (int l = 0; l < VDEV_LABELS; l++)
1077 		vdev_label_write(zio, vd, l, ubbuf,
1078 		    VDEV_UBERBLOCK_OFFSET(vd, n), VDEV_UBERBLOCK_SIZE(vd),
1079 		    vdev_uberblock_sync_done, zio->io_private,
1080 		    flags | ZIO_FLAG_DONT_PROPAGATE);
1081 
1082 	zio_buf_free(ubbuf, VDEV_UBERBLOCK_SIZE(vd));
1083 }
1084 
1085 /* Sync the uberblocks to all vdevs in svd[] */
1086 int
vdev_uberblock_sync_list(vdev_t ** svd,int svdcount,uberblock_t * ub,int flags)1087 vdev_uberblock_sync_list(vdev_t **svd, int svdcount, uberblock_t *ub, int flags)
1088 {
1089 	spa_t *spa = svd[0]->vdev_spa;
1090 	zio_t *zio;
1091 	uint64_t good_writes = 0;
1092 
1093 	zio = zio_root(spa, NULL, &good_writes, flags);
1094 
1095 	for (int v = 0; v < svdcount; v++)
1096 		vdev_uberblock_sync(zio, ub, svd[v], flags);
1097 
1098 	(void) zio_wait(zio);
1099 
1100 	/*
1101 	 * Flush the uberblocks to disk.  This ensures that the odd labels
1102 	 * are no longer needed (because the new uberblocks and the even
1103 	 * labels are safely on disk), so it is safe to overwrite them.
1104 	 */
1105 	zio = zio_root(spa, NULL, NULL, flags);
1106 
1107 	for (int v = 0; v < svdcount; v++)
1108 		zio_flush(zio, svd[v]);
1109 
1110 	(void) zio_wait(zio);
1111 
1112 	return (good_writes >= 1 ? 0 : EIO);
1113 }
1114 
1115 /*
1116  * On success, increment the count of good writes for our top-level vdev.
1117  */
1118 static void
vdev_label_sync_done(zio_t * zio)1119 vdev_label_sync_done(zio_t *zio)
1120 {
1121 	uint64_t *good_writes = zio->io_private;
1122 
1123 	if (zio->io_error == 0)
1124 		atomic_inc_64(good_writes);
1125 }
1126 
1127 /*
1128  * If there weren't enough good writes, indicate failure to the parent.
1129  */
1130 static void
vdev_label_sync_top_done(zio_t * zio)1131 vdev_label_sync_top_done(zio_t *zio)
1132 {
1133 	uint64_t *good_writes = zio->io_private;
1134 
1135 	if (*good_writes == 0)
1136 		zio->io_error = SET_ERROR(EIO);
1137 
1138 	kmem_free(good_writes, sizeof (uint64_t));
1139 }
1140 
1141 /*
1142  * We ignore errors for log and cache devices, simply free the private data.
1143  */
1144 static void
vdev_label_sync_ignore_done(zio_t * zio)1145 vdev_label_sync_ignore_done(zio_t *zio)
1146 {
1147 	kmem_free(zio->io_private, sizeof (uint64_t));
1148 }
1149 
1150 /*
1151  * Write all even or odd labels to all leaves of the specified vdev.
1152  */
1153 static void
vdev_label_sync(zio_t * zio,vdev_t * vd,int l,uint64_t txg,int flags)1154 vdev_label_sync(zio_t *zio, vdev_t *vd, int l, uint64_t txg, int flags)
1155 {
1156 	nvlist_t *label;
1157 	vdev_phys_t *vp;
1158 	char *buf;
1159 	size_t buflen;
1160 
1161 	for (int c = 0; c < vd->vdev_children; c++)
1162 		vdev_label_sync(zio, vd->vdev_child[c], l, txg, flags);
1163 
1164 	if (!vd->vdev_ops->vdev_op_leaf)
1165 		return;
1166 
1167 	if (!vdev_writeable(vd))
1168 		return;
1169 
1170 	/*
1171 	 * Generate a label describing the top-level config to which we belong.
