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,
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 *ashift = MAX(*ashift, cvd->vdev_ashift);
161 if (numerrors == vd->vdev_children) {
162 vd->vdev_stat.vs_aux = VDEV_AUX_NO_REPLICAS;
170 vdev_mirror_close(vdev_t *vd)
172 for (int c = 0; c < vd->vdev_children; c++)
173 vdev_close(vd->vdev_child[c]);
177 vdev_mirror_child_done(zio_t *zio)
179 mirror_child_t *mc = zio->io_private;
181 mc->mc_error = zio->io_error;
187 vdev_mirror_scrub_done(zio_t *zio)
189 mirror_child_t *mc = zio->io_private;
191 if (zio->io_error == 0) {
194 mutex_enter(&zio->io_lock);
195 while ((pio = zio_walk_parents(zio)) != NULL) {
196 mutex_enter(&pio->io_lock);
197 ASSERT3U(zio->io_size, >=, pio->io_size);
198 bcopy(zio->io_data, pio->io_data, pio->io_size);
199 mutex_exit(&pio->io_lock);
201 mutex_exit(&zio->io_lock);
204 zio_buf_free(zio->io_data, zio->io_size);
206 mc->mc_error = zio->io_error;
212 * Try to find a child whose DTL doesn't contain the block we want to read.
213 * If we can't, try the read on any vdev we haven't already tried.
216 vdev_mirror_child_select(zio_t *zio)
218 mirror_map_t *mm = zio->io_vsd;
220 uint64_t txg = zio->io_txg;
223 ASSERT(zio->io_bp == NULL || BP_PHYSICAL_BIRTH(zio->io_bp) == txg);
226 * Try to find a child whose DTL doesn't contain the block to read.
227 * If a child is known to be completely inaccessible (indicated by
228 * vdev_readable() returning B_FALSE), don't even try.
230 for (i = 0, c = mm->mm_preferred; i < mm->mm_children; i++, c++) {
231 if (c >= mm->mm_children)
233 mc = &mm->mm_child[c];
234 if (mc->mc_tried || mc->mc_skipped)
236 if (!vdev_readable(mc->mc_vd)) {
237 mc->mc_error = SET_ERROR(ENXIO);
238 mc->mc_tried = 1; /* don't even try */
242 if (!vdev_dtl_contains(mc->mc_vd, DTL_MISSING, txg, 1))
244 mc->mc_error = SET_ERROR(ESTALE);
246 mc->mc_speculative = 1;
250 * Every device is either missing or has this txg in its DTL.
251 * Look for any child we haven't already tried before giving up.
253 for (c = 0; c < mm->mm_children; c++)
254 if (!mm->mm_child[c].mc_tried)
258 * Every child failed. There's no place left to look.
264 vdev_mirror_io_start(zio_t *zio)
270 mm = vdev_mirror_map_alloc(zio);
272 if (zio->io_type == ZIO_TYPE_READ) {
273 if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
275 * For scrubbing reads we need to allocate a read
276 * buffer for each child and issue reads to all
277 * children. If any child succeeds, it will copy its
278 * data into zio->io_data in vdev_mirror_scrub_done.
280 for (c = 0; c < mm->mm_children; c++) {
281 mc = &mm->mm_child[c];
282 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
283 mc->mc_vd, mc->mc_offset,
284 zio_buf_alloc(zio->io_size), zio->io_size,
285 zio->io_type, zio->io_priority, 0,
286 vdev_mirror_scrub_done, mc));
288 return (ZIO_PIPELINE_CONTINUE);
291 * For normal reads just pick one child.
293 c = vdev_mirror_child_select(zio);
296 ASSERT(zio->io_type == ZIO_TYPE_WRITE ||
297 zio->io_type == ZIO_TYPE_FREE);
300 * Writes and frees go to all children.
303 children = mm->mm_children;
307 mc = &mm->mm_child[c];
308 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
309 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
310 zio->io_type, zio->io_priority, 0,
311 vdev_mirror_child_done, mc));
315 return (ZIO_PIPELINE_CONTINUE);
319 vdev_mirror_worst_error(mirror_map_t *mm)
321 int error[2] = { 0, 0 };
323 for (int c = 0; c < mm->mm_children; c++) {
324 mirror_child_t *mc = &mm->mm_child[c];
325 int s = mc->mc_speculative;
326 error[s] = zio_worst_error(error[s], mc->mc_error);
329 return (error[0] ? error[0] : error[1]);
333 vdev_mirror_io_done(zio_t *zio)
335 mirror_map_t *mm = zio->io_vsd;
339 int unexpected_errors = 0;
341 for (c = 0; c < mm->mm_children; c++) {
342 mc = &mm->mm_child[c];
347 } else if (mc->mc_tried) {
352 if (zio->io_type == ZIO_TYPE_WRITE) {
354 * XXX -- for now, treat partial writes as success.
