xref: /freebsd/sys/contrib/openzfs/module/zfs/vdev_mirror.c (revision 2f10ffc003be396f3fc23cd2888023896560252b)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * This file and its contents are supplied under the terms of the
4  * Common Development and Distribution License ("CDDL"), version 1.0.
5  * You may only use this file in accordance with the terms of version
6  * 1.0 of the CDDL.
7  *
8  * A full copy of the text of the CDDL should have accompanied this
9  * source.  A copy of the CDDL is also available via the Internet at
10  * https://opensource.org/license/CDDL-1.0.
11  */
12 /*
13  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
14  * Use is subject to license terms.
15  */
16 
17 /*
18  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
19  */
20 
21 #include <sys/zfs_context.h>
22 #include <sys/spa.h>
23 #include <sys/spa_impl.h>
24 #include <sys/dsl_pool.h>
25 #include <sys/dsl_scan.h>
26 #include <sys/vdev_impl.h>
27 #include <sys/vdev_draid.h>
28 #include <sys/zio.h>
29 #include <sys/zio_checksum.h>
30 #include <sys/abd.h>
31 #include <sys/fs/zfs.h>
32 
33 /*
34  * Vdev mirror kstats
35  */
36 static kstat_t *mirror_ksp = NULL;
37 
38 typedef struct mirror_stats {
39 	kstat_named_t vdev_mirror_stat_rotating_linear;
40 	kstat_named_t vdev_mirror_stat_rotating_offset;
41 	kstat_named_t vdev_mirror_stat_rotating_seek;
42 	kstat_named_t vdev_mirror_stat_non_rotating_linear;
43 	kstat_named_t vdev_mirror_stat_non_rotating_seek;
44 
45 	kstat_named_t vdev_mirror_stat_preferred_found;
46 	kstat_named_t vdev_mirror_stat_preferred_not_found;
47 } mirror_stats_t;
48 
49 static mirror_stats_t mirror_stats = {
50 	/* New I/O follows directly the last I/O */
51 	{ "rotating_linear",			KSTAT_DATA_UINT64 },
52 	/* New I/O is within zfs_vdev_mirror_rotating_seek_offset of the last */
53 	{ "rotating_offset",			KSTAT_DATA_UINT64 },
54 	/* New I/O requires random seek */
55 	{ "rotating_seek",			KSTAT_DATA_UINT64 },
56 	/* New I/O follows directly the last I/O  (nonrot) */
57 	{ "non_rotating_linear",		KSTAT_DATA_UINT64 },
58 	/* New I/O requires random seek (nonrot) */
59 	{ "non_rotating_seek",			KSTAT_DATA_UINT64 },
60 	/* Preferred child vdev found */
61 	{ "preferred_found",			KSTAT_DATA_UINT64 },
62 	/* Preferred child vdev not found or equal load  */
63 	{ "preferred_not_found",		KSTAT_DATA_UINT64 },
64 
65 };
66 
67 #define	MIRROR_STAT(stat)		(mirror_stats.stat.value.ui64)
68 #define	MIRROR_INCR(stat, val) 		atomic_add_64(&MIRROR_STAT(stat), val)
69 #define	MIRROR_BUMP(stat)		MIRROR_INCR(stat, 1)
70 
71 void
72 vdev_mirror_stat_init(void)
73 {
74 	mirror_ksp = kstat_create("zfs", 0, "vdev_mirror_stats",
75 	    "misc", KSTAT_TYPE_NAMED,
76 	    sizeof (mirror_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
77 	if (mirror_ksp != NULL) {
78 		mirror_ksp->ks_data = &mirror_stats;
79 		kstat_install(mirror_ksp);
80 	}
81 }
82 
83 void
84 vdev_mirror_stat_fini(void)
85 {
86 	if (mirror_ksp != NULL) {
87 		kstat_delete(mirror_ksp);
88 		mirror_ksp = NULL;
89 	}
90 }
91 
92 /*
93  * Virtual device vector for mirroring.
94  */
95 typedef struct mirror_child {
96 	vdev_t		*mc_vd;
97 	abd_t		*mc_abd;
98 	uint64_t	mc_offset;
99 	int		mc_error;
100 	int		mc_load;
101 	uint8_t		mc_tried;
102 	uint8_t		mc_skipped;
103 	uint8_t		mc_speculative;
104 	uint8_t		mc_rebuilding;
105 } mirror_child_t;
106 
107 typedef struct mirror_map {
108 	int		*mm_preferred;
109 	int		mm_preferred_cnt;
110 	int		mm_children;
111 	boolean_t	mm_resilvering;
112 	boolean_t	mm_rebuilding;
113 	boolean_t	mm_root;
114 	mirror_child_t	mm_child[];
115 } mirror_map_t;
116 
117 static const int vdev_mirror_shift = 21;
118 
119 /*
120  * The load configuration settings below are tuned by default for
121  * the case where all devices are of the same rotational type.
122  *
123  * If there is a mixture of rotating and non-rotating media, setting
124  * zfs_vdev_mirror_non_rotating_seek_inc to 0 may well provide better results
125  * as it will direct more reads to the non-rotating vdevs which are more likely
126  * to have a higher performance.
