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) 2013 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 {
46 uint8_t mc_speculative;
49 typedef struct mirror_map {
54 mirror_child_t mm_child[1];
57 int vdev_mirror_shift = 21;
60 vdev_mirror_map_free(zio_t *zio)
62 mirror_map_t *mm = zio->io_vsd;
64 kmem_free(mm, offsetof(mirror_map_t, mm_child[mm->mm_children]));
67 static const zio_vsd_ops_t vdev_mirror_vsd_ops = {
69 zio_vsd_default_cksum_report
73 vdev_mirror_map_alloc(zio_t *zio)
75 mirror_map_t *mm = NULL;
77 vdev_t *vd = zio->io_vd;
81 dva_t *dva = zio->io_bp->blk_dva;
82 spa_t *spa = zio->io_spa;
84 c = BP_GET_NDVAS(zio->io_bp);
86 mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
88 mm->mm_replacing = B_FALSE;
89 mm->mm_preferred = spa_get_random(c);
93 * Check the other, lower-index DVAs to see if they're on
94 * the same vdev as the child we picked. If they are, use
95 * them since they are likely to have been allocated from
96 * the primary metaslab in use at the time, and hence are
97 * more likely to have locality with single-copy data.
99 for (c = mm->mm_preferred, d = c - 1; d >= 0; d--) {
100 if (DVA_GET_VDEV(&dva[d]) == DVA_GET_VDEV(&dva[c]))
101 mm->mm_preferred = d;
104 for (c = 0; c < mm->mm_children; c++) {
105 mc = &mm->mm_child[c];
107 mc->mc_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[c]));
108 mc->mc_offset = DVA_GET_OFFSET(&dva[c]);
111 c = vd->vdev_children;
113 mm = kmem_zalloc(offsetof(mirror_map_t, mm_child[c]), KM_SLEEP);
115 mm->mm_replacing = (vd->vdev_ops == &vdev_replacing_ops ||
116 vd->vdev_ops == &vdev_spare_ops);
117 mm->mm_preferred = mm->mm_replacing ? 0 :
118 (zio->io_offset >> vdev_mirror_shift) % c;
119 mm->mm_root = B_FALSE;
121 for (c = 0; c < mm->mm_children; c++) {
122 mc = &mm->mm_child[c];
123 mc->mc_vd = vd->vdev_child[c];
124 mc->mc_offset = zio->io_offset;
129 zio->io_vsd_ops = &vdev_mirror_vsd_ops;
134 vdev_mirror_open(vdev_t *vd, uint64_t *asize, uint64_t *max_asize,
135 uint64_t *logical_ashift, uint64_t *physical_ashift)
140 if (vd->vdev_children == 0) {
141 vd->vdev_stat.vs_aux = VDEV_AUX_BAD_LABEL;
142 return (SET_ERROR(EINVAL));
145 vdev_open_children(vd);
147 for (int c = 0; c < vd->vdev_children; c++) {
148 vdev_t *cvd = vd->vdev_child[c];
150 if (cvd->vdev_open_error) {
151 lasterror = cvd->vdev_open_error;
156 *asize = MIN(*asize - 1, cvd->vdev_asize - 1) + 1;
157 *max_asize = MIN(*max_asize - 1, cvd->vdev_max_asize - 1) + 1;
158 *logical_ashift = MAX(*logical_ashift, cvd->vdev_ashift);
159 *physical_ashift = MAX(*physical_ashift,
160 cvd->vdev_physical_ashift);
163 if (numerrors == vd->vdev_children) {
164 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
172 vdev_mirror_close(vdev_t *vd)
174 for (int c = 0; c < vd->vdev_children; c++)
175 vdev_close(vd->vdev_child[c]);
179 vdev_mirror_child_done(zio_t *zio)
181 mirror_child_t *mc = zio->io_private;
183 mc->mc_error = zio->io_error;
189 vdev_mirror_scrub_done(zio_t *zio)
191 mirror_child_t *mc = zio->io_private;
193 if (zio->io_error == 0) {
196 mutex_enter(&zio->io_lock);
197 while ((pio = zio_walk_parents(zio)) != NULL) {
198 mutex_enter(&pio->io_lock);
199 ASSERT3U(zio->io_size, >=, pio->io_size);
200 bcopy(zio->io_data, pio->io_data, pio->io_size);
201 mutex_exit(&pio->io_lock);
203 mutex_exit(&zio->io_lock);
206 zio_buf_free(zio->io_data, zio->io_size);
208 mc->mc_error = zio->io_error;
214 * Try to find a child whose DTL doesn't contain the block we want to read.
215 * If we can't, try the read on any vdev we haven't already tried.
218 vdev_mirror_child_select(zio_t *zio)
220 mirror_map_t *mm = zio->io_vsd;
222 uint64_t txg = zio->io_txg;
225 ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
228 * Try to find a child whose DTL doesn't contain the block to read.
229 * If a child is known to be completely inaccessible (indicated by
230 * vdev_readable() returning B_FALSE), don't even try.
232 for (i = 0, c = mm->mm_preferred; i < mm->mm_children; i++, c++) {
233 if (c >= mm->mm_children)
235 mc = &mm->mm_child[c];
236 if (mc->mc_tried || mc->mc_skipped)
238 if (!vdev_readable(mc->mc_vd)) {
239 mc->mc_error = SET_ERROR(ENXIO);
240 mc->mc_tried = 1; /* don't even try */
244 if (!vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1))
246 mc->mc_error = SET_ERROR(ESTALE);
248 mc->mc_speculative = 1;
252 * Every device is either missing or has this txg in its DTL.
