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 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 #include <sys/zfs_context.h>
28 #include <sys/vdev_impl.h>
30 #include <sys/fs/zfs.h>
33 * Virtual device vector for mirroring.
36 typedef struct mirror_child
{
42 uint8_t mc_speculative
;
45 typedef struct mirror_map
{
50 mirror_child_t mm_child
[1];
53 int vdev_mirror_shift
= 21;
56 vdev_mirror_map_free(zio_t
*zio
)
58 mirror_map_t
*mm
= zio
->io_vsd
;
60 kmem_free(mm
, offsetof(mirror_map_t
, mm_child
[mm
->mm_children
]));
64 vdev_mirror_map_alloc(zio_t
*zio
)
66 mirror_map_t
*mm
= NULL
;
68 vdev_t
*vd
= zio
->io_vd
;
72 dva_t
*dva
= zio
->io_bp
->blk_dva
;
73 spa_t
*spa
= zio
->io_spa
;
75 c
= BP_GET_NDVAS(zio
->io_bp
);
77 mm
= kmem_zalloc(offsetof(mirror_map_t
, mm_child
[c
]), KM_SLEEP
);
79 mm
->mm_replacing
= B_FALSE
;
80 mm
->mm_preferred
= spa_get_random(c
);
84 * Check the other, lower-index DVAs to see if they're on
85 * the same vdev as the child we picked. If they are, use
86 * them since they are likely to have been allocated from
87 * the primary metaslab in use at the time, and hence are
88 * more likely to have locality with single-copy data.
90 for (c
= mm
->mm_preferred
, d
= c
- 1; d
>= 0; d
--) {
91 if (DVA_GET_VDEV(&dva
[d
]) == DVA_GET_VDEV(&dva
[c
]))
95 for (c
= 0; c
< mm
->mm_children
; c
++) {
96 mc
= &mm
->mm_child
[c
];
98 mc
->mc_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[c
]));
99 mc
->mc_offset
= DVA_GET_OFFSET(&dva
[c
]);
102 c
= vd
->vdev_children
;
104 mm
= kmem_zalloc(offsetof(mirror_map_t
, mm_child
[c
]), KM_SLEEP
);
106 mm
->mm_replacing
= (vd
->vdev_ops
== &vdev_replacing_ops
||
107 vd
->vdev_ops
== &vdev_spare_ops
);
108 mm
->mm_preferred
= mm
->mm_replacing
? 0 :
109 (zio
->io_offset
>> vdev_mirror_shift
) % c
;
110 mm
->mm_root
= B_FALSE
;
112 for (c
= 0; c
< mm
->mm_children
; c
++) {
113 mc
= &mm
->mm_child
[c
];
114 mc
->mc_vd
= vd
->vdev_child
[c
];
115 mc
->mc_offset
= zio
->io_offset
;
120 zio
->io_vsd_free
= vdev_mirror_map_free
;
125 vdev_mirror_open(vdev_t
*vd
, uint64_t *asize
, uint64_t *ashift
)
130 int ret
, lasterror
= 0;
132 if (vd
->vdev_children
== 0) {
133 vd
->vdev_stat
.vs_aux
= VDEV_AUX_BAD_LABEL
;
137 for (c
= 0; c
< vd
->vdev_children
; c
++) {
138 cvd
= vd
->vdev_child
[c
];
140 if ((ret
= vdev_open(cvd
)) != 0) {
146 *asize
= MIN(*asize
- 1, cvd
->vdev_asize
- 1) + 1;
147 *ashift
= MAX(*ashift
, cvd
->vdev_ashift
);
150 if (numerrors
== vd
->vdev_children
) {
151 vd
->vdev_stat
.vs_aux
= VDEV_AUX_NO_REPLICAS
;
159 vdev_mirror_close(vdev_t
*vd
)
163 for (c
= 0; c
< vd
->vdev_children
; c
++)
164 vdev_close(vd
->vdev_child
[c
]);
168 vdev_mirror_child_done(zio_t
*zio
)
170 mirror_child_t
*mc
= zio
->io_private
;
172 mc
->mc_error
= zio
->io_error
;
178 vdev_mirror_scrub_done(zio_t
*zio
)
180 mirror_child_t
*mc
= zio
->io_private
;
182 if (zio
->io_error
== 0) {
185 mutex_enter(&zio
->io_lock
);
186 while ((pio
= zio_walk_parents(zio
)) != NULL
) {
187 mutex_enter(&pio
->io_lock
);
188 ASSERT3U(zio
->io_size
, >=, pio
->io_size
);
189 bcopy(zio
->io_data
, pio
->io_data
, pio
->io_size
);
190 mutex_exit(&pio
->io_lock
);
192 mutex_exit(&zio
->io_lock
);
195 zio_buf_free(zio
->io_data
, zio
->io_size
);
197 mc
->mc_error
= zio
->io_error
;
203 * Try to find a child whose DTL doesn't contain the block we want to read.
