xref: /freebsd/sys/contrib/openzfs/module/zfs/vdev_mirror.c (revision 61145dc2b94f12f6a47344fb9aac702321880e43)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
29  */
30 
31 #include <sys/zfs_context.h>
32 #include <sys/spa.h>
33 #include <sys/spa_impl.h>
34 #include <sys/dsl_pool.h>
35 #include <sys/dsl_scan.h>
36 #include <sys/vdev_impl.h>
37 #include <sys/vdev_draid.h>
38 #include <sys/zio.h>
39 #include <sys/zio_checksum.h>
40 #include <sys/abd.h>
41 #include <sys/fs/zfs.h>
42 
43 /*
44  * Vdev mirror kstats
45  */
46 static kstat_t *mirror_ksp = NULL;
47 
48 typedef struct mirror_stats {
49 	kstat_named_t vdev_mirror_stat_rotating_linear;
50 	kstat_named_t vdev_mirror_stat_rotating_offset;
51 	kstat_named_t vdev_mirror_stat_rotating_seek;
52 	kstat_named_t vdev_mirror_stat_non_rotating_linear;
53 	kstat_named_t vdev_mirror_stat_non_rotating_seek;
54 
55 	kstat_named_t vdev_mirror_stat_preferred_found;
56 	kstat_named_t vdev_mirror_stat_preferred_not_found;
57 } mirror_stats_t;
58 
59 static mirror_stats_t mirror_stats = {
60 	/* New I/O follows directly the last I/O */
61 	{ "rotating_linear",			KSTAT_DATA_UINT64 },
62 	/* New I/O is within zfs_vdev_mirror_rotating_seek_offset of the last */
63 	{ "rotating_offset",			KSTAT_DATA_UINT64 },
64 	/* New I/O requires random seek */
65 	{ "rotating_seek",			KSTAT_DATA_UINT64 },
66 	/* New I/O follows directly the last I/O  (nonrot) */
67 	{ "non_rotating_linear",		KSTAT_DATA_UINT64 },
68 	/* New I/O requires random seek (nonrot) */
69 	{ "non_rotating_seek",			KSTAT_DATA_UINT64 },
70 	/* Preferred child vdev found */
71 	{ "preferred_found",			KSTAT_DATA_UINT64 },
72 	/* Preferred child vdev not found or equal load  */
73 	{ "preferred_not_found",		KSTAT_DATA_UINT64 },
74 
75 };
76 
77 #define	MIRROR_STAT(stat)		(mirror_stats.stat.value.ui64)
78 #define	MIRROR_INCR(stat, val) 		atomic_add_64(&MIRROR_STAT(stat), val)
79 #define	MIRROR_BUMP(stat)		MIRROR_INCR(stat, 1)
80 
81 void
vdev_mirror_stat_init(void)82 vdev_mirror_stat_init(void)
83 {
84 	mirror_ksp = kstat_create("zfs", 0, "vdev_mirror_stats",
85 	    "misc", KSTAT_TYPE_NAMED,
86 	    sizeof (mirror_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
87 	if (mirror_ksp != NULL) {
88 		mirror_ksp->ks_data = &mirror_stats;
89 		kstat_install(mirror_ksp);
90 	}
91 }
92 
93 void
vdev_mirror_stat_fini(void)94 vdev_mirror_stat_fini(void)
95 {
96 	if (mirror_ksp != NULL) {
97 		kstat_delete(mirror_ksp);
98 		mirror_ksp = NULL;
99 	}
100 }
101 
102 /*
103  * Virtual device vector for mirroring.
104  */
105 typedef struct mirror_child {
106 	vdev_t		*mc_vd;
107 	abd_t		*mc_abd;
108 	uint64_t	mc_offset;
109 	int		mc_error;
110 	int		mc_load;
111 	uint8_t		mc_tried;
112 	uint8_t		mc_skipped;
113 	uint8_t		mc_speculative;
114 	uint8_t		mc_rebuilding;
115 } mirror_child_t;
116 
117 typedef struct mirror_map {
118 	int		*mm_preferred;
119 	int		mm_preferred_cnt;
120 	int		mm_children;
121 	boolean_t	mm_resilvering;
122 	boolean_t	mm_rebuilding;
123 	boolean_t	mm_root;
124 	mirror_child_t	mm_child[];
125 } mirror_map_t;
126 
127 static const int vdev_mirror_shift = 21;
128 
129 /*
130  * The load configuration settings below are tuned by default for
131  * the case where all devices are of the same rotational type.
132  *
133  * If there is a mixture of rotating and non-rotating media, setting
134  * zfs_vdev_mirror_non_rotating_seek_inc to 0 may well provide better results
135  * as it will direct more reads to the non-rotating vdevs which are more likely
136  * to have a higher performance.