127  */
128 
129 /* Rotating media load calculation configuration. */
130 static int zfs_vdev_mirror_rotating_inc = 0;
131 static int zfs_vdev_mirror_rotating_seek_inc = 5;
132 static int zfs_vdev_mirror_rotating_seek_offset = 1 * 1024 * 1024;
133 
134 /* Non-rotating media load calculation configuration. */
135 static int zfs_vdev_mirror_non_rotating_inc = 0;
136 static int zfs_vdev_mirror_non_rotating_seek_inc = 1;
137 
138 static inline size_t
139 vdev_mirror_map_size(int children)
140 {
141 	return (offsetof(mirror_map_t, mm_child[children]) +
142 	    sizeof (int) * children);
143 }
144 
145 static inline mirror_map_t *
146 vdev_mirror_map_alloc(int children, boolean_t resilvering, boolean_t root)
147 {
148 	mirror_map_t *mm;
149 
150 	mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP);
151 	mm->mm_children = children;
152 	mm->mm_resilvering = resilvering;
153 	mm->mm_root = root;
154 	mm->mm_preferred = (int *)((uintptr_t)mm +
155 	    offsetof(mirror_map_t, mm_child[children]));
156 
157 	return (mm);
158 }
159 
160 static void
161 vdev_mirror_map_free(zio_t *zio)
162 {
163 	mirror_map_t *mm = zio->io_vsd;
164 
165 	kmem_free(mm, vdev_mirror_map_size(mm->mm_children));
166 }
167 
168 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
169 	.vsd_free = vdev_mirror_map_free,
170 };
171 
172 static int
173 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
174 {
175 	uint64_t last_offset;
176 	int64_t offset_diff;
177 	int load;
178 
179 	/* All DVAs have equal weight at the root. */
180 	if (mm->mm_root)
181 		return (INT_MAX);
182 
183 	/*
184 	 * We don't return INT_MAX if the device is resilvering i.e.
185 	 * vdev_resilver_txg != 0 as when tested performance was slightly
186 	 * worse overall when resilvering with compared to without.
187 	 */
188 
189 	/* Fix zio_offset for leaf vdevs */
190 	if (vd->vdev_ops->vdev_op_leaf)
191 		zio_offset += VDEV_LABEL_START_SIZE;
192 
193 	/* Standard load based on pending queue length. */
194 	load = vdev_queue_length(vd);
195 	last_offset = vdev_queue_last_offset(vd);
196 
197 	if (vd->vdev_nonrot) {
198 		/* Non-rotating media. */
199 		if (last_offset == zio_offset) {
200 			MIRROR_BUMP(vdev_mirror_stat_non_rotating_linear);
201 			return (load + zfs_vdev_mirror_non_rotating_inc);
202 		}
203 
204 		/*
205 		 * Apply a seek penalty even for non-rotating devices as
206 		 * sequential I/O's can be aggregated into fewer operations on
207 		 * the device, thus avoiding unnecessary per-command overhead
208 		 * and boosting performance.
209 		 */
210 		MIRROR_BUMP(vdev_mirror_stat_non_rotating_seek);
211 		return (load + zfs_vdev_mirror_non_rotating_seek_inc);
212 	}
213 
214 	/* Rotating media I/O's which directly follow the last I/O. */
215 	if (last_offset == zio_offset) {
216 		MIRROR_BUMP(vdev_mirror_stat_rotating_linear);
217 		return (load + zfs_vdev_mirror_rotating_inc);
218 	}
219 
220 	/*
221 	 * Apply half the seek increment to I/O's within seek offset
222 	 * of the last I/O issued to this vdev as they should incur less
223 	 * of a seek increment.
224 	 */
225 	offset_diff = (int64_t)(last_offset - zio_offset);
226 	if (ABS(offset_diff) < zfs_vdev_mirror_rotating_seek_offset) {
227 		MIRROR_BUMP(vdev_mirror_stat_rotating_offset);
228 		return (load + (zfs_vdev_mirror_rotating_seek_inc / 2));
229 	}
230 
231 	/* Apply the full seek increment to all other I/O's. */
232 	MIRROR_BUMP(vdev_mirror_stat_rotating_seek);
233 	return (load + zfs_vdev_mirror_rotating_seek_inc);
234 }
235 
236 static boolean_t
237 vdev_mirror_rebuilding(vdev_t *vd)
238 {
239 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_rebuild_txg)
240 		return (B_TRUE);
241 
242 	for (int i = 0; i < vd->vdev_children; i++) {
243 		if (vdev_mirror_rebuilding(vd->vdev_child[i])) {
244 			return (B_TRUE);
245 		}
246 	}
247 
248 	return (B_FALSE);
249 }
250 
251 /*
252  * Avoid inlining the function to keep vdev_mirror_io_start(), which
253  * is this functions only caller, as small as possible on the stack.
254  */
255 noinline static mirror_map_t *
256 vdev_mirror_map_init(zio_t *zio)
257 {
258 	mirror_map_t *mm = NULL;
259 	mirror_child_t *mc;
260 	vdev_t *vd = zio->io_vd;
261 	int c;
262 
263 	if (vd == NULL) {
264 		dva_t *dva = zio->io_bp->blk_dva;
265 		spa_t *spa = zio->io_spa;
266 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
267 		dva_t dva_copy[SPA_DVAS_PER_BP];
268 
269 		/*
270 		 * The sequential scrub code sorts and issues all DVAs
271 		 * of a bp separately. Each of these IOs includes all
272 		 * original DVA copies so that repairs can be performed
273 		 * in the event of an error, but we only actually want
274 		 * to check the first DVA since the others will be
275 		 * checked by their respective sorted IOs. Only if we
276 		 * hit an error will we try all DVAs upon retrying.