1172 	 */
1173 	label = spa_config_generate(vd->vdev_spa, vd, txg, B_FALSE);
1174 
1175 	vp = zio_buf_alloc(sizeof (vdev_phys_t));
1176 	bzero(vp, sizeof (vdev_phys_t));
1177 
1178 	buf = vp->vp_nvlist;
1179 	buflen = sizeof (vp->vp_nvlist);
1180 
1181 	if (nvlist_pack(label, &buf, &buflen, NV_ENCODE_XDR, KM_SLEEP) == 0) {
1182 		for (; l < VDEV_LABELS; l += 2) {
1183 			vdev_label_write(zio, vd, l, vp,
1184 			    offsetof(vdev_label_t, vl_vdev_phys),
1185 			    sizeof (vdev_phys_t),
1186 			    vdev_label_sync_done, zio->io_private,
1187 			    flags | ZIO_FLAG_DONT_PROPAGATE);
1188 		}
1189 	}
1190 
1191 	zio_buf_free(vp, sizeof (vdev_phys_t));
1192 	nvlist_free(label);
1193 }
1194 
1195 int
vdev_label_sync_list(spa_t * spa,int l,uint64_t txg,int flags)1196 vdev_label_sync_list(spa_t *spa, int l, uint64_t txg, int flags)
1197 {
1198 	list_t *dl = &spa->spa_config_dirty_list;
1199 	vdev_t *vd;
1200 	zio_t *zio;
1201 	int error;
1202 
1203 	/*
1204 	 * Write the new labels to disk.
1205 	 */
1206 	zio = zio_root(spa, NULL, NULL, flags);
1207 
1208 	for (vd = list_head(dl); vd != NULL; vd = list_next(dl, vd)) {
1209 		uint64_t *good_writes = kmem_zalloc(sizeof (uint64_t),
1210 		    KM_SLEEP);
1211 
1212 		ASSERT(!vd->vdev_ishole);
1213 
1214 		zio_t *vio = zio_null(zio, spa, NULL,
1215 		    (vd->vdev_islog || vd->vdev_aux != NULL) ?
1216 		    vdev_label_sync_ignore_done : vdev_label_sync_top_done,
1217 		    good_writes, flags);
1218 		vdev_label_sync(vio, vd, l, txg, flags);
1219 		zio_nowait(vio);
1220 	}
1221 
1222 	error = zio_wait(zio);
1223 
1224 	/*
1225 	 * Flush the new labels to disk.
1226 	 */
1227 	zio = zio_root(spa, NULL, NULL, flags);
1228 
1229 	for (vd = list_head(dl); vd != NULL; vd = list_next(dl, vd))
1230 		zio_flush(zio, vd);
1231 
1232 	(void) zio_wait(zio);
1233 
1234 	return (error);
1235 }
1236 
1237 /*
1238  * Sync the uberblock and any changes to the vdev configuration.
1239  *
1240  * The order of operations is carefully crafted to ensure that
1241  * if the system panics or loses power at any time, the state on disk
1242  * is still transactionally consistent.  The in-line comments below
1243  * describe the failure semantics at each stage.
1244  *
1245  * Moreover, vdev_config_sync() is designed to be idempotent: if it fails
1246  * at any time, you can just call it again, and it will resume its work.
1247  */
1248 int
vdev_config_sync(vdev_t ** svd,int svdcount,uint64_t txg)1249 vdev_config_sync(vdev_t **svd, int svdcount, uint64_t txg)
1250 {
1251 	spa_t *spa = svd[0]->vdev_spa;
1252 	uberblock_t *ub = &spa->spa_uberblock;
1253 	vdev_t *vd;
1254 	zio_t *zio;
1255 	int error = 0;
1256 	int flags = ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_CANFAIL;
1257 
1258 retry:
1259 	/*
1260 	 * Normally, we don't want to try too hard to write every label and
1261 	 * uberblock.  If there is a flaky disk, we don't want the rest of the
1262 	 * sync process to block while we retry.  But if we can't write a
1263 	 * single label out, we should retry with ZIO_FLAG_TRYHARD before
1264 	 * bailing out and declaring the pool faulted.