137  */
138 
139 /* Rotating media load calculation configuration. */
140 static int zfs_vdev_mirror_rotating_inc = 0;
141 static int zfs_vdev_mirror_rotating_seek_inc = 5;
142 static int zfs_vdev_mirror_rotating_seek_offset = 1 * 1024 * 1024;
143 
144 /* Non-rotating media load calculation configuration. */
145 static int zfs_vdev_mirror_non_rotating_inc = 0;
146 static int zfs_vdev_mirror_non_rotating_seek_inc = 1;
147 
148 static inline size_t
vdev_mirror_map_size(int children)149 vdev_mirror_map_size(int children)
150 {
151 	return (offsetof(mirror_map_t, mm_child[children]) +
152 	    sizeof (int) * children);
153 }
154 
155 static inline mirror_map_t *
vdev_mirror_map_alloc(int children,boolean_t resilvering,boolean_t root)156 vdev_mirror_map_alloc(int children, boolean_t resilvering, boolean_t root)
157 {
158 	mirror_map_t *mm;
159 
160 	mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP);
161 	mm->mm_children = children;
162 	mm->mm_resilvering = resilvering;
163 	mm->mm_root = root;
164 	mm->mm_preferred = (int *)((uintptr_t)mm +
165 	    offsetof(mirror_map_t, mm_child[children]));
166 
167 	return (mm);
168 }
169 
170 static void
vdev_mirror_map_free(zio_t * zio)171 vdev_mirror_map_free(zio_t *zio)
172 {
173 	mirror_map_t *mm = zio->io_vsd;
174 
175 	kmem_free(mm, vdev_mirror_map_size(mm->mm_children));
176 }
177 
178 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
179 	.vsd_free = vdev_mirror_map_free,
180 };
181 
182 static int
vdev_mirror_load(mirror_map_t * mm,vdev_t * vd,uint64_t zio_offset)183 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
184 {
185 	uint64_t last_offset;
186 	int64_t offset_diff;
187 	int load;
188 
189 	/* All DVAs have equal weight at the root. */
190 	if (mm->mm_root)
191 		return (INT_MAX);
192 
193 	/*
194 	 * We don't return INT_MAX if the device is resilvering i.e.
195 	 * vdev_resilver_txg != 0 as when tested performance was slightly
196 	 * worse overall when resilvering with compared to without.
197 	 */
198 
199 	/* Fix zio_offset for leaf vdevs */
200 	if (vd->vdev_ops->vdev_op_leaf)
201 		zio_offset += VDEV_LABEL_START_SIZE;
202 
203 	/* Standard load based on pending queue length. */
204 	load = vdev_queue_length(vd);
205 	last_offset = vdev_queue_last_offset(vd);
206 
207 	if (vd->vdev_nonrot) {
208 		/* Non-rotating media. */
209 		if (last_offset == zio_offset) {
210 			MIRROR_BUMP(vdev_mirror_stat_non_rotating_linear);
211 			return (load + zfs_vdev_mirror_non_rotating_inc);
212 		}
213 
214 		/*
215 		 * Apply a seek penalty even for non-rotating devices as
216 		 * sequential I/O's can be aggregated into fewer operations on
217 		 * the device, thus avoiding unnecessary per-command overhead
218 		 * and boosting performance.
219 		 */
220 		MIRROR_BUMP(vdev_mirror_stat_non_rotating_seek);
221 		return (load + zfs_vdev_mirror_non_rotating_seek_inc);
222 	}
223 
224 	/* Rotating media I/O's which directly follow the last I/O. */
225 	if (last_offset == zio_offset) {
226 		MIRROR_BUMP(vdev_mirror_stat_rotating_linear);
227 		return (load + zfs_vdev_mirror_rotating_inc);
228 	}
229 
230 	/*
231 	 * Apply half the seek increment to I/O's within seek offset
232 	 * of the last I/O issued to this vdev as they should incur less
233 	 * of a seek increment.
234 	 */
235 	offset_diff = (int64_t)(last_offset - zio_offset);
236 	if (ABS(offset_diff) < zfs_vdev_mirror_rotating_seek_offset) {
237 		MIRROR_BUMP(vdev_mirror_stat_rotating_offset);
238 		return (load + (zfs_vdev_mirror_rotating_seek_inc / 2));
239 	}
240 
241 	/* Apply the full seek increment to all other I/O's. */
242 	MIRROR_BUMP(vdev_mirror_stat_rotating_seek);
243 	return (load + zfs_vdev_mirror_rotating_seek_inc);
244 }
245 
246 static boolean_t
vdev_mirror_rebuilding(vdev_t * vd)247 vdev_mirror_rebuilding(vdev_t *vd)
248 {
249 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_rebuild_txg)
250 		return (B_TRUE);
251 
252 	for (int i = 0; i < vd->vdev_children; i++) {
253 		if (vdev_mirror_rebuilding(vd->vdev_child[i])) {
254 			return (B_TRUE);
255 		}
256 	}
257 
258 	return (B_FALSE);
259 }
260 
261 /*
262  * Avoid inlining the function to keep vdev_mirror_io_start(), which
263  * is this functions only caller, as small as possible on the stack.
