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 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 {
55 mirror_child_t mm_child[1];
59 * When the children are equally busy queue incoming requests to a single
60 * child for N microseconds. This is done to maximize the likelihood that
61 * the Linux elevator will be able to merge requests while it is plugged.
62 * Otherwise, requests are queued to the least busy device.
64 * For rotational disks the Linux elevator will plug for 10ms which is
65 * why zfs_vdev_mirror_switch_us is set to 10ms by default. For non-
66 * rotational disks the elevator will not plug, but 10ms is still a small
67 * enough value that the requests will get spread over all the children.
69 * For fast SSDs it may make sense to decrease zfs_vdev_mirror_switch_us
70 * significantly to bound the worst case latencies. It would probably be
71 * ideal to calculate a decaying average of the last observed latencies and
72 * use that to dynamically adjust the zfs_vdev_mirror_switch_us time.
74 int zfs_vdev_mirror_switch_us = 10000;
77 vdev_mirror_map_free(zio_t *zio)
79 mirror_map_t *mm = zio->io_vsd;
81 kmem_free(mm, offsetof(mirror_map_t, mm_child[mm->mm_children]));
84 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
86 zio_vsd_default_cksum_report
90 vdev_mirror_pending(vdev_t *vd)
92 vdev_queue_t *vq = &vd->vdev_queue;
95 mutex_enter(&vq->vq_lock);
96 pending = avl_numnodes(&vq->vq_pending_tree);
97 mutex_exit(&vq->vq_lock);
102 static mirror_map_t *
103 vdev_mirror_map_alloc(zio_t *zio)
105 mirror_map_t *mm = NULL;
107 vdev_t *vd = zio->io_vd;
111 dva_t *dva = zio->io_bp->blk_dva;
112 spa_t *spa = zio->io_spa;
114 c = BP_GET_NDVAS(zio->io_bp);
116 mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_PUSHPAGE);
118 mm->mm_replacing = B_FALSE;
119 mm->mm_preferred = spa_get_random(c);
120 mm->mm_root = B_TRUE;
123 * Check the other, lower-index DVAs to see if they're on
124 * the same vdev as the child we picked. If they are, use
125 * them since they are likely to have been allocated from
126 * the primary metaslab in use at the time, and hence are
127 * more likely to have locality with single-copy data.
129 for (c = mm->mm_preferred, d = c - 1; d >= 0; d--) {
130 if (DVA_GET_VDEV(&dva[d]) == DVA_GET_VDEV(&dva[c]))
131 mm->mm_preferred = d;
134 for (c = 0; c < mm->mm_children; c++) {
135 mc = &mm->mm_child[c];
137 mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
138 mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
141 int lowest_pending = INT_MAX;
144 c = vd->vdev_children;
146 mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_PUSHPAGE);
148 mm->mm_replacing = (vd->vdev_ops == &vdev_replacing_ops ||
149 vd->vdev_ops == &vdev_spare_ops);
150 mm->mm_preferred = 0;
151 mm->mm_root = B_FALSE;
153 for (c = 0; c < mm->mm_children; c++) {
154 mc = &mm->mm_child[c];
155 mc->mc_vd = vd->vdev_child[c];
156 mc->mc_offset = zio->io_offset;
158 if (mm->mm_replacing)
161 if (!vdev_readable(mc->mc_vd)) {
162 mc->mc_error = ENXIO;
165 mc->mc_pending = INT_MAX;
169 mc->mc_pending = vdev_mirror_pending(mc->mc_vd);
170 if (mc->mc_pending < lowest_pending) {
171 lowest_pending = mc->mc_pending;
173 } else if (mc->mc_pending == lowest_pending) {
178 d = gethrtime() / (NSEC_PER_USEC * zfs_vdev_mirror_switch_us);
179 d = (d % lowest_nr) + 1;
181 for (c = 0; c < mm->mm_children; c++) {
182 mc = &mm->mm_child[c];
184 if (mm->mm_child[c].mc_pending == lowest_pending) {