277 		 *
278 		 * Note: This check is safe even if the user switches
279 		 * from a legacy scrub to a sequential one in the middle
280 		 * of processing, since scn_is_sorted isn't updated until
281 		 * all outstanding IOs from the previous scrub pass
282 		 * complete.
283 		 */
284 		if ((zio->io_flags & ZIO_FLAG_SCRUB) &&
285 		    !(zio->io_flags & ZIO_FLAG_IO_RETRY) &&
286 		    dsl_scan_scrubbing(spa->spa_dsl_pool) &&
287 		    scn->scn_is_sorted) {
288 			c = 1;
289 		} else {
290 			c = BP_GET_NDVAS(zio->io_bp);
291 		}
292 
293 		/*
294 		 * If the pool cannot be written to, then infer that some
295 		 * DVAs might be invalid or point to vdevs that do not exist.
296 		 * We skip them.
297 		 */
298 		if (!spa_writeable(spa)) {
299 			ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
300 			int j = 0;
301 			for (int i = 0; i < c; i++) {
302 				if (zfs_dva_valid(spa, &dva[i], zio->io_bp))
303 					dva_copy[j++] = dva[i];
304 			}
305 			if (j == 0) {
306 				zio->io_vsd = NULL;
307 				zio->io_error = ENXIO;
308 				return (NULL);
309 			}
310 			if (j < c) {
311 				dva = dva_copy;
312 				c = j;
313 			}
314 		}
315 
316 		mm = vdev_mirror_map_alloc(c, B_FALSE, B_TRUE);
317 		for (c = 0; c < mm->mm_children; c++) {
318 			mc = &mm->mm_child[c];
319 
320 			mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
321 			mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
322 			if (mc->mc_vd == NULL) {
323 				kmem_free(mm, vdev_mirror_map_size(
324 				    mm->mm_children));
325 				zio->io_vsd = NULL;
326 				zio->io_error = ENXIO;
327 				return (NULL);
328 			}
329 		}
330 	} else {
331 		/*
332 		 * If we are resilvering, then we should handle scrub reads
333 		 * differently; we shouldn't issue them to the resilvering
334 		 * device because it might not have those blocks.
335 		 *
336 		 * We are resilvering iff:
337 		 * 1) We are a replacing vdev (ie our name is "replacing-1" or
338 		 *    "spare-1" or something like that), and
339 		 * 2) The pool is currently being resilvered.
340 		 *
341 		 * We cannot simply check vd->vdev_resilver_txg, because it's
342 		 * not set in this path.
343 		 *
344 		 * Nor can we just check our vdev_ops; there are cases (such as
345 		 * when a user types "zpool replace pool odev spare_dev" and
346 		 * spare_dev is in the spare list, or when a spare device is
347 		 * automatically used to replace a DEGRADED device) when
348 		 * resilvering is complete but both the original vdev and the
349 		 * spare vdev remain in the pool.  That behavior is intentional.
350 		 * It helps implement the policy that a spare should be
351 		 * automatically removed from the pool after the user replaces
352 		 * the device that originally failed.
353 		 *
354 		 * If a spa load is in progress, then spa_dsl_pool may be
355 		 * uninitialized.  But we shouldn't be resilvering during a spa
356 		 * load anyway.
