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/vdev_impl.h>
35 #include <sys/fs/zfs.h>
38 * Virtual device vector for mirroring.
41 typedef struct mirror_child {
48 uint8_t mc_speculative;
51 typedef struct mirror_map {
55 boolean_t mm_replacing;
57 mirror_child_t mm_child[];
60 static int vdev_mirror_shift = 21;
63 SYSCTL_DECL(_vfs_zfs_vdev);
64 static SYSCTL_NODE(_vfs_zfs_vdev, OID_AUTO, mirror, CTLFLAG_RD, 0,
69 * The load configuration settings below are tuned by default for
70 * the case where all devices are of the same rotational type.
72 * If there is a mixture of rotating and non-rotating media, setting
73 * non_rotating_seek_inc to 0 may well provide better results as it
74 * will direct more reads to the non-rotating vdevs which are more
75 * likely to have a higher performance.
78 /* Rotating media load calculation configuration. */
79 static int rotating_inc = 0;
81 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, rotating_inc, CTLFLAG_RWTUN,
82 &rotating_inc, 0, "Rotating media load increment for non-seeking I/O's");
85 static int rotating_seek_inc = 5;
87 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, rotating_seek_inc, CTLFLAG_RWTUN,
88 &rotating_seek_inc, 0, "Rotating media load increment for seeking I/O's");
91 static int rotating_seek_offset = 1 * 1024 * 1024;
93 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, rotating_seek_offset, CTLFLAG_RWTUN,
94 &rotating_seek_offset, 0, "Offset in bytes from the last I/O which "
95 "triggers a reduced rotating media seek increment");
98 /* Non-rotating media load calculation configuration. */
99 static int non_rotating_inc = 0;
101 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, non_rotating_inc, CTLFLAG_RWTUN,
102 &non_rotating_inc, 0,
103 "Non-rotating media load increment for non-seeking I/O's");
106 static int non_rotating_seek_inc = 1;
108 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, non_rotating_seek_inc, CTLFLAG_RWTUN,
109 &non_rotating_seek_inc, 0,
110 "Non-rotating media load increment for seeking I/O's");
115 vdev_mirror_map_size(int children)
117 return (offsetof(mirror_map_t, mm_child[children]) +
118 sizeof(int) * children);
121 static inline mirror_map_t *
122 vdev_mirror_map_alloc(int children, boolean_t replacing, boolean_t root)
126 mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP);
127 mm->mm_children = children;
128 mm->mm_replacing = replacing;
130 mm->mm_preferred = (int *)((uintptr_t)mm +
131 offsetof(mirror_map_t, mm_child[children]));
137 vdev_mirror_map_free(zio_t *zio)
139 mirror_map_t *mm = zio->io_vsd;
141 kmem_free(mm, vdev_mirror_map_size(mm->mm_children));
144 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
145 vdev_mirror_map_free,
146 zio_vsd_default_cksum_report
150 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
155 /* All DVAs have equal weight at the root. */
160 * We don't return INT_MAX if the device is resilvering i.e.
161 * vdev_resilver_txg != 0 as when tested performance was slightly
162 * worse overall when resilvering with compared to without.
165 /* Standard load based on pending queue length. */
166 load = vdev_queue_length(vd);
167 lastoffset = vdev_queue_lastoffset(vd);
169 if (vd->vdev_rotation_rate == VDEV_RATE_NON_ROTATING) {
170 /* Non-rotating media. */
171 if (lastoffset == zio_offset)
172 return (load + non_rotating_inc);
175 * Apply a seek penalty even for non-rotating devices as
176 * sequential I/O'a can be aggregated into fewer operations
177 * on the device, thus avoiding unnecessary per-command
178 * overhead and boosting performance.
180 return (load + non_rotating_seek_inc);
183 /* Rotating media I/O's which directly follow the last I/O. */
184 if (lastoffset == zio_offset)
185 return (load + rotating_inc);
188 * Apply half the seek increment to I/O's within seek offset
189 * of the last I/O queued to this vdev as they should incure less
190 * of a seek increment.