253 * Look for any child we haven't already tried before giving up.
255 for (c = 0; c < mm->mm_children; c++)
256 if (!mm->mm_child[c].mc_tried)
260 * Every child failed. There's no place left to look.
266 vdev_mirror_io_start(zio_t *zio)
272 mm = vdev_mirror_map_alloc(zio);
274 if (zio->io_type == ZIO_TYPE_READ) {
275 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
277 * For scrubbing reads we need to allocate a read
278 * buffer for each child and issue reads to all
279 * children. If any child succeeds, it will copy its
280 * data into zio->io_data in vdev_mirror_scrub_done.
282 for (c = 0; c < mm->mm_children; c++) {
283 mc = &mm->mm_child[c];
284 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
285 mc->mc_vd, mc->mc_offset,
286 zio_buf_alloc(zio->io_size), zio->io_size,
287 zio->io_type, zio->io_priority, 0,
288 vdev_mirror_scrub_done, mc));
290 return (ZIO_PIPELINE_CONTINUE);
293 * For normal reads just pick one child.
295 c = vdev_mirror_child_select(zio);
298 ASSERT(zio->io_type == ZIO_TYPE_WRITE ||
299 zio->io_type == ZIO_TYPE_FREE);
302 * Writes and frees go to all children.
305 children = mm->mm_children;
309 mc = &mm->mm_child[c];
310 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
311 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
312 zio->io_type, zio->io_priority, 0,
313 vdev_mirror_child_done, mc));
317 return (ZIO_PIPELINE_CONTINUE);
321 vdev_mirror_worst_error(mirror_map_t *mm)
323 int error[2] = { 0, 0 };
325 for (int c = 0; c < mm->mm_children; c++) {
326 mirror_child_t *mc = &mm->mm_child[c];
327 int s = mc->mc_speculative;
328 error[s] = zio_worst_error(error[s], mc->mc_error);
331 return (error[0] ? error[0] : error[1]);
335 vdev_mirror_io_done(zio_t *zio)
337 mirror_map_t *mm = zio->io_vsd;
341 int unexpected_errors = 0;
343 for (c = 0; c < mm->mm_children; c++) {
344 mc = &mm->mm_child[c];
349 } else if (mc->mc_tried) {
354 if (zio->io_type == ZIO_TYPE_WRITE) {
356 * XXX -- for now, treat partial writes as success.
358 * Now that we support write reallocation, it would be better
359 * to treat partial failure as real failure unless there are
360 * no non-degraded top-level vdevs left, and not update DTLs
361 * if we intend to reallocate.
364 if (good_copies != mm->mm_children) {
366 * Always require at least one good copy.
368 * For ditto blocks (io_vd == NULL), require
369 * all copies to be good.
371 * XXX -- for replacing vdevs, there's no great answer.
372 * If the old device is really dead, we may not even
373 * be able to access it -- so we only want to
374 * require good writes to the new device. But if
375 * the new device turns out to be flaky, we want
376 * to be able to detach it -- which requires all
377 * writes to the old device to have succeeded.
379 if (good_copies == 0 || zio->io_vd == NULL)
380 zio->io_error = vdev_mirror_worst_error(mm);
383 } else if (zio->io_type == ZIO_TYPE_FREE) {
387 ASSERT(zio->io_type == ZIO_TYPE_READ);
390 * If we don't have a good copy yet, keep trying other children.
393 if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
394 ASSERT(c >= 0 && c < mm->mm_children);
395 mc = &mm->mm_child[c];
396 zio_vdev_io_redone(zio);
397 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
398 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
399 ZIO_TYPE_READ, zio->io_priority, 0,
400 vdev_mirror_child_done, mc));
405 if (good_copies == 0) {
406 zio->io_error = vdev_mirror_worst_error(mm);
407 ASSERT(zio->io_error != 0);
410 if (good_copies && spa_writeable(zio->io_spa) &&
411 (unexpected_errors ||
412 (zio->io_flags & ZIO_FLAG_RESILVER) ||
413 ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) {
415 * Use the good data we have in hand to repair damaged children.
417 for (c = 0; c < mm->mm_children; c++) {
419 * Don't rewrite known good children.
420 * Not only is it unnecessary, it could
421 * actually be harmful: if the system lost
422 * power while rewriting the only good copy,
423 * there would be no good copies left!
425 mc = &mm->mm_child[c];
427 if (mc->mc_error == 0) {
430 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
431 !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
434 mc->mc_error = SET_ERROR(ESTALE);
437 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
438 mc->mc_vd, mc->mc_offset,
439 zio->io_data, zio->io_size,
440 ZIO_TYPE_WRITE, zio->io_priority,
441 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
442 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
448 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
450 if (faulted == vd->vdev_children)
451 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
452 VDEV_AUX_NO_REPLICAS);
453 else if (degraded + faulted != 0)
454 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
456 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
459 vdev_ops_t vdev_mirror_ops = {
463 vdev_mirror_io_start,
465 vdev_mirror_state_change,
468 VDEV_TYPE_MIRROR, /* name of this vdev type */
469 B_FALSE /* not a leaf vdev */
472 vdev_ops_t vdev_replacing_ops = {
476 vdev_mirror_io_start,
478 vdev_mirror_state_change,
481 VDEV_TYPE_REPLACING, /* name of this vdev type */
482 B_FALSE /* not a leaf vdev */
485 vdev_ops_t vdev_spare_ops = {
489 vdev_mirror_io_start,
491 vdev_mirror_state_change,
494 VDEV_TYPE_SPARE, /* name of this vdev type */
495 B_FALSE /* not a leaf vdev */