357 		 */
358 		boolean_t replacing = (vd->vdev_ops == &vdev_replacing_ops ||
359 		    vd->vdev_ops == &vdev_spare_ops) &&
360 		    spa_load_state(vd->vdev_spa) == SPA_LOAD_NONE &&
361 		    dsl_scan_resilvering(vd->vdev_spa->spa_dsl_pool);
362 		mm = vdev_mirror_map_alloc(vd->vdev_children, replacing,
363 		    B_FALSE);
364 		for (c = 0; c < mm->mm_children; c++) {
365 			mc = &mm->mm_child[c];
366 			mc->mc_vd = vd->vdev_child[c];
367 			mc->mc_offset = zio->io_offset;
368 
369 			if (vdev_mirror_rebuilding(mc->mc_vd))
370 				mm->mm_rebuilding = mc->mc_rebuilding = B_TRUE;
371 		}
372 	}
373 
374 	return (mm);
375 }
376 
377 static int
378 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
379     uint64_t *logical_ashift, uint64_t *physical_ashift, cred_t *cr)
380 {
381 	int numerrors = 0;
382 	int lasterror = 0;
383 
384 	if (vd->vdev_children == 0) {
385 		vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
386 		return (SET_ERROR(EINVAL));
387 	}
388 
389 	vdev_open_children(vd, cr);
390 
391 	for (int c = 0; c < vd->vdev_children; c++) {
392 		vdev_t *cvd = vd->vdev_child[c];
393 
394 		if (cvd->vdev_open_error) {
395 			lasterror = cvd->vdev_open_error;
396 			numerrors++;
397 			continue;
398 		}
399 
400 		*asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
401 		*max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
402 		*logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift);
403 	}
404 	for (int c = 0; c < vd->vdev_children; c++) {
405 		vdev_t *cvd = vd->vdev_child[c];
406 
407 		if (cvd->vdev_open_error)
408 			continue;
409 		*physical_ashift = vdev_best_ashift(*logical_ashift,
410 		    *physical_ashift, cvd->vdev_physical_ashift);
411 	}
412 
413 	if (numerrors == vd->vdev_children) {
414 		if (vdev_children_are_offline(vd))
415 			vd->vdev_stat.vs_aux = VDEV_AUX_CHILDREN_OFFLINE;
416 		else
417 			vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
418 		return (lasterror);
419 	}
420 
421 	return (0);
422 }
423 
424 static void
425 vdev_mirror_close(vdev_t *vd)
426 {
427 	for (int c = 0; c < vd->vdev_children; c++)
428 		vdev_close(vd->vdev_child[c]);
429 }
430 
431 static void
432 vdev_mirror_child_done(zio_t *zio)
433 {
434 	mirror_child_t *mc = zio->io_private;
435 
436 	mc->mc_error = zio->io_error;
437 	mc->mc_tried = 1;
438 	mc->mc_skipped = 0;
439 }
440 
441 /*
442  * Check the other, lower-index DVAs to see if they're on the same
443  * vdev as the child we picked.  If they are, use them since they
444  * are likely to have been allocated from the primary metaslab in
445  * use at the time, and hence are more likely to have locality with
446  * single-copy data.
447  */
448 static int
449 vdev_mirror_dva_select(zio_t *zio, int p)
450 {
451 	dva_t *dva = zio->io_bp->blk_dva;
452 	mirror_map_t *mm = zio->io_vsd;
453 	int preferred;
454 	int c;
455 
456 	preferred = mm->mm_preferred[p];
457 	for (p--; p >= 0; p--) {
458 		c = mm->mm_preferred[p];
459 		if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
460 			preferred = c;
461 	}
462 	return (preferred);
463 }
464 
465 static int
466 vdev_mirror_preferred_child_randomize(zio_t *zio)
467 {
468 	mirror_map_t *mm = zio->io_vsd;
469 	int p;
470 
471 	if (mm->mm_root) {
472 		p = random_in_range(mm->mm_preferred_cnt);
473 		return (vdev_mirror_dva_select(zio, p));
474 	}
475 
476 	/*
477 	 * To ensure we don't always favour the first matching vdev,
478 	 * which could lead to wear leveling issues on SSD's, we
479 	 * use the I/O offset as a pseudo random seed into the vdevs
480 	 * which have the lowest load.
481 	 */
482 	p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt;
483 	return (mm->mm_preferred[p]);
484 }
485 
486 static boolean_t
487 vdev_mirror_child_readable(mirror_child_t *mc)
488 {
489 	vdev_t *vd = mc->mc_vd;
490 
491 	if (vd->vdev_top != NULL && vd->vdev_top->vdev_ops == &vdev_draid_ops)
492 		return (vdev_draid_readable(vd, mc->mc_offset));
493 	else
494 		return (vdev_readable(vd));
495 }
496 
497 static boolean_t
498 vdev_mirror_child_missing(mirror_child_t *mc, uint64_t txg, uint64_t size)
499 {
500 	vdev_t *vd = mc->mc_vd;
501 
502 	if (vd->vdev_top != NULL && vd->vdev_top->vdev_ops == &vdev_draid_ops)
503 		return (vdev_draid_missing(vd, mc->mc_offset, txg, size));
504 	else
505 		return (vdev_dtl_contains(vd, DTL_MISSING, txg, size));
506 }
507 
508 /*
509  * Try to find a vdev whose DTL doesn't contain the block we want to read
510  * preferring vdevs based on determined load. If we can't, try the read on
511  * any vdev we haven't already tried.
512  *
513  * Distributed spares are an exception to the above load rule. They are
514  * always preferred in order to detect gaps in the distributed spare which
515  * are created when another disk in the dRAID fails. In order to restore
516  * redundancy those gaps must be read to trigger the required repair IO.
