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.
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
]));
63 static const zio_vsd_ops_t vdev_mirror_vsd_ops
= {
65 zio_vsd_default_cksum_report
69 vdev_mirror_map_alloc(zio_t
*zio
)
71 mirror_map_t
*mm
= NULL
;
73 vdev_t
*vd
= zio
->io_vd
;
77 dva_t
*dva
= zio
->io_bp
->blk_dva
;
78 spa_t
*spa
= zio
->io_spa
;
80 c
= BP_GET_NDVAS(zio
->io_bp
);
82 mm
= kmem_zalloc(offsetof(mirror_map_t
, mm_child
[c
]), KM_SLEEP
);
84 mm
->mm_replacing
= B_FALSE
;
85 mm
->mm_preferred
= spa_get_random(c
);
89 * Check the other, lower-index DVAs to see if they're on
90 * the same vdev as the child we picked. If they are, use
91 * them since they are likely to have been allocated from
92 * the primary metaslab in use at the time, and hence are
93 * more likely to have locality with single-copy data.
95 for (c
= mm
->mm_preferred
, d
= c
- 1; d
>= 0; d
--) {
96 if (DVA_GET_VDEV(&dva
[d
]) == DVA_GET_VDEV(&dva
[c
]))
100 for (c
= 0; c
< mm
->mm_children
; c
++) {
101 mc
= &mm
->mm_child
[c
];
103 mc
->mc_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[c
]));
104 mc
->mc_offset
= DVA_GET_OFFSET(&dva
[c
]);
107 c
= vd
->vdev_children
;
109 mm
= kmem_zalloc(offsetof(mirror_map_t
, mm_child
[c
]), KM_SLEEP
);
111 mm
->mm_replacing
= (vd
->vdev_ops
== &vdev_replacing_ops
||
112 vd
->vdev_ops
== &vdev_spare_ops
);
113 mm
->mm_preferred
= mm
->mm_replacing
? 0 :
114 (zio
->io_offset
>> vdev_mirror_shift
) % c
;
115 mm
->mm_root
= B_FALSE
;
117 for (c
= 0; c
< mm
->mm_children
; c
++) {
118 mc
= &mm
->mm_child
[c
];
119 mc
->mc_vd
= vd
->vdev_child
[c
];
120 mc
->mc_offset
= zio
->io_offset
;
125 zio
->io_vsd_ops
= &vdev_mirror_vsd_ops
;
130 vdev_mirror_open(vdev_t
*vd
, uint64_t *asize
, uint64_t *ashift
)
135 if (vd
->vdev_children
== 0) {
136 vd
->vdev_stat
.vs_aux
= VDEV_AUX_BAD_LABEL
;
140 vdev_open_children(vd
);
142 for (int c
= 0; c
< vd
->vdev_children
; c
++) {
143 vdev_t
*cvd
= vd
->vdev_child
[c
];
145 if (cvd
->vdev_open_error
) {
146 lasterror
= cvd
->vdev_open_error
;
151 *asize
= MIN(*asize
- 1, cvd
->vdev_asize
- 1) + 1;
152 *ashift
= MAX(*ashift
, cvd
->vdev_ashift
);
155 if (numerrors
== vd
->vdev_children
) {
156 vd
->vdev_stat
.vs_aux
= VDEV_AUX_NO_REPLICAS
;
164 vdev_mirror_close(vdev_t
*vd
)
166 for (int c
= 0; c
< vd
->vdev_children
; c
++)
167 vdev_close(vd
->vdev_child
[c
]);
171 vdev_mirror_child_done(zio_t
*zio
)
173 mirror_child_t
*mc
= zio
->io_private
;
175 mc
->mc_error
= zio
->io_error
;
181 vdev_mirror_scrub_done(zio_t
*zio
)
183 mirror_child_t
*mc
= zio
->io_private
;
185 if (zio
->io_error
== 0) {
188 mutex_enter(&zio
->io_lock
);
189 while ((pio
= zio_walk_parents(zio
)) != NULL
) {
190 mutex_enter(&pio
->io_lock
);
191 ASSERT3U(zio
->io_size
, >=, pio
->io_size
);
192 bcopy(zio
->io_data
, pio
->io_data
, pio
->io_size
);
193 mutex_exit(&pio
->io_lock
);
195 mutex_exit(&zio
->io_lock
);
198 zio_buf_free(zio
->io_data
, zio
->io_size
);
200 mc
->mc_error
= zio
->io_error
;
206 * Try to find a child whose DTL doesn't contain the block we want to read.
