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, 2014 by Delphix. All rights reserved.
30 #include <sys/zfs_context.h>
32 #include <sys/vdev_impl.h>
34 #include <sys/fs/zfs.h>
37 * Virtual device vector for mirroring.
40 typedef struct mirror_child {
47 uint8_t mc_speculative;
50 typedef struct mirror_map {
54 boolean_t mm_replacing;
56 mirror_child_t mm_child[];
59 static int vdev_mirror_shift = 21;
61 SYSCTL_DECL(_vfs_zfs_vdev);
62 static SYSCTL_NODE(_vfs_zfs_vdev, OID_AUTO, mirror, CTLFLAG_RD, 0,
66 * The load configuration settings below are tuned by default for
67 * the case where all devices are of the same rotational type.
69 * If there is a mixture of rotating and non-rotating media, setting
70 * non_rotating_seek_inc to 0 may well provide better results as it
71 * will direct more reads to the non-rotating vdevs which are more
72 * likely to have a higher performance.
75 /* Rotating media load calculation configuration. */
76 static int rotating_inc = 0;
77 TUNABLE_INT("vfs.zfs.vdev.mirror.rotating_inc", &rotating_inc);
78 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, rotating_inc, CTLFLAG_RW,
79 &rotating_inc, 0, "Rotating media load increment for non-seeking I/O's");
81 static int rotating_seek_inc = 5;
82 TUNABLE_INT("vfs.zfs.vdev.mirror.rotating_seek_inc", &rotating_seek_inc);
83 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, rotating_seek_inc, CTLFLAG_RW,
84 &rotating_seek_inc, 0, "Rotating media load increment for seeking I/O's");
86 static int rotating_seek_offset = 1 * 1024 * 1024;
87 TUNABLE_INT("vfs.zfs.vdev.mirror.rotating_seek_offset", &rotating_seek_offset);
88 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, rotating_seek_offset, CTLFLAG_RW,
89 &rotating_seek_offset, 0, "Offset in bytes from the last I/O which "
90 "triggers a reduced rotating media seek increment");
92 /* Non-rotating media load calculation configuration. */
93 static int non_rotating_inc = 0;
94 TUNABLE_INT("vfs.zfs.vdev.mirror.non_rotating_inc", &non_rotating_inc);
95 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, non_rotating_inc, CTLFLAG_RW,
97 "Non-rotating media load increment for non-seeking I/O's");
99 static int non_rotating_seek_inc = 1;
100 TUNABLE_INT("vfs.zfs.vdev.mirror.non_rotating_seek_inc",
101 &non_rotating_seek_inc);
102 SYSCTL_INT(_vfs_zfs_vdev_mirror, OID_AUTO, non_rotating_seek_inc, CTLFLAG_RW,
103 &non_rotating_seek_inc, 0,
104 "Non-rotating media load increment for seeking I/O's");
108 vdev_mirror_map_size(int children)
110 return (offsetof(mirror_map_t, mm_child[children]) +
111 sizeof(int) * children);
114 static inline mirror_map_t *
115 vdev_mirror_map_alloc(int children, boolean_t replacing, boolean_t root)
119 mm = kmem_zalloc(vdev_mirror_map_size(children), KM_SLEEP);
120 mm->mm_children = children;
121 mm->mm_replacing = replacing;
123 mm->mm_preferred = (int *)((uintptr_t)mm +
124 offsetof(mirror_map_t, mm_child[children]));
130 vdev_mirror_map_free(zio_t *zio)
132 mirror_map_t *mm = zio->io_vsd;
134 kmem_free(mm, vdev_mirror_map_size(mm->mm_children));
137 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
138 vdev_mirror_map_free,
139 zio_vsd_default_cksum_report
143 vdev_mirror_load(mirror_map_t *mm, vdev_t *vd, uint64_t zio_offset)
148 /* All DVAs have equal weight at the root. */
153 * We don't return INT_MAX if the device is resilvering i.e.
154 * vdev_resilver_txg != 0 as when tested performance was slightly
155 * worse overall when resilvering with compared to without.
