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.
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.
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]
22 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
30 #include <sys/zfs_context.h>
32 #include <sys/spa_impl.h>
33 #include <sys/dsl_pool.h>
34 #include <sys/dsl_scan.h>
35 #include <sys/vdev_impl.h>
38 #include <sys/fs/zfs.h>
43 static kstat_t *mirror_ksp = NULL;
45 typedef struct mirror_stats {
46 kstat_named_t vdev_mirror_stat_rotating_linear;
47 kstat_named_t vdev_mirror_stat_rotating_offset;
48 kstat_named_t vdev_mirror_stat_rotating_seek;
49 kstat_named_t vdev_mirror_stat_non_rotating_linear;
50 kstat_named_t vdev_mirror_stat_non_rotating_seek;
52 kstat_named_t vdev_mirror_stat_preferred_found;
53 kstat_named_t vdev_mirror_stat_preferred_not_found;
56 static mirror_stats_t mirror_stats = {
57 /* New I/O follows directly the last I/O */
58 { "rotating_linear", KSTAT_DATA_UINT64 },
59 /* New I/O is within zfs_vdev_mirror_rotating_seek_offset of the last */
60 { "rotating_offset", KSTAT_DATA_UINT64 },
61 /* New I/O requires random seek */
62 { "rotating_seek", KSTAT_DATA_UINT64 },
63 /* New I/O follows directly the last I/O (nonrot) */
64 { "non_rotating_linear", KSTAT_DATA_UINT64 },
65 /* New I/O requires random seek (nonrot) */
66 { "non_rotating_seek", KSTAT_DATA_UINT64 },
67 /* Preferred child vdev found */
68 { "preferred_found", KSTAT_DATA_UINT64 },
69 /* Preferred child vdev not found or equal load */
70 { "preferred_not_found", KSTAT_DATA_UINT64 },
74 #define MIRROR_STAT(stat) (mirror_stats.stat.value.ui64)
75 #define MIRROR_INCR(stat, val) atomic_add_64(&MIRROR_STAT(stat), val)
76 #define MIRROR_BUMP(stat) MIRROR_INCR(stat, 1)
79 vdev_mirror_stat_init(void)
81 mirror_ksp = kstat_create("zfs", 0, "vdev_mirror_stats",
82 "misc", KSTAT_TYPE_NAMED,
83 sizeof (mirror_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
84 if (mirror_ksp != NULL) {
85 mirror_ksp->ks_data = &mirror_stats;
86 kstat_install(mirror_ksp);
91 vdev_mirror_stat_fini(void)
93 if (mirror_ksp != NULL) {
94 kstat_delete(mirror_ksp);
100 * Virtual device vector for mirroring.
103 typedef struct mirror_child {
110 uint8_t mc_speculative;
113 typedef struct mirror_map {
115 int mm_preferred_cnt;
117 boolean_t mm_resilvering;
119 mirror_child_t mm_child[];
122 static int vdev_mirror_shift = 21;
125 * The load configuration settings below are tuned by default for
126 * the case where all devices are of the same rotational type.
128 * If there is a mixture of rotating and non-rotating media, setting
129 * zfs_vdev_mirror_non_rotating_seek_inc to 0 may well provide better results
130 * as it will direct more reads to the non-rotating vdevs which are more likely
131 * to have a higher performance.
134 /* Rotating media load calculation configuration. */
135 static int zfs_vdev_mirror_rotating_inc = 0;
136 static int zfs_vdev_mirror_rotating_seek_inc = 5;
137 static int zfs_vdev_mirror_rotating_seek_offset = 1 * 1024 * 1024;
139 /* Non-rotating media load calculation configuration. */
140 static int zfs_vdev_mirror_non_rotating_inc = 0;
141 static int zfs_vdev_mirror_non_rotating_seek_inc = 1;
144 vdev_mirror_map_size(int children)
146 return (offsetof(mirror_map_t, mm_child[children]) +
147 sizeof (int) * children);
150 static inline mirror_map_t *
151 vdev_mirror_map_alloc(int children, boolean_t resilvering, boolean_t root)
155 mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP);
156 mm->mm_children = children;
157 mm->mm_resilvering = resilvering;
159 mm->mm_preferred = (int *)((uintptr_t)mm +
160 offsetof(mirror_map_t, mm_child[children]));
166 vdev_mirror_map_free(zio_t *zio)
168 mirror_map_t *mm = zio->io_vsd;
170 kmem_free(mm, vdev_mirror_map_size(mm->mm_children));
173 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
174 .vsd_free = vdev_mirror_map_free,
175 .vsd_cksum_report = zio_vsd_default_cksum_report
179 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
181 uint64_t last_offset;
185 /* All DVAs have equal weight at the root. */
190 * We don't return INT_MAX if the device is resilvering i.e.