186 mm->mm_preferred = c;
194 zio->io_vsd_ops = &vdev_mirror_vsd_ops;
199 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
206 if (vd->vdev_children == 0) {
207 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
211 vdev_open_children(vd);
213 for (c = 0; c < vd->vdev_children; c++) {
214 vdev_t *cvd = vd->vdev_child[c];
216 if (cvd->vdev_open_error) {
217 lasterror = cvd->vdev_open_error;
222 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
223 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
224 *ashift = MAX(*ashift, cvd->vdev_ashift);
227 if (numerrors == vd->vdev_children) {
228 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
236 vdev_mirror_close(vdev_t *vd)
240 for (c = 0; c < vd->vdev_children; c++)
241 vdev_close(vd->vdev_child[c]);
245 vdev_mirror_child_done(zio_t *zio)
247 mirror_child_t *mc = zio->io_private;
249 mc->mc_error = zio->io_error;
255 vdev_mirror_scrub_done(zio_t *zio)
257 mirror_child_t *mc = zio->io_private;
259 if (zio->io_error == 0) {
262 mutex_enter(&zio->io_lock);
263 while ((pio = zio_walk_parents(zio)) != NULL) {
264 mutex_enter(&pio->io_lock);
265 ASSERT3U(zio->io_size, >=, pio->io_size);
266 bcopy(zio->io_data, pio->io_data, pio->io_size);
267 mutex_exit(&pio->io_lock);
269 mutex_exit(&zio->io_lock);
272 zio_buf_free(zio->io_data, zio->io_size);
274 mc->mc_error = zio->io_error;
280 * Try to find a child whose DTL doesn't contain the block we want to read.
281 * If we can't, try the read on any vdev we haven't already tried.
284 vdev_mirror_child_select(zio_t *zio)
286 mirror_map_t *mm = zio->io_vsd;
288 uint64_t txg = zio->io_txg;
291 ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
294 * Try to find a child whose DTL doesn't contain the block to read.
295 * If a child is known to be completely inaccessible (indicated by
296 * vdev_readable() returning B_FALSE), don't even try.
298 for (i = 0, c = mm->mm_preferred; i < mm->mm_children; i++, c++) {
299 if (c >= mm->mm_children)
301 mc = &mm->mm_child[c];
302 if (mc->mc_tried || mc->mc_skipped)
304 if (!vdev_readable(mc->mc_vd)) {
305 mc->mc_error = ENXIO;
306 mc->mc_tried = 1; /* don't even try */
310 if (!vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1))
312 mc->mc_error = ESTALE;
314 mc->mc_speculative = 1;
318 * Every device is either missing or has this txg in its DTL.
319 * Look for any child we haven't already tried before giving up.
321 for (c = 0; c < mm->mm_children; c++)
322 if (!mm->mm_child[c].mc_tried)
326 * Every child failed. There's no place left to look.
332 vdev_mirror_io_start(zio_t *zio)
338 mm = vdev_mirror_map_alloc(zio);
340 if (zio->io_type == ZIO_TYPE_READ) {
341 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
343 * For scrubbing reads we need to allocate a read
344 * buffer for each child and issue reads to all
345 * children. If any child succeeds, it will copy its
346 * data into zio->io_data in vdev_mirror_scrub_done.
348 for (c = 0; c < mm->mm_children; c++) {
349 mc = &mm->mm_child[c];
350 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
351 mc->mc_vd, mc->mc_offset,
352 zio_buf_alloc(zio->io_size), zio->io_size,
353 zio->io_type, zio->io_priority, 0,
354 vdev_mirror_scrub_done, mc));
356 return (ZIO_PIPELINE_CONTINUE);
359 * For normal reads just pick one child.
361 c = vdev_mirror_child_select(zio);
364 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
367 * Writes go to all children.