264  */
265 noinline static mirror_map_t *
vdev_mirror_map_init(zio_t * zio)266 vdev_mirror_map_init(zio_t *zio)
267 {
268 	mirror_map_t *mm = NULL;
269 	mirror_child_t *mc;
270 	vdev_t *vd = zio->io_vd;
271 	int c;
272 
273 	if (vd == NULL) {
274 		dva_t *dva = zio->io_bp->blk_dva;
275 		spa_t *spa = zio->io_spa;
276 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
277 		dva_t dva_copy[SPA_DVAS_PER_BP];
278 
279 		/*
280 		 * The sequential scrub code sorts and issues all DVAs
281 		 * of a bp separately. Each of these IOs includes all
282 		 * original DVA copies so that repairs can be performed
283 		 * in the event of an error, but we only actually want
284 		 * to check the first DVA since the others will be
285 		 * checked by their respective sorted IOs. Only if we
286 		 * hit an error will we try all DVAs upon retrying.
287 		 *
288 		 * Note: This check is safe even if the user switches
289 		 * from a legacy scrub to a sequential one in the middle
290 		 * of processing, since scn_is_sorted isn't updated until
291 		 * all outstanding IOs from the previous scrub pass
292 		 * complete.
293 		 */
294 		if ((zio->io_flags & ZIO_FLAG_SCRUB) &&
295 		    !(zio->io_flags & ZIO_FLAG_IO_RETRY) &&
296 		    dsl_scan_scrubbing(spa->spa_dsl_pool) &&
297 		    scn->scn_is_sorted) {
298 			c = 1;
299 		} else {
300 			c = BP_GET_NDVAS(zio->io_bp);
301 		}
302 
303 		/*
304 		 * If the pool cannot be written to, then infer that some
305 		 * DVAs might be invalid or point to vdevs that do not exist.
306 		 * We skip them.
307 		 */
308 		if (!spa_writeable(spa)) {
309 			ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
310 			int j = 0;
311 			for (int i = 0; i < c; i++) {
312 				if (zfs_dva_valid(spa, &dva[i], zio->io_bp))
313 					dva_copy[j++] = dva[i];
314 			}
315 			if (j == 0) {
316 				zio->io_vsd = NULL;
317 				zio->io_error = ENXIO;
318 				return (NULL);
319 			}
320 			if (j < c) {
321 				dva = dva_copy;
322 				c = j;
323 			}
324 		}
325 
326 		mm = vdev_mirror_map_alloc(c, B_FALSE, B_TRUE);
327 		for (c = 0; c < mm->mm_children; c++) {
328 			mc = &mm->mm_child[c];
329 
330 			mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
331 			mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
332 			if (mc->mc_vd == NULL) {
333 				kmem_free(mm, vdev_mirror_map_size(
334 				    mm->mm_children));
335 				zio->io_vsd = NULL;
336 				zio->io_error = ENXIO;
337 				return (NULL);
338 			}
339 		}
340 	} else {
341 		/*
342 		 * If we are resilvering, then we should handle scrub reads
343 		 * differently; we shouldn't issue them to the resilvering
344 		 * device because it might not have those blocks.
345 		 *
346 		 * We are resilvering iff:
347 		 * 1) We are a replacing vdev (ie our name is "replacing-1" or
348 		 *    "spare-1" or something like that), and
349 		 * 2) The pool is currently being resilvered.
350 		 *
351 		 * We cannot simply check vd->vdev_resilver_txg, because it's
352 		 * not set in this path.
353 		 *
354 		 * Nor can we just check our vdev_ops; there are cases (such as
355 		 * when a user types "zpool replace pool odev spare_dev" and
356 		 * spare_dev is in the spare list, or when a spare device is
357 		 * automatically used to replace a DEGRADED device) when
358 		 * resilvering is complete but both the original vdev and the
359 		 * spare vdev remain in the pool.  That behavior is intentional.
360 		 * It helps implement the policy that a spare should be
361 		 * automatically removed from the pool after the user replaces
362 		 * the device that originally failed.
363 		 *
364 		 * If a spa load is in progress, then spa_dsl_pool may be
365 		 * uninitialized.  But we shouldn't be resilvering during a spa
366 		 * load anyway.