191 * vdev_resilver_txg != 0 as when tested performance was slightly
192 * worse overall when resilvering with compared to without.
195 /* Fix zio_offset for leaf vdevs */
196 if (vd->vdev_ops->vdev_op_leaf)
197 zio_offset += VDEV_LABEL_START_SIZE;
199 /* Standard load based on pending queue length. */
200 load = vdev_queue_length(vd);
201 last_offset = vdev_queue_last_offset(vd);
203 if (vd->vdev_nonrot) {
204 /* Non-rotating media. */
205 if (last_offset == zio_offset) {
206 MIRROR_BUMP(vdev_mirror_stat_non_rotating_linear);
207 return (load + zfs_vdev_mirror_non_rotating_inc);
211 * Apply a seek penalty even for non-rotating devices as
212 * sequential I/O's can be aggregated into fewer operations on
213 * the device, thus avoiding unnecessary per-command overhead
214 * and boosting performance.
216 MIRROR_BUMP(vdev_mirror_stat_non_rotating_seek);
217 return (load + zfs_vdev_mirror_non_rotating_seek_inc);
220 /* Rotating media I/O's which directly follow the last I/O. */
221 if (last_offset == zio_offset) {
222 MIRROR_BUMP(vdev_mirror_stat_rotating_linear);
223 return (load + zfs_vdev_mirror_rotating_inc);
227 * Apply half the seek increment to I/O's within seek offset
228 * of the last I/O issued to this vdev as they should incur less
229 * of a seek increment.
231 offset_diff = (int64_t)(last_offset - zio_offset);
232 if (ABS(offset_diff) < zfs_vdev_mirror_rotating_seek_offset) {
233 MIRROR_BUMP(vdev_mirror_stat_rotating_offset);
234 return (load + (zfs_vdev_mirror_rotating_seek_inc / 2));
237 /* Apply the full seek increment to all other I/O's. */
238 MIRROR_BUMP(vdev_mirror_stat_rotating_seek);
239 return (load + zfs_vdev_mirror_rotating_seek_inc);
243 * Avoid inlining the function to keep vdev_mirror_io_start(), which
244 * is this functions only caller, as small as possible on the stack.
246 noinline static mirror_map_t *
247 vdev_mirror_map_init(zio_t *zio)
249 mirror_map_t *mm = NULL;
251 vdev_t *vd = zio->io_vd;
255 dva_t *dva = zio->io_bp->blk_dva;
256 spa_t *spa = zio->io_spa;
257 dva_t dva_copy[SPA_DVAS_PER_BP];
259 c = BP_GET_NDVAS(zio->io_bp);
262 * If we do not trust the pool config, some DVAs might be
263 * invalid or point to vdevs that do not exist. We skip them.
265 if (!spa_trust_config(spa)) {
266 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
268 for (int i = 0; i < c; i++) {
269 if (zfs_dva_valid(spa, &dva[i], zio->io_bp))
270 dva_copy[j++] = dva[i];
274 zio->io_error = ENXIO;
283 mm = vdev_mirror_map_alloc(c, B_FALSE, B_TRUE);
284 for (c = 0; c < mm->mm_children; c++) {
285 mc = &mm->mm_child[c];
287 mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
288 mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
292 * If we are resilvering, then we should handle scrub reads
293 * differently; we shouldn't issue them to the resilvering
294 * device because it might not have those blocks.
296 * We are resilvering iff:
297 * 1) We are a replacing vdev (ie our name is "replacing-1" or
298 * "spare-1" or something like that), and
299 * 2) The pool is currently being resilvered.
301 * We cannot simply check vd->vdev_resilver_txg, because it's
302 * not set in this path.
304 * Nor can we just check our vdev_ops; there are cases (such as
305 * when a user types "zpool replace pool odev spare_dev" and
306 * spare_dev is in the spare list, or when a spare device is
307 * automatically used to replace a DEGRADED device) when
308 * resilvering is complete but both the original vdev and the
309 * spare vdev remain in the pool. That behavior is intentional.