1265 	 */
1266 	if (error != 0) {
1267 		if ((flags & ZIO_FLAG_TRYHARD) != 0)
1268 			return (error);
1269 		flags |= ZIO_FLAG_TRYHARD;
1270 	}
1271 
1272 	ASSERT(ub->ub_txg <= txg);
1273 
1274 	/*
1275 	 * If this isn't a resync due to I/O errors,
1276 	 * and nothing changed in this transaction group,
1277 	 * and the vdev configuration hasn't changed,
1278 	 * then there's nothing to do.
1279 	 */
1280 	if (ub->ub_txg < txg &&
1281 	    uberblock_update(ub, spa->spa_root_vdev, txg) == B_FALSE &&
1282 	    list_is_empty(&spa->spa_config_dirty_list))
1283 		return (0);
1284 
1285 	if (txg > spa_freeze_txg(spa))
1286 		return (0);
1287 
1288 	ASSERT(txg <= spa->spa_final_txg);
1289 
1290 	/*
1291 	 * Flush the write cache of every disk that's been written to
1292 	 * in this transaction group.  This ensures that all blocks
1293 	 * written in this txg will be committed to stable storage
1294 	 * before any uberblock that references them.
1295 	 */
1296 	zio = zio_root(spa, NULL, NULL, flags);
1297 
1298 	for (vd = txg_list_head(&spa->spa_vdev_txg_list, TXG_CLEAN(txg)); vd;
1299 	    vd = txg_list_next(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg)))
1300 		zio_flush(zio, vd);
1301 
1302 	(void) zio_wait(zio);
1303 
1304 	/*
1305 	 * Sync out the even labels (L0, L2) for every dirty vdev.  If the
1306 	 * system dies in the middle of this process, that's OK: all of the
1307 	 * even labels that made it to disk will be newer than any uberblock,
1308 	 * and will therefore be considered invalid.  The odd labels (L1, L3),
1309 	 * which have not yet been touched, will still be valid.  We flush
1310 	 * the new labels to disk to ensure that all even-label updates
1311 	 * are committed to stable storage before the uberblock update.
1312 	 */
1313 	if ((error = vdev_label_sync_list(spa, 0, txg, flags)) != 0)
1314 		goto retry;
1315 
1316 	/*
1317 	 * Sync the uberblocks to all vdevs in svd[].
1318 	 * If the system dies in the middle of this step, there are two cases
1319 	 * to consider, and the on-disk state is consistent either way:
1320 	 *
1321 	 * (1)	If none of the new uberblocks made it to disk, then the
1322 	 *	previous uberblock will be the newest, and the odd labels
1323 	 *	(which had not yet been touched) will be valid with respect
1324 	 *	to that uberblock.
1325 	 *
1326 	 * (2)	If one or more new uberblocks made it to disk, then they
1327 	 *	will be the newest, and the even labels (which had all
1328 	 *	been successfully committed) will be valid with respect
1329 	 *	to the new uberblocks.
1330 	 */
1331 	if ((error = vdev_uberblock_sync_list(svd, svdcount, ub, flags)) != 0)
1332 		goto retry;
1333 
1334 	/*
1335 	 * Sync out odd labels for every dirty vdev.  If the system dies
1336 	 * in the middle of this process, the even labels and the new
1337 	 * uberblocks will suffice to open the pool.  The next time
1338 	 * the pool is opened, the first thing we'll do -- before any
1339 	 * user data is modified -- is mark every vdev dirty so that
1340 	 * all labels will be brought up to date.  We flush the new labels
1341 	 * to disk to ensure that all odd-label updates are committed to
1342 	 * stable storage before the next transaction group begins.
1343 	 */
1344 	if ((error = vdev_label_sync_list(spa, 1, txg, flags)) != 0)
1345 		goto retry;;
1346 
1347 	trim_thread_wakeup(spa);
1348 
1349 	return (0);
1350 }
1351