192 if (ABS(lastoffset - zio_offset) < rotating_seek_offset)
193 return (load + (rotating_seek_inc / 2));
195 /* Apply the full seek increment to all other I/O's. */
196 return (load + rotating_seek_inc);
200 static mirror_map_t *
201 vdev_mirror_map_init(zio_t *zio)
203 mirror_map_t *mm = NULL;
205 vdev_t *vd = zio->io_vd;
209 dva_t *dva = zio->io_bp->blk_dva;
210 spa_t *spa = zio->io_spa;
212 mm = vdev_mirror_map_alloc(BP_GET_NDVAS(zio->io_bp), B_FALSE,
214 for (c = 0; c < mm->mm_children; c++) {
215 mc = &mm->mm_child[c];
216 mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
217 mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
220 mm = vdev_mirror_map_alloc(vd->vdev_children,
221 (vd->vdev_ops == &vdev_replacing_ops ||
222 vd->vdev_ops == &vdev_spare_ops), B_FALSE);
223 for (c = 0; c < mm->mm_children; c++) {
224 mc = &mm->mm_child[c];
225 mc->mc_vd = vd->vdev_child[c];
226 mc->mc_offset = zio->io_offset;
231 zio->io_vsd_ops = &vdev_mirror_vsd_ops;
236 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
237 uint64_t *logical_ashift, uint64_t *physical_ashift)
242 if (vd->vdev_children == 0) {
243 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
244 return (SET_ERROR(EINVAL));
247 vdev_open_children(vd);
249 for (int c = 0; c < vd->vdev_children; c++) {
250 vdev_t *cvd = vd->vdev_child[c];
252 if (cvd->vdev_open_error) {
253 lasterror = cvd->vdev_open_error;
258 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
259 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
260 *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift);
261 *physical_ashift = MAX(*physical_ashift,
262 cvd->vdev_physical_ashift);
265 if (numerrors == vd->vdev_children) {
266 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
274 vdev_mirror_close(vdev_t *vd)
276 for (int c = 0; c < vd->vdev_children; c++)
277 vdev_close(vd->vdev_child[c]);
281 vdev_mirror_child_done(zio_t *zio)
283 mirror_child_t *mc = zio->io_private;
285 mc->mc_error = zio->io_error;
291 vdev_mirror_scrub_done(zio_t *zio)
293 mirror_child_t *mc = zio->io_private;
295 if (zio->io_error == 0) {
297 zio_link_t *zl = NULL;
299 mutex_enter(&zio->io_lock);
300 while ((pio = zio_walk_parents(zio, &zl)) != NULL) {
301 mutex_enter(&pio->io_lock);
302 ASSERT3U(zio->io_size, >=, pio->io_size);
303 abd_copy(pio->io_abd, zio->io_abd, pio->io_size);
304 mutex_exit(&pio->io_lock);
306 mutex_exit(&zio->io_lock);
308 abd_free(zio->io_abd);
310 mc->mc_error = zio->io_error;
316 * Check the other, lower-index DVAs to see if they're on the same
317 * vdev as the child we picked. If they are, use them since they
318 * are likely to have been allocated from the primary metaslab in
319 * use at the time, and hence are more likely to have locality with
323 vdev_mirror_dva_select(zio_t *zio, int p)
325 dva_t *dva = zio->io_bp->blk_dva;
326 mirror_map_t *mm = zio->io_vsd;
330 preferred = mm->mm_preferred[p];
331 for (p-- ; p >= 0; p--) {
332 c = mm->mm_preferred[p];
333 if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
340 vdev_mirror_preferred_child_randomize(zio_t *zio)
342 mirror_map_t *mm = zio->io_vsd;
346 p = spa_get_random(mm->mm_preferred_cnt);
347 return (vdev_mirror_dva_select(zio, p));
351 * To ensure we don't always favour the first matching vdev,
352 * which could lead to wear leveling issues on SSD's, we
353 * use the I/O offset as a pseudo random seed into the vdevs
354 * which have the lowest load.
356 p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt;
357 return (mm->mm_preferred[p]);
361 * Try to find a vdev whose DTL doesn't contain the block we want to read
362 * prefering vdevs based on determined load.
364 * If we can't, try the read on any vdev we haven't already tried.
367 vdev_mirror_child_select(zio_t *zio)
369 mirror_map_t *mm = zio->io_vsd;
370 uint64_t txg = zio->io_txg;
373 ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
375 lowest_load = INT_MAX;
376 mm->mm_preferred_cnt = 0;
377 for (c = 0; c < mm->mm_children; c++) {
380 mc = &mm->mm_child[c];
381 if (mc->mc_tried || mc->mc_skipped)
384 if (!vdev_readable(mc->mc_vd)) {
385 mc->mc_error = SET_ERROR(ENXIO);
386 mc->mc_tried = 1; /* don't even try */
391 if (vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1)) {
392 mc->mc_error = SET_ERROR(ESTALE);
394 mc->mc_speculative = 1;
398 mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
399 if (mc->mc_load > lowest_load)
402 if (mc->mc_load < lowest_load) {
403 lowest_load = mc->mc_load;
404 mm->mm_preferred_cnt = 0;
406 mm->mm_preferred[mm->mm_preferred_cnt] = c;
407 mm->mm_preferred_cnt++;
410 if (mm->mm_preferred_cnt == 1) {
411 vdev_queue_register_lastoffset(
412 mm->mm_child[mm->mm_preferred[0]].mc_vd, zio);
413 return (mm->mm_preferred[0]);
416 if (mm->mm_preferred_cnt > 1) {
417 int c = vdev_mirror_preferred_child_randomize(zio);
419 vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd, zio);
424 * Every device is either missing or has this txg in its DTL.
425 * Look for any child we haven't already tried before giving up.
427 for (c = 0; c < mm->mm_children; c++) {
428 if (!mm->mm_child[c].mc_tried) {
429 vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd,
436 * Every child failed. There's no place left to look.
442 vdev_mirror_io_start(zio_t *zio)
448 mm = vdev_mirror_map_init(zio);
450 if (zio->io_type == ZIO_TYPE_READ) {
451 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing &&
452 mm->mm_children > 1) {
454 * For scrubbing reads we need to allocate a read
455 * buffer for each child and issue reads to all
456 * children. If any child succeeds, it will copy its
457 * data into zio->io_data in vdev_mirror_scrub_done.