517  */
518 static int
519 vdev_mirror_child_select(zio_t *zio)
520 {
521 	mirror_map_t *mm = zio->io_vsd;
522 	uint64_t txg = zio->io_txg;
523 	int c, lowest_load;
524 
525 	ASSERT(zio->io_bp == NULL || BP_GET_PHYSICAL_BIRTH(zio->io_bp) == txg);
526 
527 	lowest_load = INT_MAX;
528 	mm->mm_preferred_cnt = 0;
529 	for (c = 0; c < mm->mm_children; c++) {
530 		mirror_child_t *mc;
531 
532 		mc = &mm->mm_child[c];
533 		if (mc->mc_tried || mc->mc_skipped)
534 			continue;
535 
536 		if (mc->mc_vd == NULL ||
537 		    !vdev_mirror_child_readable(mc)) {
538 			mc->mc_error = SET_ERROR(ENXIO);
539 			mc->mc_tried = 1;	/* don't even try */
540 			mc->mc_skipped = 1;
541 			continue;
542 		}
543 
544 		if (vdev_mirror_child_missing(mc, txg, 1)) {
545 			mc->mc_error = SET_ERROR(ESTALE);
546 			mc->mc_skipped = 1;
547 			mc->mc_speculative = 1;
548 			continue;
549 		}
550 
551 		if (mc->mc_vd->vdev_ops == &vdev_draid_spare_ops) {
552 			mm->mm_preferred[0] = c;
553 			mm->mm_preferred_cnt = 1;
554 			break;
555 		}
556 
557 		mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
558 		if (mc->mc_load > lowest_load)
559 			continue;
560 
561 		if (mc->mc_load < lowest_load) {
562 			lowest_load = mc->mc_load;
563 			mm->mm_preferred_cnt = 0;
564 		}
565 		mm->mm_preferred[mm->mm_preferred_cnt] = c;
566 		mm->mm_preferred_cnt++;
567 	}
568 
569 	if (mm->mm_preferred_cnt == 1) {
570 		MIRROR_BUMP(vdev_mirror_stat_preferred_found);
571 		return (mm->mm_preferred[0]);
572 	}
573 
574 	if (mm->mm_preferred_cnt > 1) {
575 		MIRROR_BUMP(vdev_mirror_stat_preferred_not_found);
576 		return (vdev_mirror_preferred_child_randomize(zio));
577 	}
578 
579 	/*
580 	 * Every device is either missing or has this txg in its DTL.
581 	 * Look for any child we haven't already tried before giving up.
582 	 */
583 	for (c = 0; c < mm->mm_children; c++) {
584 		if (!mm->mm_child[c].mc_tried)
585 			return (c);
586 	}
587 
588 	/*
589 	 * Every child failed.  There's no place left to look.
590 	 */
591 	return (-1);
592 }
593 
594 static void
595 vdev_mirror_io_start(zio_t *zio)
596 {
597 	mirror_map_t *mm;
598 	mirror_child_t *mc;
599 	int c, children;
600 
601 	mm = vdev_mirror_map_init(zio);
602 	zio->io_vsd = mm;
603 	zio->io_vsd_ops = &vdev_mirror_vsd_ops;
604 
605 	if (mm == NULL) {
606 		ASSERT(!spa_trust_config(zio->io_spa));
607 		ASSERT(zio->io_type == ZIO_TYPE_READ);
608 		zio_execute(zio);
609 		return;
610 	}
611 
612 	if (zio->io_type == ZIO_TYPE_READ) {
613 		if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_resilvering) {
614 			/*
615 			 * For scrubbing reads we need to issue reads to all
616 			 * children.  One child can reuse parent buffer, but
617 			 * for others we have to allocate separate ones to
618 			 * verify checksums if io_bp is non-NULL, or compare
619 			 * them in vdev_mirror_io_done() otherwise.
620 			 */
621 			boolean_t first = B_TRUE;
622 
623 			if (mm->mm_children > 1)
624 				zio_batch_create(zio);
625 
626 			for (c = 0; c < mm->mm_children; c++) {
627 				mc = &mm->mm_child[c];
628 
629 				/* Don't issue ZIOs to offline children */
630 				if (!vdev_mirror_child_readable(mc)) {
631 					mc->mc_error = SET_ERROR(ENXIO);
632 					mc->mc_tried = 1;
633 					mc->mc_skipped = 1;
634 					continue;
635 				}
636 
637 				mc->mc_abd = first ? zio->io_abd :
638 				    abd_alloc_sametype(zio->io_abd,
639 				    zio->io_size);
640 				zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
641 				    mc->mc_vd, mc->mc_offset, mc->mc_abd,
642 				    zio->io_size, zio->io_type,
643 				    zio->io_priority, 0,
644 				    vdev_mirror_child_done, mc));
645 				first = B_FALSE;
646 			}
647 
648 			zio_execute(zio_batch_rele(zio));
649 			return;
650 		}
651 		/*
652 		 * For normal reads just pick one child.
653 		 */
654 		c = vdev_mirror_child_select(zio);
655 		children = (c >= 0);
656 	} else {
657 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
658 
659 		/*
660 		 * Writes go to all children.
661 		 */
662 		c = 0;
663 		children = mm->mm_children;
664 	}
665 
666 	if (children > 1)
667 		zio_batch_create(zio);
668 
669 	while (children--) {
670 		mc = &mm->mm_child[c++];
671 
672 		/*
673 		 * When sequentially resilvering and the integrity of the data
674 		 * is speculative (ZIO_FLAG_SPECULATIVE), issue write repair IOs
675 		 * only to the vdev which is being rebuilt. Existing data on
676 		 * other children must never be overwritten with unconfirmed
677 		 * data to avoid unrecoverable damage to the pool.