356 * Now that we support write reallocation, it would be better
357 * to treat partial failure as real failure unless there are
358 * no non-degraded top-level vdevs left, and not update DTLs
359 * if we intend to reallocate.
362 if (good_copies != mm->mm_children) {
364 * Always require at least one good copy.
366 * For ditto blocks (io_vd == NULL), require
367 * all copies to be good.
369 * XXX -- for replacing vdevs, there's no great answer.
370 * If the old device is really dead, we may not even
371 * be able to access it -- so we only want to
372 * require good writes to the new device. But if
373 * the new device turns out to be flaky, we want
374 * to be able to detach it -- which requires all
375 * writes to the old device to have succeeded.
377 if (good_copies == 0 || zio->io_vd == NULL)
378 zio->io_error = vdev_mirror_worst_error(mm);
381 } else if (zio->io_type == ZIO_TYPE_FREE) {
385 ASSERT(zio->io_type == ZIO_TYPE_READ);
388 * If we don't have a good copy yet, keep trying other children.
391 if (good_copies == 0 && (c = vdev_mirror_child_select(zio)) != -1) {
392 ASSERT(c >= 0 && c < mm->mm_children);
393 mc = &mm->mm_child[c];
394 zio_vdev_io_redone(zio);
395 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
396 mc->mc_vd, mc->mc_offset, zio->io_data, zio->io_size,
397 ZIO_TYPE_READ, zio->io_priority, 0,
398 vdev_mirror_child_done, mc));
403 if (good_copies == 0) {
404 zio->io_error = vdev_mirror_worst_error(mm);
405 ASSERT(zio->io_error != 0);
408 if (good_copies && spa_writeable(zio->io_spa) &&
409 (unexpected_errors ||
410 (zio->io_flags & ZIO_FLAG_RESILVER) ||
411 ((zio->io_flags & ZIO_FLAG_SCRUB) && mm->mm_replacing))) {
413 * Use the good data we have in hand to repair damaged children.
415 for (c = 0; c < mm->mm_children; c++) {
417 * Don't rewrite known good children.
418 * Not only is it unnecessary, it could
419 * actually be harmful: if the system lost
420 * power while rewriting the only good copy,
421 * there would be no good copies left!
423 mc = &mm->mm_child[c];
425 if (mc->mc_error == 0) {
428 if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
429 !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
432 mc->mc_error = SET_ERROR(ESTALE);
435 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
436 mc->mc_vd, mc->mc_offset,
437 zio->io_data, zio->io_size,
438 ZIO_TYPE_WRITE, zio->io_priority,
439 ZIO_FLAG_IO_REPAIR | (unexpected_errors ?
440 ZIO_FLAG_SELF_HEAL : 0), NULL, NULL));
446 vdev_mirror_state_change(vdev_t *vd, int faulted, int degraded)
448 if (faulted == vd->vdev_children)
449 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
450 VDEV_AUX_NO_REPLICAS);
451 else if (degraded + faulted != 0)
452 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, VDEV_AUX_NONE);
454 vdev_set_state(vd, B_FALSE, VDEV_STATE_HEALTHY, VDEV_AUX_NONE);
457 vdev_ops_t vdev_mirror_ops = {
461 vdev_mirror_io_start,
463 vdev_mirror_state_change,
466 VDEV_TYPE_MIRROR, /* name of this vdev type */
467 B_FALSE /* not a leaf vdev */
470 vdev_ops_t vdev_replacing_ops = {
474 vdev_mirror_io_start,
476 vdev_mirror_state_change,
479 VDEV_TYPE_REPLACING, /* name of this vdev type */
480 B_FALSE /* not a leaf vdev */
483 vdev_ops_t vdev_spare_ops = {
487 vdev_mirror_io_start,
489 vdev_mirror_state_change,
492 VDEV_TYPE_SPARE, /* name of this vdev type */
493 B_FALSE /* not a leaf vdev */