367 		 */
368 		boolean_t replacing = (vd->vdev_ops == &vdev_replacing_ops ||
369 		    vd->vdev_ops == &vdev_spare_ops) &&
370 		    spa_load_state(vd->vdev_spa) == SPA_LOAD_NONE &&
371 		    dsl_scan_resilvering(vd->vdev_spa->spa_dsl_pool);
372 		mm = vdev_mirror_map_alloc(vd->vdev_children, replacing,
373 		    B_FALSE);
374 		for (c = 0; c < mm->mm_children; c++) {
375 			mc = &mm->mm_child[c];
376 			mc->mc_vd = vd->vdev_child[c];
377 			mc->mc_offset = zio->io_offset;
378 
379 			if (vdev_mirror_rebuilding(mc->mc_vd))
380 				mm->mm_rebuilding = mc->mc_rebuilding = B_TRUE;
381 		}
382 	}
383 
384 	return (mm);
385 }
386 
387 static int
vdev_mirror_open(vdev_t * vd,uint64_t * asize,uint64_t * max_asize,uint64_t * logical_ashift,uint64_t * physical_ashift)388 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
389     uint64_t *logical_ashift, uint64_t *physical_ashift)
390 {
391 	int numerrors = 0;
392 	int lasterror = 0;
393 
394 	if (vd->vdev_children == 0) {
395 		vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
396 		return (SET_ERROR(EINVAL));
397 	}
398 
399 	vdev_open_children(vd);
400 
401 	for (int c = 0; c < vd->vdev_children; c++) {
402 		vdev_t *cvd = vd->vdev_child[c];
403 
404 		if (cvd->vdev_open_error) {
405 			lasterror = cvd->vdev_open_error;
406 			numerrors++;
407 			continue;
408 		}
409 
410 		*asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
411 		*max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
412 		*logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift);
413 	}
414 	for (int c = 0; c < vd->vdev_children; c++) {
415 		vdev_t *cvd = vd->vdev_child[c];
416 
417 		if (cvd->vdev_open_error)
418 			continue;
419 		*physical_ashift = vdev_best_ashift(*logical_ashift,
420 		    *physical_ashift, cvd->vdev_physical_ashift);
421 	}
422 
423 	if (numerrors == vd->vdev_children) {
424 		if (vdev_children_are_offline(vd))
425 			vd->vdev_stat.vs_aux = VDEV_AUX_CHILDREN_OFFLINE;
426 		else
427 			vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
428 		return (lasterror);
429 	}
430 
431 	return (0);
432 }
433 
434 static void
vdev_mirror_close(vdev_t * vd)435 vdev_mirror_close(vdev_t *vd)
436 {
437 	for (int c = 0; c < vd->vdev_children; c++)
438 		vdev_close(vd->vdev_child[c]);
439 }
440 
441 static void
vdev_mirror_child_done(zio_t * zio)442 vdev_mirror_child_done(zio_t *zio)
443 {
444 	mirror_child_t *mc = zio->io_private;
445 
446 	mc->mc_error = zio->io_error;
447 	mc->mc_tried = 1;
448 	mc->mc_skipped = 0;
449 }
450 
451 /*
452  * Check the other, lower-index DVAs to see if they're on the same
453  * vdev as the child we picked.  If they are, use them since they
454  * are likely to have been allocated from the primary metaslab in
455  * use at the time, and hence are more likely to have locality with
456  * single-copy data.
457  */
458 static int
vdev_mirror_dva_select(zio_t * zio,int p)459 vdev_mirror_dva_select(zio_t *zio, int p)
460 {
461 	dva_t *dva = zio->io_bp->blk_dva;
462 	mirror_map_t *mm = zio->io_vsd;
463 	int preferred;
464 	int c;
465 
466 	preferred = mm->mm_preferred[p];
467 	for (p--; p >= 0; p--) {
468 		c = mm->mm_preferred[p];
469 		if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
470 			preferred = c;
471 	}
472 	return (preferred);
473 }
474 
475 static int
vdev_mirror_preferred_child_randomize(zio_t * zio)476 vdev_mirror_preferred_child_randomize(zio_t *zio)
477 {
478 	mirror_map_t *mm = zio->io_vsd;
479 	int p;
480 
481 	if (mm->mm_root) {
482 		p = random_in_range(mm->mm_preferred_cnt);
483 		return (vdev_mirror_dva_select(zio, p));
484 	}
485 
486 	/*
487 	 * To ensure we don't always favour the first matching vdev,
488 	 * which could lead to wear leveling issues on SSD's, we
489 	 * use the I/O offset as a pseudo random seed into the vdevs
490 	 * which have the lowest load.
491 	 */
492 	p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt;
493 	return (mm->mm_preferred[p]);
494 }
495 
496 static boolean_t
vdev_mirror_child_readable(mirror_child_t * mc)497 vdev_mirror_child_readable(mirror_child_t *mc)
498 {
499 	vdev_t *vd = mc->mc_vd;
500 
501 	if (vd->vdev_top != NULL && vd->vdev_top->vdev_ops == &vdev_draid_ops)
502 		return (vdev_draid_readable(vd, mc->mc_offset));
503 	else
504 		return (vdev_readable(vd));
505 }
506 
507 static boolean_t
vdev_mirror_child_missing(mirror_child_t * mc,uint64_t txg,uint64_t size)508 vdev_mirror_child_missing(mirror_child_t *mc, uint64_t txg, uint64_t size)
509 {
510 	vdev_t *vd = mc->mc_vd;
511 
512 	if (vd->vdev_top != NULL && vd->vdev_top->vdev_ops == &vdev_draid_ops)
513 		return (vdev_draid_missing(vd, mc->mc_offset, txg, size));
514 	else
515 		return (vdev_dtl_contains(vd, DTL_MISSING, txg, size));
516 }
517 
518 /*
519  * Try to find a vdev whose DTL doesn't contain the block we want to read
520  * preferring vdevs based on determined load. If we can't, try the read on
521  * any vdev we haven't already tried.
522  *
523  * Distributed spares are an exception to the above load rule. They are
524  * always preferred in order to detect gaps in the distributed spare which
525  * are created when another disk in the dRAID fails. In order to restore
526  * redundancy those gaps must be read to trigger the required repair IO.