158 /* Standard load based on pending queue length. */
159 load = vdev_queue_length(vd);
160 lastoffset = vdev_queue_lastoffset(vd);
162 if (vd->vdev_rotation_rate == VDEV_RATE_NON_ROTATING) {
163 /* Non-rotating media. */
164 if (lastoffset == zio_offset)
165 return (load + non_rotating_inc);
168 * Apply a seek penalty even for non-rotating devices as
169 * sequential I/O'a can be aggregated into fewer operations
170 * on the device, thus avoiding unnecessary per-command
171 * overhead and boosting performance.
173 return (load + non_rotating_seek_inc);
176 /* Rotating media I/O's which directly follow the last I/O. */
177 if (lastoffset == zio_offset)
178 return (load + rotating_inc);
181 * Apply half the seek increment to I/O's within seek offset
182 * of the last I/O queued to this vdev as they should incure less
183 * of a seek increment.
185 if (ABS(lastoffset - zio_offset) < rotating_seek_offset)
186 return (load + (rotating_seek_inc / 2));
188 /* Apply the full seek increment to all other I/O's. */
189 return (load + rotating_seek_inc);
193 static mirror_map_t *
194 vdev_mirror_map_init(zio_t *zio)
196 mirror_map_t *mm = NULL;
198 vdev_t *vd = zio->io_vd;
202 dva_t *dva = zio->io_bp->blk_dva;
203 spa_t *spa = zio->io_spa;
205 mm = vdev_mirror_map_alloc(BP_GET_NDVAS(zio->io_bp), B_FALSE,
207 for (c = 0; c < mm->mm_children; c++) {
208 mc = &mm->mm_child[c];
209 mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
210 mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
213 mm = vdev_mirror_map_alloc(vd->vdev_children,
214 (vd->vdev_ops == &vdev_replacing_ops ||
215 vd->vdev_ops == &vdev_spare_ops), B_FALSE);
216 for (c = 0; c < mm->mm_children; c++) {
217 mc = &mm->mm_child[c];
218 mc->mc_vd = vd->vdev_child[c];
219 mc->mc_offset = zio->io_offset;
224 zio->io_vsd_ops = &vdev_mirror_vsd_ops;
229 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
230 uint64_t *logical_ashift, uint64_t *physical_ashift)
235 if (vd->vdev_children == 0) {
236 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
237 return (SET_ERROR(EINVAL));
240 vdev_open_children(vd);
242 for (int c = 0; c < vd->vdev_children; c++) {
243 vdev_t *cvd = vd->vdev_child[c];
245 if (cvd->vdev_open_error) {
246 lasterror = cvd->vdev_open_error;
251 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
252 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
253 *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift);
254 *physical_ashift = MAX(*physical_ashift,
255 cvd->vdev_physical_ashift);
258 if (numerrors == vd->vdev_children) {
259 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
267 vdev_mirror_close(vdev_t *vd)
269 for (int c = 0; c < vd->vdev_children; c++)
270 vdev_close(vd->vdev_child[c]);
274 vdev_mirror_child_done(zio_t *zio)
276 mirror_child_t *mc = zio->io_private;
278 mc->mc_error = zio->io_error;
284 vdev_mirror_scrub_done(zio_t *zio)
286 mirror_child_t *mc = zio->io_private;
288 if (zio->io_error == 0) {
291 mutex_enter(&zio->io_lock);
292 while ((pio = zio_walk_parents(zio)) != NULL) {
293 mutex_enter(&pio->io_lock);
294 ASSERT3U(zio->io_size, >=, pio->io_size);
295 bcopy(zio->io_data, pio->io_data, pio->io_size);
296 mutex_exit(&pio->io_lock);
298 mutex_exit(&zio->io_lock);
301 zio_buf_free(zio->io_data, zio->io_size);
303 mc->mc_error = zio->io_error;
309 * Check the other, lower-index DVAs to see if they're on the same
310 * vdev as the child we picked. If they are, use them since they
311 * are likely to have been allocated from the primary metaslab in
312 * use at the time, and hence are more likely to have locality with
316 vdev_mirror_dva_select(zio_t *zio, int p)
318 dva_t *dva = zio->io_bp->blk_dva;
319 mirror_map_t *mm = zio->io_vsd;
323 preferred = mm->mm_preferred[p];
324 for (p-- ; p >= 0; p--) {
325 c = mm->mm_preferred[p];
326 if (DVA_GET_VDEV(&dva[c]) == DVA_GET_VDEV(&dva[preferred]))
333 vdev_mirror_preferred_child_randomize(zio_t *zio)
335 mirror_map_t *mm = zio->io_vsd;
339 p = spa_get_random(mm->mm_preferred_cnt);
340 return (vdev_mirror_dva_select(zio, p));
344 * To ensure we don't always favour the first matching vdev,
345 * which could lead to wear leveling issues on SSD's, we
346 * use the I/O offset as a pseudo random seed into the vdevs
347 * which have the lowest load.