204 * If we can't, try the read on any vdev we haven't already tried.
207 vdev_mirror_child_select(zio_t
*zio
)
209 mirror_map_t
*mm
= zio
->io_vsd
;
211 uint64_t txg
= zio
->io_txg
;
214 ASSERT(zio
->io_bp
== NULL
|| zio
->io_bp
->blk_birth
== txg
);
217 * Try to find a child whose DTL doesn't contain the block to read.
218 * If a child is known to be completely inaccessible (indicated by
219 * vdev_readable() returning B_FALSE), don't even try.
221 for (i
= 0, c
= mm
->mm_preferred
; i
< mm
->mm_children
; i
++, c
++) {
222 if (c
>= mm
->mm_children
)
224 mc
= &mm
->mm_child
[c
];
225 if (mc
->mc_tried
|| mc
->mc_skipped
)
227 if (!vdev_readable(mc
->mc_vd
)) {
228 mc
->mc_error
= ENXIO
;
229 mc
->mc_tried
= 1; /* don't even try */
233 if (!vdev_dtl_contains(mc
->mc_vd
, DTL_MISSING
, txg
, 1))
235 mc
->mc_error
= ESTALE
;
237 mc
->mc_speculative
= 1;
241 * Every device is either missing or has this txg in its DTL.
242 * Look for any child we haven't already tried before giving up.
244 for (c
= 0; c
< mm
->mm_children
; c
++)
245 if (!mm
->mm_child
[c
].mc_tried
)
249 * Every child failed. There's no place left to look.
255 vdev_mirror_io_start(zio_t
*zio
)
261 mm
= vdev_mirror_map_alloc(zio
);
263 if (zio
->io_type
== ZIO_TYPE_READ
) {
264 if ((zio
->io_flags
& ZIO_FLAG_SCRUB
) && !mm
->mm_replacing
) {
266 * For scrubbing reads we need to allocate a read
267 * buffer for each child and issue reads to all
268 * children. If any child succeeds, it will copy its
269 * data into zio->io_data in vdev_mirror_scrub_done.
271 for (c
= 0; c
< mm
->mm_children
; c
++) {
272 mc
= &mm
->mm_child
[c
];
273 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
274 mc
->mc_vd
, mc
->mc_offset
,
275 zio_buf_alloc(zio
->io_size
), zio
->io_size
,
276 zio
->io_type
, zio
->io_priority
, 0,
277 vdev_mirror_scrub_done
, mc
));
279 return (ZIO_PIPELINE_CONTINUE
);
282 * For normal reads just pick one child.
284 c
= vdev_mirror_child_select(zio
);
287 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
290 * Writes go to all children.
293 children
= mm
->mm_children
;
297 mc
= &mm
->mm_child
[c
];
298 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
299 mc
->mc_vd
, mc
->mc_offset
, zio
->io_data
, zio
->io_size
,
300 zio
->io_type
, zio
->io_priority
, 0,
301 vdev_mirror_child_done
, mc
));
305 return (ZIO_PIPELINE_CONTINUE
);
309 vdev_mirror_worst_error(mirror_map_t
*mm
)
311 int error
[2] = { 0, 0 };
313 for (int c
= 0; c
< mm
->mm_children
; c
++) {
314 mirror_child_t
*mc
= &mm
->mm_child
[c
];
315 int s
= mc
->mc_speculative
;
316 error
[s
] = zio_worst_error(error
[s
], mc
->mc_error
);
319 return (error
[0] ? error
[0] : error
[1]);
323 vdev_mirror_io_done(zio_t
*zio
)
325 mirror_map_t
*mm
= zio
->io_vsd
;
329 int unexpected_errors
= 0;
331 for (c
= 0; c
< mm
->mm_children
; c
++) {
332 mc
= &mm
->mm_child
[c
];
337 } else if (mc
->mc_tried
) {
342 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
344 * XXX -- for now, treat partial writes as success.