678 		 */
679 		if ((zio->io_priority == ZIO_PRIORITY_REBUILD) &&
680 		    (zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
681 		    !(zio->io_flags & ZIO_FLAG_SCRUB) &&
682 		    (zio->io_flags & ZIO_FLAG_SPECULATIVE) &&
683 		    mm->mm_rebuilding && !mc->mc_rebuilding) {
684 			continue;
685 		}
686 
687 		zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
688 		    mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
689 		    zio->io_type, zio->io_priority, 0,
690 		    vdev_mirror_child_done, mc));
691 	}
692 
693 	zio_execute(zio_batch_rele(zio));
694 }
695 
696 static int
697 vdev_mirror_worst_error(mirror_map_t *mm)
698 {
699 	int error[2] = { 0, 0 };
700 
701 	for (int c = 0; c < mm->mm_children; c++) {
702 		mirror_child_t *mc = &mm->mm_child[c];
703 		int s = mc->mc_speculative;
704 		error[s] = zio_worst_error(error[s], mc->mc_error);
705 	}
706 
707 	return (error[0] ? error[0] : error[1]);
708 }
709 
710 static void
711 vdev_mirror_io_done(zio_t *zio)
712 {
713 	mirror_map_t *mm = zio->io_vsd;
714 	mirror_child_t *mc;
715 	int c;
716 	int good_copies = 0;
717 	int unexpected_errors = 0;
718 	int last_good_copy = -1;
719 
720 	if (mm == NULL)
721 		return;
722 
723 	for (c = 0; c < mm->mm_children; c++) {
724 		mc = &mm->mm_child[c];
725 
726 		if (mc->mc_error) {
727 			if (!mc->mc_skipped)
728 				unexpected_errors++;
729 		} else if (mc->mc_tried) {
730 			last_good_copy = c;
731 			good_copies++;
732 		}
733 	}
734 
735 	if (zio->io_type == ZIO_TYPE_WRITE) {
736 		/*
737 		 * XXX -- for now, treat partial writes as success.
738 		 *
739 		 * Now that we support write reallocation, it would be better
740 		 * to treat partial failure as real failure unless there are
741 		 * no non-degraded top-level vdevs left, and not update DTLs
742 		 * if we intend to reallocate.
743 		 */
744 		if (good_copies != mm->mm_children) {
745 			/*
746 			 * Always require at least one good copy.
747 			 *
748 			 * For ditto blocks (io_vd == NULL), require
749 			 * all copies to be good.
750 			 *
751 			 * XXX -- for replacing vdevs, there's no great answer.
752 			 * If the old device is really dead, we may not even
753 			 * be able to access it -- so we only want to
754 			 * require good writes to the new device.  But if
755 			 * the new device turns out to be flaky, we want
756 			 * to be able to detach it -- which requires all
757 			 * writes to the old device to have succeeded.
758 			 */
759 			if (good_copies == 0 || zio->io_vd == NULL)
760 				zio->io_error = vdev_mirror_worst_error(mm);
761 		}
762 		return;
763 	}
764 
765 	ASSERT(zio->io_type == ZIO_TYPE_READ);
766 
767 	/*
768 	 * Any Direct I/O read that has a checksum error must be treated as
769 	 * suspicious as the contents of the buffer could be getting
770 	 * manipulated while the I/O is taking place. The checksum verify error
771 	 * will be reported to the top-level Mirror VDEV.
772 	 *
773 	 * There will be no attampt at reading any additional data copies. If
774 	 * the buffer is still being manipulated while attempting to read from
775 	 * another child, there exists a possibly that the checksum could be
776 	 * verified as valid. However, the buffer contents could again get
777 	 * manipulated after verifying the checksum. This would lead to bad data
778 	 * being written out during self healing.
779 	 */
780 	if ((zio->io_flags & ZIO_FLAG_DIO_READ) &&
781 	    (zio->io_post & ZIO_POST_DIO_CHKSUM_ERR)) {
782 		zio_dio_chksum_verify_error_report(zio);
783 		zio->io_error = vdev_mirror_worst_error(mm);
784 		ASSERT3U(zio->io_error, ==, ECKSUM);
785 		return;
786 	}
787 
788 	/*
789 	 * If we don't have a good copy yet, keep trying other children.
790 	 */
791 	if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
792 		ASSERT(c >= 0 && c < mm->mm_children);
793 		mc = &mm->mm_child[c];
794 		zio_vdev_io_redone(zio);
795 		zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
796 		    mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
797 		    ZIO_TYPE_READ, zio->io_priority, 0,
798 		    vdev_mirror_child_done, mc));
799 		return;
800 	}
801 
802 	if (zio->io_flags & ZIO_FLAG_SCRUB && !mm->mm_resilvering) {
803 		abd_t *best_abd = NULL;
804 		if (last_good_copy >= 0)
805 			best_abd = mm->mm_child[last_good_copy].mc_abd;
806 
807 		/*
808 		 * If we're scrubbing but don't have a BP available (because
809 		 * this vdev is under a raidz or draid vdev) then the best we
810 		 * can do is compare all of the copies read.  If they're not
811 		 * identical then return a checksum error and the most likely
812 		 * correct data.  The raidz code will issue a repair I/O if
813 		 * possible.
814 		 */
815 		if (zio->io_bp == NULL) {
816 			ASSERT(zio->io_vd->vdev_ops == &vdev_replacing_ops ||
817 			    zio->io_vd->vdev_ops == &vdev_spare_ops);
818 
819 			abd_t *pref_abd = NULL;
820 			for (c = 0; c < last_good_copy; c++) {
821 				mc = &mm->mm_child[c];
822 				if (mc->mc_error || !mc->mc_tried)
823 					continue;
824 
825 				if (abd_cmp(mc->mc_abd, best_abd) != 0)
826 					zio->io_error = SET_ERROR(ECKSUM);
827 
828 				/*
829 				 * The distributed spare is always prefered
830 				 * by vdev_mirror_child_select() so it's
831 				 * considered to be the best candidate.