527  */
528 static int
vdev_mirror_child_select(zio_t * zio)529 vdev_mirror_child_select(zio_t *zio)
530 {
531 	mirror_map_t *mm = zio->io_vsd;
532 	uint64_t txg = zio->io_txg;
533 	int c, lowest_load;
534 
535 	ASSERT(zio->io_bp == NULL || BP_GET_BIRTH(zio->io_bp) == txg);
536 
537 	lowest_load = INT_MAX;
538 	mm->mm_preferred_cnt = 0;
539 	for (c = 0; c < mm->mm_children; c++) {
540 		mirror_child_t *mc;
541 
542 		mc = &mm->mm_child[c];
543 		if (mc->mc_tried || mc->mc_skipped)
544 			continue;
545 
546 		if (mc->mc_vd == NULL ||
547 		    !vdev_mirror_child_readable(mc)) {
548 			mc->mc_error = SET_ERROR(ENXIO);
549 			mc->mc_tried = 1;	/* don't even try */
550 			mc->mc_skipped = 1;
551 			continue;
552 		}
553 
554 		if (vdev_mirror_child_missing(mc, txg, 1)) {
555 			mc->mc_error = SET_ERROR(ESTALE);
556 			mc->mc_skipped = 1;
557 			mc->mc_speculative = 1;
558 			continue;
559 		}
560 
561 		if (mc->mc_vd->vdev_ops == &vdev_draid_spare_ops) {
562 			mm->mm_preferred[0] = c;
563 			mm->mm_preferred_cnt = 1;
564 			break;
565 		}
566 
567 		mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
568 		if (mc->mc_load > lowest_load)
569 			continue;
570 
571 		if (mc->mc_load < lowest_load) {
572 			lowest_load = mc->mc_load;
573 			mm->mm_preferred_cnt = 0;
574 		}
575 		mm->mm_preferred[mm->mm_preferred_cnt] = c;
576 		mm->mm_preferred_cnt++;
577 	}
578 
579 	if (mm->mm_preferred_cnt == 1) {
580 		MIRROR_BUMP(vdev_mirror_stat_preferred_found);
581 		return (mm->mm_preferred[0]);
582 	}
583 
584 	if (mm->mm_preferred_cnt > 1) {
585 		MIRROR_BUMP(vdev_mirror_stat_preferred_not_found);
586 		return (vdev_mirror_preferred_child_randomize(zio));
587 	}
588 
589 	/*
590 	 * Every device is either missing or has this txg in its DTL.
591 	 * Look for any child we haven't already tried before giving up.
592 	 */
593 	for (c = 0; c < mm->mm_children; c++) {
594 		if (!mm->mm_child[c].mc_tried)
595 			return (c);
596 	}
597 
598 	/*
599 	 * Every child failed.  There's no place left to look.
600 	 */
601 	return (-1);
602 }
603 
604 static void
vdev_mirror_io_start(zio_t * zio)605 vdev_mirror_io_start(zio_t *zio)
606 {
607 	mirror_map_t *mm;
608 	mirror_child_t *mc;
609 	int c, children;
610 
611 	mm = vdev_mirror_map_init(zio);
612 	zio->io_vsd = mm;
613 	zio->io_vsd_ops = &vdev_mirror_vsd_ops;
614 
615 	if (mm == NULL) {
616 		ASSERT(!spa_trust_config(zio->io_spa));
617 		ASSERT(zio->io_type == ZIO_TYPE_READ);
618 		zio_execute(zio);
619 		return;
620 	}
621 
622 	if (zio->io_type == ZIO_TYPE_READ) {
623 		if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_resilvering) {
624 			/*
625 			 * For scrubbing reads we need to issue reads to all
626 			 * children.  One child can reuse parent buffer, but
627 			 * for others we have to allocate separate ones to
628 			 * verify checksums if io_bp is non-NULL, or compare
629 			 * them in vdev_mirror_io_done() otherwise.
630 			 */
631 			boolean_t first = B_TRUE;
632 			for (c = 0; c < mm->mm_children; c++) {
633 				mc = &mm->mm_child[c];
634 
635 				/* Don't issue ZIOs to offline children */
636 				if (!vdev_mirror_child_readable(mc)) {
637 					mc->mc_error = SET_ERROR(ENXIO);
638 					mc->mc_tried = 1;
639 					mc->mc_skipped = 1;
640 					continue;
641 				}
642 
643 				mc->mc_abd = first ? zio->io_abd :
644 				    abd_alloc_sametype(zio->io_abd,
645 				    zio->io_size);
646 				zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
647 				    mc->mc_vd, mc->mc_offset, mc->mc_abd,
648 				    zio->io_size, zio->io_type,
649 				    zio->io_priority, 0,
650 				    vdev_mirror_child_done, mc));
651 				first = B_FALSE;
652 			}
653 			zio_execute(zio);
654 			return;
655 		}
656 		/*
657 		 * For normal reads just pick one child.
658 		 */
659 		c = vdev_mirror_child_select(zio);
660 		children = (c >= 0);
661 	} else {
662 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
663 
664 		/*
665 		 * Writes go to all children.