346 * Now that we support write reallocation, it would be better
347 * to treat partial failure as real failure unless there are
348 * no non-degraded top-level vdevs left, and not update DTLs
349 * if we intend to reallocate.
352 if (good_copies
!= mm
->mm_children
) {
354 * Always require at least one good copy.
356 * For ditto blocks (io_vd == NULL), require
357 * all copies to be good.
359 * XXX -- for replacing vdevs, there's no great answer.
360 * If the old device is really dead, we may not even
361 * be able to access it -- so we only want to
362 * require good writes to the new device. But if
363 * the new device turns out to be flaky, we want
364 * to be able to detach it -- which requires all
365 * writes to the old device to have succeeded.
367 if (good_copies
== 0 || zio
->io_vd
== NULL
)
368 zio
->io_error
= vdev_mirror_worst_error(mm
);
373 ASSERT(zio
->io_type
== ZIO_TYPE_READ
);
376 * If we don't have a good copy yet, keep trying other children.
379 if (good_copies
== 0 && (c
= vdev_mirror_child_select(zio
)) != -1) {
380 ASSERT(c
>= 0 && c
< mm
->mm_children
);
381 mc
= &mm
->mm_child
[c
];
382 zio_vdev_io_redone(zio
);
383 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
384 mc
->mc_vd
, mc
->mc_offset
, zio
->io_data
, zio
->io_size
,
385 ZIO_TYPE_READ
, zio
->io_priority
, 0,
386 vdev_mirror_child_done
, mc
));
391 if (good_copies
== 0) {
392 zio
->io_error
= vdev_mirror_worst_error(mm
);
393 ASSERT(zio
->io_error
!= 0);
396 if (good_copies
&& spa_writeable(zio
->io_spa
) &&
397 (unexpected_errors
||
398 (zio
->io_flags
& ZIO_FLAG_RESILVER
) ||
399 ((zio
->io_flags
& ZIO_FLAG_SCRUB
) && mm
->mm_replacing
))) {
401 * Use the good data we have in hand to repair damaged children.
403 for (c
= 0; c
< mm
->mm_children
; c
++) {
405 * Don't rewrite known good children.
406 * Not only is it unnecessary, it could
407 * actually be harmful: if the system lost
408 * power while rewriting the only good copy,
409 * there would be no good copies left!
411 mc
= &mm
->mm_child
[c
];
413 if (mc
->mc_error
== 0) {
416 if (!(zio
->io_flags
& ZIO_FLAG_SCRUB
) &&
417 !vdev_dtl_contains(mc
->mc_vd
, DTL_PARTIAL
,
420 mc
->mc_error
= ESTALE
;
423 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
424 mc
->mc_vd
, mc
->mc_offset
,
425 zio
->io_data
, zio
->io_size
,
426 ZIO_TYPE_WRITE
, zio
->io_priority
,
427 ZIO_FLAG_IO_REPAIR
| (unexpected_errors
?
428 ZIO_FLAG_SELF_HEAL
: 0), NULL
, NULL
));
434 vdev_mirror_state_change(vdev_t
*vd
, int faulted
, int degraded
)
436 if (faulted
== vd
->vdev_children
)
437 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_CANT_OPEN
,
438 VDEV_AUX_NO_REPLICAS
);
439 else if (degraded
+ faulted
!= 0)
440 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_DEGRADED
, VDEV_AUX_NONE
);
442 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_HEALTHY
, VDEV_AUX_NONE
);
445 vdev_ops_t vdev_mirror_ops
= {
449 vdev_mirror_io_start
,
451 vdev_mirror_state_change
,
452 VDEV_TYPE_MIRROR
, /* name of this vdev type */
453 B_FALSE
/* not a leaf vdev */
456 vdev_ops_t vdev_replacing_ops
= {
460 vdev_mirror_io_start
,
462 vdev_mirror_state_change
,
463 VDEV_TYPE_REPLACING
, /* name of this vdev type */
464 B_FALSE
/* not a leaf vdev */
467 vdev_ops_t vdev_spare_ops
= {
471 vdev_mirror_io_start
,
473 vdev_mirror_state_change
,
474 VDEV_TYPE_SPARE
, /* name of this vdev type */
475 B_FALSE
/* not a leaf vdev */