459 for (c = 0; c < mm->mm_children; c++) {
460 mc = &mm->mm_child[c];
461 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
462 mc->mc_vd, mc->mc_offset,
463 abd_alloc_sametype(zio->io_abd,
464 zio->io_size), zio->io_size,
465 zio->io_type, zio->io_priority, 0,
466 vdev_mirror_scrub_done, mc));
472 * For normal reads just pick one child.
474 c = vdev_mirror_child_select(zio);
477 ASSERT(zio->io_type == ZIO_TYPE_WRITE ||
478 zio->io_type == ZIO_TYPE_FREE);
481 * Writes and frees go to all children.
484 children = mm->mm_children;
488 mc = &mm->mm_child[c];
489 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
490 mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
491 zio->io_type, zio->io_priority, 0,
492 vdev_mirror_child_done, mc));
500 vdev_mirror_worst_error(mirror_map_t *mm)
502 int error[2] = { 0, 0 };
504 for (int c = 0; c < mm->mm_children; c++) {
505 mirror_child_t *mc = &mm->mm_child[c];
506 int s = mc->mc_speculative;
507 error[s] = zio_worst_error(error[s], mc->mc_error);
510 return (error[0] ? error[0] : error[1]);
514 vdev_mirror_io_done(zio_t *zio)
516 mirror_map_t *mm = zio->io_vsd;
520 int unexpected_errors = 0;
522 for (c = 0; c < mm->mm_children; c++) {
523 mc = &mm->mm_child[c];
528 } else if (mc->mc_tried) {
533 if (zio->io_type == ZIO_TYPE_WRITE) {
535 * XXX -- for now, treat partial writes as success.
537 * Now that we support write reallocation, it would be better
538 * to treat partial failure as real failure unless there are
539 * no non-degraded top-level vdevs left, and not update DTLs
540 * if we intend to reallocate.
543 if (good_copies != mm->mm_children) {
545 * Always require at least one good copy.
547 * For ditto blocks (io_vd == NULL), require
548 * all copies to be good.
550 * XXX -- for replacing vdevs, there's no great answer.
551 * If the old device is really dead, we may not even
552 * be able to access it -- so we only want to
553 * require good writes to the new device. But if
554 * the new device turns out to be flaky, we want
555 * to be able to detach it -- which requires all
556 * writes to the old device to have succeeded.
558 if (good_copies == 0 || zio->io_vd == NULL)
559 zio->io_error = vdev_mirror_worst_error(mm);
562 } else if (zio->io_type == ZIO_TYPE_FREE) {
566 ASSERT(zio->io_type == ZIO_TYPE_READ);
569 * If we don't have a good copy yet, keep trying other children.
572 if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
573 ASSERT(c >= 0 && c < mm->mm_children);
574 mc = &mm->mm_child[c];
575 zio_vdev_io_redone(zio);
576 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
577 mc->mc_vd, mc->mc_offset, zio->io_abd, zio->io_size,
578 ZIO_TYPE_READ, zio->io_priority, 0,
579 vdev_mirror_child_done, mc));
584 if (good_copies == 0) {
585 zio->io_error = vdev_mirror_worst_error(mm);
586 ASSERT(zio->io_error != 0);
589 if (good_copies && spa_writeable(zio->io_spa) &&
590 (unexpected_errors ||
591 (zio->io_flags & ZIO_FLAG_RESILVER) ||
592 ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) {
594 * Use the good data we have in hand to repair damaged children.
596 for (c = 0; c < mm->mm_children; c++) {
598 * Don't rewrite known good children.
599 * Not only is it unnecessary, it could
600 * actually be harmful: if the system lost
601 * power while rewriting the only good copy,
602 * there would be no good copies left!
604 mc = &mm->mm_child[c];
606 if (mc->mc_error == 0) {
609 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
610 !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
613 mc->mc_error = SET_ERROR(ESTALE);
616 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
617 mc->mc_vd, mc->mc_offset,
618 zio->io_abd, zio->io_size,
619 ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
620 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
621 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
627 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
629 if (faulted == vd->vdev_children)
630 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
631 VDEV_AUX_NO_REPLICAS);
632 else if (degraded + faulted != 0)
633 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
635 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
638 vdev_ops_t vdev_mirror_ops = {
642 vdev_mirror_io_start,
644 vdev_mirror_state_change,
647 VDEV_TYPE_MIRROR, /* name of this vdev type */
648 B_FALSE /* not a leaf vdev */
651 vdev_ops_t vdev_replacing_ops = {
655 vdev_mirror_io_start,
657 vdev_mirror_state_change,
660 VDEV_TYPE_REPLACING, /* name of this vdev type */
661 B_FALSE /* not a leaf vdev */
664 vdev_ops_t vdev_spare_ops = {
668 vdev_mirror_io_start,
670 vdev_mirror_state_change,
673 VDEV_TYPE_SPARE, /* name of this vdev type */
674 B_FALSE /* not a leaf vdev */