310 * It helps implement the policy that a spare should be
311 * automatically removed from the pool after the user replaces
312 * the device that originally failed.
314 * If a spa load is in progress, then spa_dsl_pool may be
315 * uninitialized. But we shouldn't be resilvering during a spa
318 boolean_t replacing = (vd->vdev_ops == &vdev_replacing_ops ||
319 vd->vdev_ops == &vdev_spare_ops) &&
320 spa_load_state(vd->vdev_spa) == SPA_LOAD_NONE &&
321 dsl_scan_resilvering(vd->vdev_spa->spa_dsl_pool);
322 mm = vdev_mirror_map_alloc(vd->vdev_children, replacing,
324 for (c = 0; c < mm->mm_children; c++) {
325 mc = &mm->mm_child[c];
326 mc->mc_vd = vd->vdev_child[c];
327 mc->mc_offset = zio->io_offset;
332 zio->io_vsd_ops = &vdev_mirror_vsd_ops;
337 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
343 if (vd->vdev_children == 0) {
344 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
345 return (SET_ERROR(EINVAL));
348 vdev_open_children(vd);
350 for (int c = 0; c < vd->vdev_children; c++) {
351 vdev_t *cvd = vd->vdev_child[c];
353 if (cvd->vdev_open_error) {
354 lasterror = cvd->vdev_open_error;
359 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
360 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
361 *ashift = MAX(*ashift, cvd->vdev_ashift);
364 if (numerrors == vd->vdev_children) {
365 if (vdev_children_are_offline(vd))
366 vd->vdev_stat.vs_aux = VDEV_AUX_CHILDREN_OFFLINE;
368 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
376 vdev_mirror_close(vdev_t *vd)
378 for (int c = 0; c < vd->vdev_children; c++)
379 vdev_close(vd->vdev_child[c]);
383 vdev_mirror_child_done(zio_t *zio)
385 mirror_child_t *mc = zio->io_private;
387 mc->mc_error = zio->io_error;
393 vdev_mirror_scrub_done(zio_t *zio)
395 mirror_child_t *mc = zio->io_private;
397 if (zio->io_error == 0) {
399 zio_link_t *zl = NULL;
401 mutex_enter(&zio->io_lock);
402 while ((pio = zio_walk_parents(zio, &zl)) != NULL) {
403 mutex_enter(&pio->io_lock);
404 ASSERT3U(zio->io_size, >=, pio->io_size);
405 abd_copy(pio->io_abd, zio->io_abd, pio->io_size);
406 mutex_exit(&pio->io_lock);
408 mutex_exit(&zio->io_lock);
411 abd_free(zio->io_abd);
413 mc->mc_error = zio->io_error;
419 * Check the other, lower-index DVAs to see if they're on the same
420 * vdev as the child we picked. If they are, use them since they
421 * are likely to have been allocated from the primary metaslab in
422 * use at the time, and hence are more likely to have locality with
426 vdev_mirror_dva_select(zio_t *zio, int p)
428 dva_t *dva = zio->io_bp->blk_dva;
429 mirror_map_t *mm = zio->io_vsd;
433 preferred = mm->mm_preferred[p];
434 for (p--; p >= 0; p--) {
435 c = mm->mm_preferred[p];
436 if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
443 vdev_mirror_preferred_child_randomize(zio_t *zio)
445 mirror_map_t *mm = zio->io_vsd;
449 p = spa_get_random(mm->mm_preferred_cnt);
450 return (vdev_mirror_dva_select(zio, p));
454 * To ensure we don't always favour the first matching vdev,
455 * which could lead to wear leveling issues on SSD's, we
456 * use the I/O offset as a pseudo random seed into the vdevs
457 * which have the lowest load.
459 p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt;
460 return (mm->mm_preferred[p]);
464 * Try to find a vdev whose DTL doesn't contain the block we want to read
465 * prefering vdevs based on determined load.
467 * Try to find a child whose DTL doesn't contain the block we want to read.
468 * If we can't, try the read on any vdev we haven't already tried.