832 				 */
833 				if (pref_abd == NULL &&
834 				    mc->mc_vd->vdev_ops ==
835 				    &vdev_draid_spare_ops)
836 					pref_abd = mc->mc_abd;
837 
838 				/*
839 				 * In the absence of a preferred copy, use
840 				 * the parent pointer to avoid a memory copy.
841 				 */
842 				if (mc->mc_abd == zio->io_abd)
843 					best_abd = mc->mc_abd;
844 			}
845 			if (pref_abd)
846 				best_abd = pref_abd;
847 		} else {
848 
849 			/*
850 			 * If we have a BP available, then checksums are
851 			 * already verified and we just need a buffer
852 			 * with valid data, preferring parent one to
853 			 * avoid a memory copy.
854 			 */
855 			for (c = 0; c < last_good_copy; c++) {
856 				mc = &mm->mm_child[c];
857 				if (mc->mc_error || !mc->mc_tried)
858 					continue;
859 				if (mc->mc_abd == zio->io_abd) {
860 					best_abd = mc->mc_abd;
861 					break;
862 				}
863 			}
864 		}
865 
866 		if (best_abd && best_abd != zio->io_abd)
867 			abd_copy(zio->io_abd, best_abd, zio->io_size);
868 		for (c = 0; c < mm->mm_children; c++) {
869 			mc = &mm->mm_child[c];
870 			if (mc->mc_abd != zio->io_abd)
871 				abd_free(mc->mc_abd);
872 			mc->mc_abd = NULL;
873 		}
874 	}
875 
876 	if (good_copies == 0) {
877 		zio->io_error = vdev_mirror_worst_error(mm);
878 		ASSERT(zio->io_error != 0);
879 	}
880 
881 	if (good_copies && spa_writeable(zio->io_spa) &&
882 	    (unexpected_errors ||
883 	    (zio->io_flags & ZIO_FLAG_RESILVER) ||
884 	    ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_resilvering))) {
885 		/*
886 		 * Use the good data we have in hand to repair damaged children.
887 		 */
888 		for (c = 0; c < mm->mm_children; c++) {
889 			/*
890 			 * Don't rewrite known good children.
891 			 * Not only is it unnecessary, it could
892 			 * actually be harmful: if the system lost
893 			 * power while rewriting the only good copy,
894 			 * there would be no good copies left!
895 			 */
896 			mc = &mm->mm_child[c];
897 
898 			if (mc->mc_error == 0) {
899 				vdev_ops_t *ops = mc->mc_vd->vdev_ops;
900 
901 				if (mc->mc_tried)
902 					continue;
903 				/*
904 				 * We didn't try this child.  We need to
905 				 * repair it if:
906 				 * 1. it's a scrub (in which case we have
907 				 * tried everything that was healthy)
908 				 *  - or -
909 				 * 2. it's an indirect or distributed spare
910 				 * vdev (in which case it could point to any
911 				 * other vdev, which might have a bad DTL)
912 				 *  - or -
913 				 * 3. the DTL indicates that this data is
914 				 * missing from this vdev
915 				 */
916 				if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
917 				    ops != &vdev_indirect_ops &&
918 				    ops != &vdev_draid_spare_ops &&
919 				    !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
920 				    zio->io_txg, 1))
921 					continue;
922 				mc->mc_error = SET_ERROR(ESTALE);
923 			}
924 
925 			zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
926 			    mc->mc_vd, mc->mc_offset,
927 			    zio->io_abd, zio->io_size, ZIO_TYPE_WRITE,
928 			    zio->io_priority == ZIO_PRIORITY_REBUILD ?
929 			    ZIO_PRIORITY_REBUILD : ZIO_PRIORITY_ASYNC_WRITE,
930 			    ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
931 			    ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
932 		}
933 	}
934 }
935 
936 static void
937 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
938 {
939 	if (faulted == vd->vdev_children) {
940 		if (vdev_children_are_offline(vd)) {
941 			vdev_set_state(vd, B_FALSE, VDEV_STATE_OFFLINE,
942 			    VDEV_AUX_CHILDREN_OFFLINE);
943 		} else {
944 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
945 			    VDEV_AUX_NO_REPLICAS);
946 		}
947 	} else if (degraded + faulted != 0) {
948 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
949 	} else {
950 		vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
951 	}
952 }
953 
954 /*
955  * Return the maximum asize for a rebuild zio in the provided range.