349 p = (zio->io_offset >> vdev_mirror_shift) % mm->mm_preferred_cnt;
350 return (mm->mm_preferred[p]);
354 * Try to find a vdev whose DTL doesn't contain the block we want to read
355 * prefering vdevs based on determined load.
357 * If we can't, try the read on any vdev we haven't already tried.
360 vdev_mirror_child_select(zio_t *zio)
362 mirror_map_t *mm = zio->io_vsd;
363 uint64_t txg = zio->io_txg;
366 ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
368 lowest_load = INT_MAX;
369 mm->mm_preferred_cnt = 0;
370 for (c = 0; c < mm->mm_children; c++) {
373 mc = &mm->mm_child[c];
374 if (mc->mc_tried || mc->mc_skipped)
377 if (!vdev_readable(mc->mc_vd)) {
378 mc->mc_error = SET_ERROR(ENXIO);
379 mc->mc_tried = 1; /* don't even try */
384 if (vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1)) {
385 mc->mc_error = SET_ERROR(ESTALE);
387 mc->mc_speculative = 1;
391 mc->mc_load = vdev_mirror_load(mm, mc->mc_vd, mc->mc_offset);
392 if (mc->mc_load > lowest_load)
395 if (mc->mc_load < lowest_load) {
396 lowest_load = mc->mc_load;
397 mm->mm_preferred_cnt = 0;
399 mm->mm_preferred[mm->mm_preferred_cnt] = c;
400 mm->mm_preferred_cnt++;
403 if (mm->mm_preferred_cnt == 1) {
404 vdev_queue_register_lastoffset(
405 mm->mm_child[mm->mm_preferred[0]].mc_vd, zio);
406 return (mm->mm_preferred[0]);
409 if (mm->mm_preferred_cnt > 1) {
410 int c = vdev_mirror_preferred_child_randomize(zio);
412 vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd, zio);
417 * Every device is either missing or has this txg in its DTL.
418 * Look for any child we haven't already tried before giving up.
420 for (c = 0; c < mm->mm_children; c++) {
421 if (!mm->mm_child[c].mc_tried) {
422 vdev_queue_register_lastoffset(mm->mm_child[c].mc_vd,
429 * Every child failed. There's no place left to look.
435 vdev_mirror_io_start(zio_t *zio)
441 mm = vdev_mirror_map_init(zio);
443 if (zio->io_type == ZIO_TYPE_READ) {
444 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing &&
445 mm->mm_children > 1) {
447 * For scrubbing reads we need to allocate a read
448 * buffer for each child and issue reads to all
449 * children. If any child succeeds, it will copy its
450 * data into zio->io_data in vdev_mirror_scrub_done.
452 for (c = 0; c < mm->mm_children; c++) {
453 mc = &mm->mm_child[c];
454 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
455 mc->mc_vd, mc->mc_offset,
456 zio_buf_alloc(zio->io_size), zio->io_size,
457 zio->io_type, zio->io_priority, 0,
458 vdev_mirror_scrub_done, mc));
464 * For normal reads just pick one child.
466 c = vdev_mirror_child_select(zio);
469 ASSERT(zio->io_type == ZIO_TYPE_WRITE ||
470 zio->io_type == ZIO_TYPE_FREE);
473 * Writes and frees go to all children.