666 		 */
667 		c = 0;
668 		children = mm->mm_children;
669 	}
670 
671 	while (children--) {
672 		mc = &mm->mm_child[c];
673 		c++;
674 
675 		/*
676 		 * When sequentially resilvering only issue write repair
677 		 * IOs to the vdev which is being rebuilt since performance
678 		 * is limited by the slowest child.  This is an issue for
679 		 * faster replacement devices such as distributed spares.
680 		 */
681 		if ((zio->io_priority == ZIO_PRIORITY_REBUILD) &&
682 		    (zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
683 		    !(zio->io_flags & ZIO_FLAG_SCRUB) &&
684 		    mm->mm_rebuilding && !mc->mc_rebuilding) {
685 			continue;
686 		}
687 
688 		zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
689 		    mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
690 		    zio->io_type, zio->io_priority, 0,
691 		    vdev_mirror_child_done, mc));
692 	}
693 
694 	zio_execute(zio);
695 }
696 
697 static int
vdev_mirror_worst_error(mirror_map_t * mm)698 vdev_mirror_worst_error(mirror_map_t *mm)
699 {
700 	int error[2] = { 0, 0 };
701 
702 	for (int c = 0; c < mm->mm_children; c++) {
703 		mirror_child_t *mc = &mm->mm_child[c];
704 		int s = mc->mc_speculative;
705 		error[s] = zio_worst_error(error[s], mc->mc_error);
706 	}
707 
708 	return (error[0] ? error[0] : error[1]);
709 }
710 
711 static void
vdev_mirror_io_done(zio_t * zio)712 vdev_mirror_io_done(zio_t *zio)
713 {
714 	mirror_map_t *mm = zio->io_vsd;
715 	mirror_child_t *mc;
716 	int c;
717 	int good_copies = 0;
718 	int unexpected_errors = 0;
719 	int last_good_copy = -1;
720 
721 	if (mm == NULL)
722 		return;
723 
724 	for (c = 0; c < mm->mm_children; c++) {
725 		mc = &mm->mm_child[c];
726 
727 		if (mc->mc_error) {
728 			if (!mc->mc_skipped)
729 				unexpected_errors++;
730 		} else if (mc->mc_tried) {
731 			last_good_copy = c;
732 			good_copies++;
733 		}
734 	}
735 
736 	if (zio->io_type == ZIO_TYPE_WRITE) {
737 		/*
738 		 * XXX -- for now, treat partial writes as success.
739 		 *
740 		 * Now that we support write reallocation, it would be better
741 		 * to treat partial failure as real failure unless there are
742 		 * no non-degraded top-level vdevs left, and not update DTLs
743 		 * if we intend to reallocate.
744 		 */
745 		if (good_copies != mm->mm_children) {
746 			/*
747 			 * Always require at least one good copy.
748 			 *
749 			 * For ditto blocks (io_vd == NULL), require
750 			 * all copies to be good.
751 			 *
752 			 * XXX -- for replacing vdevs, there's no great answer.
753 			 * If the old device is really dead, we may not even
754 			 * be able to access it -- so we only want to
755 			 * require good writes to the new device.  But if
756 			 * the new device turns out to be flaky, we want
757 			 * to be able to detach it -- which requires all
758 			 * writes to the old device to have succeeded.
759 			 */
760 			if (good_copies == 0 || zio->io_vd == NULL)
761 				zio->io_error = vdev_mirror_worst_error(mm);
762 		}
763 		return;
764 	}
765 
766 	ASSERT(zio->io_type == ZIO_TYPE_READ);
767 
768 	/*
769 	 * Any Direct I/O read that has a checksum error must be treated as
770 	 * suspicious as the contents of the buffer could be getting
771 	 * manipulated while the I/O is taking place. The checksum verify error
772 	 * will be reported to the top-level Mirror VDEV.
773 	 *
774 	 * There will be no attampt at reading any additional data copies. If
775 	 * the buffer is still being manipulated while attempting to read from
776 	 * another child, there exists a possibly that the checksum could be
777 	 * verified as valid. However, the buffer contents could again get
778 	 * manipulated after verifying the checksum. This would lead to bad data
779 	 * being written out during self healing.
780 	 */
781 	if ((zio->io_flags & ZIO_FLAG_DIO_READ) &&
782 	    (zio->io_flags & ZIO_FLAG_DIO_CHKSUM_ERR)) {
783 		zio_dio_chksum_verify_error_report(zio);
784 		zio->io_error = vdev_mirror_worst_error(mm);
785 		ASSERT3U(zio->io_error, ==, ECKSUM);
786 		return;
787 	}
788 
789 	/*
790 	 * If we don't have a good copy yet, keep trying other children.