956  */
957 static uint64_t
958 vdev_mirror_rebuild_asize(vdev_t *vd, uint64_t start, uint64_t asize,
959     uint64_t max_segment)
960 {
961 	(void) start;
962 
963 	uint64_t psize = MIN(P2ROUNDUP(max_segment, 1 << vd->vdev_ashift),
964 	    SPA_MAXBLOCKSIZE);
965 
966 	return (MIN(asize, vdev_psize_to_asize(vd, psize)));
967 }
968 
969 vdev_ops_t vdev_mirror_ops = {
970 	.vdev_op_init = NULL,
971 	.vdev_op_fini = NULL,
972 	.vdev_op_open = vdev_mirror_open,
973 	.vdev_op_close = vdev_mirror_close,
974 	.vdev_op_psize_to_asize = vdev_default_asize,
975 	.vdev_op_asize_to_psize = vdev_default_psize,
976 	.vdev_op_min_asize = vdev_default_min_asize,
977 	.vdev_op_min_alloc = NULL,
978 	.vdev_op_io_start = vdev_mirror_io_start,
979 	.vdev_op_io_done = vdev_mirror_io_done,
980 	.vdev_op_state_change = vdev_mirror_state_change,
981 	.vdev_op_need_resilver = vdev_default_need_resilver,
982 	.vdev_op_hold = NULL,
983 	.vdev_op_rele = NULL,
984 	.vdev_op_remap = NULL,
985 	.vdev_op_xlate = vdev_default_xlate,
986 	.vdev_op_rebuild_asize = vdev_mirror_rebuild_asize,
987 	.vdev_op_metaslab_init = NULL,
988 	.vdev_op_config_generate = NULL,
989 	.vdev_op_nparity = NULL,
990 	.vdev_op_ndisks = NULL,
991 	.vdev_op_type = VDEV_TYPE_MIRROR,	/* name of this vdev type */
992 	.vdev_op_leaf = B_FALSE			/* not a leaf vdev */
993 };
994 
995 vdev_ops_t vdev_replacing_ops = {
996 	.vdev_op_init = NULL,
997 	.vdev_op_fini = NULL,
998 	.vdev_op_open = vdev_mirror_open,
999 	.vdev_op_close = vdev_mirror_close,
1000 	.vdev_op_psize_to_asize = vdev_default_asize,
1001 	.vdev_op_asize_to_psize = vdev_default_psize,
1002 	.vdev_op_min_asize = vdev_default_min_asize,
1003 	.vdev_op_min_alloc = NULL,
1004 	.vdev_op_io_start = vdev_mirror_io_start,
1005 	.vdev_op_io_done = vdev_mirror_io_done,
1006 	.vdev_op_state_change = vdev_mirror_state_change,
1007 	.vdev_op_need_resilver = vdev_default_need_resilver,
1008 	.vdev_op_hold = NULL,
1009 	.vdev_op_rele = NULL,
1010 	.vdev_op_remap = NULL,
1011 	.vdev_op_xlate = vdev_default_xlate,
1012 	.vdev_op_rebuild_asize = vdev_mirror_rebuild_asize,
1013 	.vdev_op_metaslab_init = NULL,
1014 	.vdev_op_config_generate = NULL,
1015 	.vdev_op_nparity = NULL,
1016 	.vdev_op_ndisks = NULL,
1017 	.vdev_op_type = VDEV_TYPE_REPLACING,	/* name of this vdev type */
1018 	.vdev_op_leaf = B_FALSE			/* not a leaf vdev */
1019 };
1020 
1021 vdev_ops_t vdev_spare_ops = {
1022 	.vdev_op_init = NULL,
1023 	.vdev_op_fini = NULL,
1024 	.vdev_op_open = vdev_mirror_open,
1025 	.vdev_op_close = vdev_mirror_close,
1026 	.vdev_op_psize_to_asize = vdev_default_asize,
1027 	.vdev_op_asize_to_psize = vdev_default_psize,
1028 	.vdev_op_min_asize = vdev_default_min_asize,
1029 	.vdev_op_min_alloc = NULL,
1030 	.vdev_op_io_start = vdev_mirror_io_start,
1031 	.vdev_op_io_done = vdev_mirror_io_done,
1032 	.vdev_op_state_change = vdev_mirror_state_change,
1033 	.vdev_op_need_resilver = vdev_default_need_resilver,
1034 	.vdev_op_hold = NULL,
1035 	.vdev_op_rele = NULL,
1036 	.vdev_op_remap = NULL,
1037 	.vdev_op_xlate = vdev_default_xlate,
1038 	.vdev_op_rebuild_asize = vdev_mirror_rebuild_asize,
1039 	.vdev_op_metaslab_init = NULL,
1040 	.vdev_op_config_generate = NULL,
1041 	.vdev_op_nparity = NULL,
1042 	.vdev_op_ndisks = NULL,
1043 	.vdev_op_type = VDEV_TYPE_SPARE,	/* name of this vdev type */
1044 	.vdev_op_leaf = B_FALSE			/* not a leaf vdev */
1045 };
1046 
1047 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_inc, INT, ZMOD_RW,
1048 	"Rotating media load increment for non-seeking I/Os");
1049 
1050 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_seek_inc, INT,
1051 	ZMOD_RW, "Rotating media load increment for seeking I/Os");
1052 
1053 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_seek_offset, INT,
1054 	ZMOD_RW,
1055 	"Offset in bytes from the last I/O which triggers "
1056 	"a reduced rotating media seek increment");
1057 
1058 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, non_rotating_inc, INT,
1059 	ZMOD_RW, "Non-rotating media load increment for non-seeking I/Os");
1060 
1061 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, non_rotating_seek_inc, INT,
1062 	ZMOD_RW, "Non-rotating media load increment for seeking I/Os");
1063