476 children = mm->mm_children;
480 mc = &mm->mm_child[c];
481 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
482 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
483 zio->io_type, zio->io_priority, 0,
484 vdev_mirror_child_done, mc));
492 vdev_mirror_worst_error(mirror_map_t *mm)
494 int error[2] = { 0, 0 };
496 for (int c = 0; c < mm->mm_children; c++) {
497 mirror_child_t *mc = &mm->mm_child[c];
498 int s = mc->mc_speculative;
499 error[s] = zio_worst_error(error[s], mc->mc_error);
502 return (error[0] ? error[0] : error[1]);
506 vdev_mirror_io_done(zio_t *zio)
508 mirror_map_t *mm = zio->io_vsd;
512 int unexpected_errors = 0;
514 for (c = 0; c < mm->mm_children; c++) {
515 mc = &mm->mm_child[c];
520 } else if (mc->mc_tried) {
525 if (zio->io_type == ZIO_TYPE_WRITE) {
527 * XXX -- for now, treat partial writes as success.
529 * Now that we support write reallocation, it would be better
530 * to treat partial failure as real failure unless there are
531 * no non-degraded top-level vdevs left, and not update DTLs
532 * if we intend to reallocate.
535 if (good_copies != mm->mm_children) {
537 * Always require at least one good copy.
539 * For ditto blocks (io_vd == NULL), require
540 * all copies to be good.
542 * XXX -- for replacing vdevs, there's no great answer.
543 * If the old device is really dead, we may not even
544 * be able to access it -- so we only want to
545 * require good writes to the new device. But if
546 * the new device turns out to be flaky, we want
547 * to be able to detach it -- which requires all
548 * writes to the old device to have succeeded.
550 if (good_copies == 0 || zio->io_vd == NULL)
551 zio->io_error = vdev_mirror_worst_error(mm);
554 } else if (zio->io_type == ZIO_TYPE_FREE) {
558 ASSERT(zio->io_type == ZIO_TYPE_READ);
561 * If we don't have a good copy yet, keep trying other children.
564 if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
565 ASSERT(c >= 0 && c < mm->mm_children);
566 mc = &mm->mm_child[c];
567 zio_vdev_io_redone(zio);
568 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
569 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
570 ZIO_TYPE_READ, zio->io_priority, 0,
571 vdev_mirror_child_done, mc));
576 if (good_copies == 0) {
577 zio->io_error = vdev_mirror_worst_error(mm);
578 ASSERT(zio->io_error != 0);
581 if (good_copies && spa_writeable(zio->io_spa) &&
582 (unexpected_errors ||
583 (zio->io_flags & ZIO_FLAG_RESILVER) ||
584 ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) {
586 * Use the good data we have in hand to repair damaged children.
588 for (c = 0; c < mm->mm_children; c++) {
590 * Don't rewrite known good children.
591 * Not only is it unnecessary, it could
592 * actually be harmful: if the system lost
593 * power while rewriting the only good copy,
594 * there would be no good copies left!
596 mc = &mm->mm_child[c];
598 if (mc->mc_error == 0) {
601 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
602 !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
605 mc->mc_error = SET_ERROR(ESTALE);
608 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
609 mc->mc_vd, mc->mc_offset,
610 zio->io_data, zio->io_size,
611 ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
612 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
613 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
619 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
621 if (faulted == vd->vdev_children)
622 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
623 VDEV_AUX_NO_REPLICAS);
624 else if (degraded + faulted != 0)
625 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
627 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
630 vdev_ops_t vdev_mirror_ops = {
634 vdev_mirror_io_start,
636 vdev_mirror_state_change,
639 VDEV_TYPE_MIRROR, /* name of this vdev type */
640 B_FALSE /* not a leaf vdev */
643 vdev_ops_t vdev_replacing_ops = {
647 vdev_mirror_io_start,
649 vdev_mirror_state_change,
652 VDEV_TYPE_REPLACING, /* name of this vdev type */
653 B_FALSE /* not a leaf vdev */
656 vdev_ops_t vdev_spare_ops = {
660 vdev_mirror_io_start,
662 vdev_mirror_state_change,
665 VDEV_TYPE_SPARE, /* name of this vdev type */
666 B_FALSE /* not a leaf vdev */