791 	 */
792 	if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
793 		ASSERT(c >= 0 && c < mm->mm_children);
794 		mc = &mm->mm_child[c];
795 		zio_vdev_io_redone(zio);
796 		zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
797 		    mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
798 		    ZIO_TYPE_READ, zio->io_priority, 0,
799 		    vdev_mirror_child_done, mc));
800 		return;
801 	}
802 
803 	if (zio->io_flags & ZIO_FLAG_SCRUB && !mm->mm_resilvering) {
804 		abd_t *best_abd = NULL;
805 		if (last_good_copy >= 0)
806 			best_abd = mm->mm_child[last_good_copy].mc_abd;
807 
808 		/*
809 		 * If we're scrubbing but don't have a BP available (because
810 		 * this vdev is under a raidz or draid vdev) then the best we
811 		 * can do is compare all of the copies read.  If they're not
812 		 * identical then return a checksum error and the most likely
813 		 * correct data.  The raidz code will issue a repair I/O if
814 		 * possible.
815 		 */
816 		if (zio->io_bp == NULL) {
817 			ASSERT(zio->io_vd->vdev_ops == &vdev_replacing_ops ||
818 			    zio->io_vd->vdev_ops == &vdev_spare_ops);
819 
820 			abd_t *pref_abd = NULL;
821 			for (c = 0; c < last_good_copy; c++) {
822 				mc = &mm->mm_child[c];
823 				if (mc->mc_error || !mc->mc_tried)
824 					continue;
825 
826 				if (abd_cmp(mc->mc_abd, best_abd) != 0)
827 					zio->io_error = SET_ERROR(ECKSUM);
828 
829 				/*
830 				 * The distributed spare is always prefered
831 				 * by vdev_mirror_child_select() so it's
832 				 * considered to be the best candidate.
833 				 */
834 				if (pref_abd == NULL &&
835 				    mc->mc_vd->vdev_ops ==
836 				    &vdev_draid_spare_ops)
837 					pref_abd = mc->mc_abd;
838 
839 				/*
840 				 * In the absence of a preferred copy, use
841 				 * the parent pointer to avoid a memory copy.
842 				 */
843 				if (mc->mc_abd == zio->io_abd)
844 					best_abd = mc->mc_abd;
845 			}
846 			if (pref_abd)
847 				best_abd = pref_abd;
848 		} else {
849 
850 			/*
851 			 * If we have a BP available, then checksums are
852 			 * already verified and we just need a buffer
853 			 * with valid data, preferring parent one to
854 			 * avoid a memory copy.
855 			 */
856 			for (c = 0; c < last_good_copy; c++) {
857 				mc = &mm->mm_child[c];
858 				if (mc->mc_error || !mc->mc_tried)
859 					continue;
860 				if (mc->mc_abd == zio->io_abd) {
861 					best_abd = mc->mc_abd;
862 					break;
863 				}
864 			}
865 		}
866 
867 		if (best_abd && best_abd != zio->io_abd)
868 			abd_copy(zio->io_abd, best_abd, zio->io_size);
869 		for (c = 0; c < mm->mm_children; c++) {
870 			mc = &mm->mm_child[c];
871 			if (mc->mc_abd != zio->io_abd)
872 				abd_free(mc->mc_abd);
873 			mc->mc_abd = NULL;
874 		}
875 	}
876 
877 	if (good_copies == 0) {
878 		zio->io_error = vdev_mirror_worst_error(mm);
879 		ASSERT(zio->io_error != 0);
880 	}
881 
882 	if (good_copies && spa_writeable(zio->io_spa) &&
883 	    (unexpected_errors ||
884 	    (zio->io_flags & ZIO_FLAG_RESILVER) ||
885 	    ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_resilvering))) {
886 		/*
887 		 * Use the good data we have in hand to repair damaged children.
888 		 */
889 		for (c = 0; c < mm->mm_children; c++) {
890 			/*
891 			 * Don't rewrite known good children.
892 			 * Not only is it unnecessary, it could
893 			 * actually be harmful: if the system lost
894 			 * power while rewriting the only good copy,
895 			 * there would be no good copies left!
896 			 */
897 			mc = &mm->mm_child[c];
898 
899 			if (mc->mc_error == 0) {
900 				vdev_ops_t *ops = mc->mc_vd->vdev_ops;
901 
902 				if (mc->mc_tried)
903 					continue;
904 				/*
905 				 * We didn't try this child.  We need to
906 				 * repair it if:
907 				 * 1. it's a scrub (in which case we have
908 				 * tried everything that was healthy)
909 				 *  - or -
910 				 * 2. it's an indirect or distributed spare
911 				 * vdev (in which case it could point to any
912 				 * other vdev, which might have a bad DTL)
913 				 *  - or -
914 				 * 3. the DTL indicates that this data is
915 				 * missing from this vdev
916 				 */
917 				if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
918 				    ops != &vdev_indirect_ops &&
919 				    ops != &vdev_draid_spare_ops &&
920 				    !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
921 				    zio->io_txg, 1))
922 					continue;
923 				mc->mc_error = SET_ERROR(ESTALE);
924 			}
925 
926 			zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
927 			    mc->mc_vd, mc->mc_offset,
928 			    zio->io_abd, zio->io_size, ZIO_TYPE_WRITE,
929 			    zio->io_priority == ZIO_PRIORITY_REBUILD ?