370 children = mm->mm_children;
374 mc = &mm->mm_child[c];
375 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
376 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
377 zio->io_type, zio->io_priority, 0,
378 vdev_mirror_child_done, mc));
382 return (ZIO_PIPELINE_CONTINUE);
386 vdev_mirror_worst_error(mirror_map_t *mm)
388 int c, error[2] = { 0, 0 };
390 for (c = 0; c < mm->mm_children; c++) {
391 mirror_child_t *mc = &mm->mm_child[c];
392 int s = mc->mc_speculative;
393 error[s] = zio_worst_error(error[s], mc->mc_error);
396 return (error[0] ? error[0] : error[1]);
400 vdev_mirror_io_done(zio_t *zio)
402 mirror_map_t *mm = zio->io_vsd;
406 int unexpected_errors = 0;
408 for (c = 0; c < mm->mm_children; c++) {
409 mc = &mm->mm_child[c];
414 } else if (mc->mc_tried) {
419 if (zio->io_type == ZIO_TYPE_WRITE) {
421 * XXX -- for now, treat partial writes as success.
423 * Now that we support write reallocation, it would be better
424 * to treat partial failure as real failure unless there are
425 * no non-degraded top-level vdevs left, and not update DTLs
426 * if we intend to reallocate.
429 if (good_copies != mm->mm_children) {
431 * Always require at least one good copy.
433 * For ditto blocks (io_vd == NULL), require
434 * all copies to be good.
436 * XXX -- for replacing vdevs, there's no great answer.
437 * If the old device is really dead, we may not even
438 * be able to access it -- so we only want to
439 * require good writes to the new device. But if
440 * the new device turns out to be flaky, we want
441 * to be able to detach it -- which requires all
442 * writes to the old device to have succeeded.
444 if (good_copies == 0 || zio->io_vd == NULL)
445 zio->io_error = vdev_mirror_worst_error(mm);
450 ASSERT(zio->io_type == ZIO_TYPE_READ);
453 * If we don't have a good copy yet, keep trying other children.
456 if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
457 ASSERT(c >= 0 && c < mm->mm_children);
458 mc = &mm->mm_child[c];
459 zio_vdev_io_redone(zio);
460 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
461 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
462 ZIO_TYPE_READ, zio->io_priority, 0,
463 vdev_mirror_child_done, mc));
468 if (good_copies == 0) {
469 zio->io_error = vdev_mirror_worst_error(mm);
470 ASSERT(zio->io_error != 0);
473 if (good_copies && spa_writeable(zio->io_spa) &&
474 (unexpected_errors ||
475 (zio->io_flags & ZIO_FLAG_RESILVER) ||
476 ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) {
478 * Use the good data we have in hand to repair damaged children.
480 for (c = 0; c < mm->mm_children; c++) {
482 * Don't rewrite known good children.
483 * Not only is it unnecessary, it could
484 * actually be harmful: if the system lost
485 * power while rewriting the only good copy,
486 * there would be no good copies left!
488 mc = &mm->mm_child[c];
490 if (mc->mc_error == 0) {
493 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
494 !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
497 mc->mc_error = ESTALE;
500 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
501 mc->mc_vd, mc->mc_offset,
502 zio->io_data, zio->io_size,
503 ZIO_TYPE_WRITE, zio->io_priority,
504 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
505 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
511 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
513 if (faulted == vd->vdev_children)
514 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
515 VDEV_AUX_NO_REPLICAS);
516 else if (degraded + faulted != 0)
517 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
519 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
522 vdev_ops_t vdev_mirror_ops = {
526 vdev_mirror_io_start,
528 vdev_mirror_state_change,
531 VDEV_TYPE_MIRROR, /* name of this vdev type */
532 B_FALSE /* not a leaf vdev */
535 vdev_ops_t vdev_replacing_ops = {
539 vdev_mirror_io_start,
541 vdev_mirror_state_change,
544 VDEV_TYPE_REPLACING, /* name of this vdev type */
545 B_FALSE /* not a leaf vdev */
548 vdev_ops_t vdev_spare_ops = {
552 vdev_mirror_io_start,
554 vdev_mirror_state_change,
557 VDEV_TYPE_SPARE, /* name of this vdev type */
558 B_FALSE /* not a leaf vdev */
561 #if defined(_KERNEL) && defined(HAVE_SPL)
562 module_param(zfs_vdev_mirror_switch_us, int, 0644);
563 MODULE_PARM_DESC(zfs_vdev_mirror_switch_us, "Switch mirrors every N usecs");