471 vdev_mirror_child_select(zio_t *zio)
473 mirror_map_t *mm = zio->io_vsd;
474 uint64_t txg = zio->io_txg;
477 ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
479 lowest_load = INT_MAX;
480 mm->mm_preferred_cnt = 0;
481 for (c = 0; c < mm->mm_children; c++) {
484 mc = &mm->mm_child[c];
485 if (mc->mc_tried || mc->mc_skipped)
488 if (mc->mc_vd == NULL || !vdev_readable(mc->mc_vd)) {
489 mc->mc_error = SET_ERROR(ENXIO);
490 mc->mc_tried = 1; /* don't even try */
495 if (vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1)) {
496 mc->mc_error = SET_ERROR(ESTALE);
498 mc->mc_speculative = 1;
502 mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
503 if (mc->mc_load > lowest_load)
506 if (mc->mc_load < lowest_load) {
507 lowest_load = mc->mc_load;
508 mm->mm_preferred_cnt = 0;
510 mm->mm_preferred[mm->mm_preferred_cnt] = c;
511 mm->mm_preferred_cnt++;
514 if (mm->mm_preferred_cnt == 1) {
515 MIRROR_BUMP(vdev_mirror_stat_preferred_found);
516 return (mm->mm_preferred[0]);
519 if (mm->mm_preferred_cnt > 1) {
520 MIRROR_BUMP(vdev_mirror_stat_preferred_not_found);
521 return (vdev_mirror_preferred_child_randomize(zio));
525 * Every device is either missing or has this txg in its DTL.
526 * Look for any child we haven't already tried before giving up.
528 for (c = 0; c < mm->mm_children; c++) {
529 if (!mm->mm_child[c].mc_tried)
534 * Every child failed. There's no place left to look.
540 vdev_mirror_io_start(zio_t *zio)
546 mm = vdev_mirror_map_init(zio);
549 ASSERT(!spa_trust_config(zio->io_spa));
550 ASSERT(zio->io_type == ZIO_TYPE_READ);
555 if (zio->io_type == ZIO_TYPE_READ) {
556 if (zio->io_bp != NULL &&
557 (zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_resilvering) {
559 * For scrubbing reads (if we can verify the
560 * checksum here, as indicated by io_bp being
561 * non-NULL) we need to allocate a read buffer for
562 * each child and issue reads to all children. If
563 * any child succeeds, it will copy its data into
564 * zio->io_data in vdev_mirror_scrub_done.
566 for (c = 0; c < mm->mm_children; c++) {
567 mc = &mm->mm_child[c];
568 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
569 mc->mc_vd, mc->mc_offset,
570 abd_alloc_sametype(zio->io_abd,
571 zio->io_size), zio->io_size,
572 zio->io_type, zio->io_priority, 0,
573 vdev_mirror_scrub_done, mc));
579 * For normal reads just pick one child.
581 c = vdev_mirror_child_select(zio);
584 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
587 * Writes go to all children.
590 children = mm->mm_children;
594 mc = &mm->mm_child[c];
595 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
596 mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
597 zio->io_type, zio->io_priority, 0,
598 vdev_mirror_child_done, mc));
606 vdev_mirror_worst_error(mirror_map_t *mm)
608 int error[2] = { 0, 0 };
610 for (int c = 0; c < mm->mm_children; c++) {
611 mirror_child_t *mc = &mm->mm_child[c];
612 int s = mc->mc_speculative;
613 error[s] = zio_worst_error(error[s], mc->mc_error);
616 return (error[0] ? error[0] : error[1]);
620 vdev_mirror_io_done(zio_t *zio)
622 mirror_map_t *mm = zio->io_vsd;
626 int unexpected_errors = 0;
631 for (c = 0; c < mm->mm_children; c++) {
632 mc = &mm->mm_child[c];
637 } else if (mc->mc_tried) {
642 if (zio->io_type == ZIO_TYPE_WRITE) {
644 * XXX -- for now, treat partial writes as success.
646 * Now that we support write reallocation, it would be better
647 * to treat partial failure as real failure unless there are
648 * no non-degraded top-level vdevs left, and not update DTLs
649 * if we intend to reallocate.
652 if (good_copies != mm->mm_children) {
654 * Always require at least one good copy.
656 * For ditto blocks (io_vd == NULL), require
657 * all copies to be good.
659 * XXX -- for replacing vdevs, there's no great answer.
660 * If the old device is really dead, we may not even
661 * be able to access it -- so we only want to
662 * require good writes to the new device. But if
663 * the new device turns out to be flaky, we want
664 * to be able to detach it -- which requires all
665 * writes to the old device to have succeeded.