930 			    ZIO_PRIORITY_REBUILD : ZIO_PRIORITY_ASYNC_WRITE,
931 			    ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
932 			    ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
933 		}
934 	}
935 }
936 
937 static void
vdev_mirror_state_change(vdev_t * vd,int faulted,int degraded)938 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
939 {
940 	if (faulted == vd->vdev_children) {
941 		if (vdev_children_are_offline(vd)) {
942 			vdev_set_state(vd, B_FALSE, VDEV_STATE_OFFLINE,
943 			    VDEV_AUX_CHILDREN_OFFLINE);
944 		} else {
945 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
946 			    VDEV_AUX_NO_REPLICAS);
947 		}
948 	} else if (degraded + faulted != 0) {
949 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
950 	} else {
951 		vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
952 	}
953 }
954 
955 /*
956  * Return the maximum asize for a rebuild zio in the provided range.
957  */
958 static uint64_t
vdev_mirror_rebuild_asize(vdev_t * vd,uint64_t start,uint64_t asize,uint64_t max_segment)959 vdev_mirror_rebuild_asize(vdev_t *vd, uint64_t start, uint64_t asize,
960     uint64_t max_segment)
961 {
962 	(void) start;
963 
964 	uint64_t psize = MIN(P2ROUNDUP(max_segment, 1 << vd->vdev_ashift),
965 	    SPA_MAXBLOCKSIZE);
966 
967 	return (MIN(asize, vdev_psize_to_asize(vd, psize)));
968 }
969 
970 vdev_ops_t vdev_mirror_ops = {
971 	.vdev_op_init = NULL,
972 	.vdev_op_fini = NULL,
973 	.vdev_op_open = vdev_mirror_open,
974 	.vdev_op_close = vdev_mirror_close,
975 	.vdev_op_asize = vdev_default_asize,
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_asize = vdev_default_asize,
1001 	.vdev_op_min_asize = vdev_default_min_asize,
1002 	.vdev_op_min_alloc = NULL,
1003 	.vdev_op_io_start = vdev_mirror_io_start,
1004 	.vdev_op_io_done = vdev_mirror_io_done,
1005 	.vdev_op_state_change = vdev_mirror_state_change,
1006 	.vdev_op_need_resilver = vdev_default_need_resilver,
1007 	.vdev_op_hold = NULL,
1008 	.vdev_op_rele = NULL,
1009 	.vdev_op_remap = NULL,
1010 	.vdev_op_xlate = vdev_default_xlate,
1011 	.vdev_op_rebuild_asize = vdev_mirror_rebuild_asize,
1012 	.vdev_op_metaslab_init = NULL,
1013 	.vdev_op_config_generate = NULL,
1014 	.vdev_op_nparity = NULL,
1015 	.vdev_op_ndisks = NULL,
1016 	.vdev_op_type = VDEV_TYPE_REPLACING,	/* name of this vdev type */
1017 	.vdev_op_leaf = B_FALSE			/* not a leaf vdev */
1018 };
1019 
1020 vdev_ops_t vdev_spare_ops = {
1021 	.vdev_op_init = NULL,
1022 	.vdev_op_fini = NULL,
1023 	.vdev_op_open = vdev_mirror_open,
1024 	.vdev_op_close = vdev_mirror_close,
1025 	.vdev_op_asize = vdev_default_asize,
1026 	.vdev_op_min_asize = vdev_default_min_asize,
1027 	.vdev_op_min_alloc = NULL,
1028 	.vdev_op_io_start = vdev_mirror_io_start,
1029 	.vdev_op_io_done = vdev_mirror_io_done,
1030 	.vdev_op_state_change = vdev_mirror_state_change,
1031 	.vdev_op_need_resilver = vdev_default_need_resilver,
1032 	.vdev_op_hold = NULL,
1033 	.vdev_op_rele = NULL,
1034 	.vdev_op_remap = NULL,
1035 	.vdev_op_xlate = vdev_default_xlate,
1036 	.vdev_op_rebuild_asize = vdev_mirror_rebuild_asize,
1037 	.vdev_op_metaslab_init = NULL,
1038 	.vdev_op_config_generate = NULL,
1039 	.vdev_op_nparity = NULL,
1040 	.vdev_op_ndisks = NULL,
1041 	.vdev_op_type = VDEV_TYPE_SPARE,	/* name of this vdev type */
1042 	.vdev_op_leaf = B_FALSE			/* not a leaf vdev */
1043 };
1044 
1045 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_inc, INT, ZMOD_RW,
1046 	"Rotating media load increment for non-seeking I/Os");
1047 
1048 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_seek_inc, INT,
1049 	ZMOD_RW, "Rotating media load increment for seeking I/Os");
1050 
1051 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, rotating_seek_offset, INT,
1052 	ZMOD_RW,
1053 	"Offset in bytes from the last I/O which triggers "
1054 	"a reduced rotating media seek increment");
1055 
1056 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, non_rotating_inc, INT,
1057 	ZMOD_RW, "Non-rotating media load increment for non-seeking I/Os");
1058 
1059 ZFS_MODULE_PARAM(zfs_vdev_mirror, zfs_vdev_mirror_, non_rotating_seek_inc, INT,
1060 	ZMOD_RW, "Non-rotating media load increment for seeking I/Os");
1061