667 if (good_copies == 0 || zio->io_vd == NULL)
668 zio->io_error = vdev_mirror_worst_error(mm);
673 ASSERT(zio->io_type == ZIO_TYPE_READ);
676 * If we don't have a good copy yet, keep trying other children.
679 if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
680 ASSERT(c >= 0 && c < mm->mm_children);
681 mc = &mm->mm_child[c];
682 zio_vdev_io_redone(zio);
683 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
684 mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
685 ZIO_TYPE_READ, zio->io_priority, 0,
686 vdev_mirror_child_done, mc));
691 if (good_copies == 0) {
692 zio->io_error = vdev_mirror_worst_error(mm);
693 ASSERT(zio->io_error != 0);
696 if (good_copies && spa_writeable(zio->io_spa) &&
697 (unexpected_errors ||
698 (zio->io_flags & ZIO_FLAG_RESILVER) ||
699 ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_resilvering))) {
701 * Use the good data we have in hand to repair damaged children.
703 for (c = 0; c < mm->mm_children; c++) {
705 * Don't rewrite known good children.
706 * Not only is it unnecessary, it could
707 * actually be harmful: if the system lost
708 * power while rewriting the only good copy,
709 * there would be no good copies left!
711 mc = &mm->mm_child[c];
713 if (mc->mc_error == 0) {
717 * We didn't try this child. We need to
719 * 1. it's a scrub (in which case we have
720 * tried everything that was healthy)
722 * 2. it's an indirect vdev (in which case
723 * it could point to any other vdev, which
724 * might have a bad DTL)
726 * 3. the DTL indicates that this data is
727 * missing from this vdev
729 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
730 mc->mc_vd->vdev_ops != &vdev_indirect_ops &&
731 !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
734 mc->mc_error = SET_ERROR(ESTALE);
737 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
738 mc->mc_vd, mc->mc_offset,
739 zio->io_abd, zio->io_size,
740 ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
741 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
742 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
748 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
750 if (faulted == vd->vdev_children) {
751 if (vdev_children_are_offline(vd)) {
752 vdev_set_state(vd, B_FALSE, VDEV_STATE_OFFLINE,
753 VDEV_AUX_CHILDREN_OFFLINE);
755 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
756 VDEV_AUX_NO_REPLICAS);
758 } else if (degraded + faulted != 0) {
759 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
761 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
765 vdev_ops_t vdev_mirror_ops = {
769 vdev_mirror_io_start,
771 vdev_mirror_state_change,
777 VDEV_TYPE_MIRROR, /* name of this vdev type */
778 B_FALSE /* not a leaf vdev */
781 vdev_ops_t vdev_replacing_ops = {
785 vdev_mirror_io_start,
787 vdev_mirror_state_change,
793 VDEV_TYPE_REPLACING, /* name of this vdev type */
794 B_FALSE /* not a leaf vdev */
797 vdev_ops_t vdev_spare_ops = {
801 vdev_mirror_io_start,
803 vdev_mirror_state_change,
809 VDEV_TYPE_SPARE, /* name of this vdev type */
810 B_FALSE /* not a leaf vdev */
815 module_param(zfs_vdev_mirror_rotating_inc, int, 0644);
816 MODULE_PARM_DESC(zfs_vdev_mirror_rotating_inc,
817 "Rotating media load increment for non-seeking I/O's");
819 module_param(zfs_vdev_mirror_rotating_seek_inc, int, 0644);
820 MODULE_PARM_DESC(zfs_vdev_mirror_rotating_seek_inc,
821 "Rotating media load increment for seeking I/O's");
823 module_param(zfs_vdev_mirror_rotating_seek_offset, int, 0644);
825 MODULE_PARM_DESC(zfs_vdev_mirror_rotating_seek_offset,
826 "Offset in bytes from the last I/O which "
827 "triggers a reduced rotating media seek increment");
829 module_param(zfs_vdev_mirror_non_rotating_inc, int, 0644);
830 MODULE_PARM_DESC(zfs_vdev_mirror_non_rotating_inc,
831 "Non-rotating media load increment for non-seeking I/O's");
833 module_param(zfs_vdev_mirror_non_rotating_seek_inc, int, 0644);
834 MODULE_PARM_DESC(zfs_vdev_mirror_non_rotating_seek_inc,
835 "Non-rotating media load increment for seeking I/O's");