207 * If we can't, try the read on any vdev we haven't already tried.
210 vdev_mirror_child_select(zio_t
*zio
)
212 mirror_map_t
*mm
= zio
->io_vsd
;
214 uint64_t txg
= zio
->io_txg
;
217 ASSERT(zio
->io_bp
== NULL
|| BP_PHYSICAL_BIRTH(zio
->io_bp
) == txg
);
220 * Try to find a child whose DTL doesn't contain the block to read.
221 * If a child is known to be completely inaccessible (indicated by
222 * vdev_readable() returning B_FALSE), don't even try.
224 for (i
= 0, c
= mm
->mm_preferred
; i
< mm
->mm_children
; i
++, c
++) {
225 if (c
>= mm
->mm_children
)
227 mc
= &mm
->mm_child
[c
];
228 if (mc
->mc_tried
|| mc
->mc_skipped
)
230 if (!vdev_readable(mc
->mc_vd
)) {
231 mc
->mc_error
= ENXIO
;
232 mc
->mc_tried
= 1; /* don't even try */
236 if (!vdev_dtl_contains(mc
->mc_vd
, DTL_MISSING
, txg
, 1))
238 mc
->mc_error
= ESTALE
;
240 mc
->mc_speculative
= 1;
244 * Every device is either missing or has this txg in its DTL.
245 * Look for any child we haven't already tried before giving up.
247 for (c
= 0; c
< mm
->mm_children
; c
++)
248 if (!mm
->mm_child
[c
].mc_tried
)
252 * Every child failed. There's no place left to look.
258 vdev_mirror_io_start(zio_t
*zio
)
264 mm
= vdev_mirror_map_alloc(zio
);
266 if (zio
->io_type
== ZIO_TYPE_READ
) {
267 if ((zio
->io_flags
& ZIO_FLAG_SCRUB
) && !mm
->mm_replacing
) {
269 * For scrubbing reads we need to allocate a read
270 * buffer for each child and issue reads to all
271 * children. If any child succeeds, it will copy its
272 * data into zio->io_data in vdev_mirror_scrub_done.
274 for (c
= 0; c
< mm
->mm_children
; c
++) {
275 mc
= &mm
->mm_child
[c
];
276 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
277 mc
->mc_vd
, mc
->mc_offset
,
278 zio_buf_alloc(zio
->io_size
), zio
->io_size
,
279 zio
->io_type
, zio
->io_priority
, 0,
280 vdev_mirror_scrub_done
, mc
));
282 return (ZIO_PIPELINE_CONTINUE
);
285 * For normal reads just pick one child.
287 c
= vdev_mirror_child_select(zio
);
290 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
293 * Writes go to all children.
296 children
= mm
->mm_children
;
300 mc
= &mm
->mm_child
[c
];
301 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
302 mc
->mc_vd
, mc
->mc_offset
, zio
->io_data
, zio
->io_size
,
303 zio
->io_type
, zio
->io_priority
, 0,
304 vdev_mirror_child_done
, mc
));
308 return (ZIO_PIPELINE_CONTINUE
);
312 vdev_mirror_worst_error(mirror_map_t
*mm
)
314 int error
[2] = { 0, 0 };
316 for (int c
= 0; c
< mm
->mm_children
; c
++) {
317 mirror_child_t
*mc
= &mm
->mm_child
[c
];
318 int s
= mc
->mc_speculative
;
319 error
[s
] = zio_worst_error(error
[s
], mc
->mc_error
);
322 return (error
[0] ? error
[0] : error
[1]);
326 vdev_mirror_io_done(zio_t
*zio
)
328 mirror_map_t
*mm
= zio
->io_vsd
;
332 int unexpected_errors
= 0;
334 for (c
= 0; c
< mm
->mm_children
; c
++) {
335 mc
= &mm
->mm_child
[c
];
340 } else if (mc
->mc_tried
) {
345 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
347 * XXX -- for now, treat partial writes as success.
349 * Now that we support write reallocation, it would be better
350 * to treat partial failure as real failure unless there are
351 * no non-degraded top-level vdevs left, and not update DTLs
352 * if we intend to reallocate.
355 if (good_copies
!= mm
->mm_children
) {
357 * Always require at least one good copy.
359 * For ditto blocks (io_vd == NULL), require
360 * all copies to be good.
362 * XXX -- for replacing vdevs, there's no great answer.
363 * If the old device is really dead, we may not even
364 * be able to access it -- so we only want to
365 * require good writes to the new device. But if
366 * the new device turns out to be flaky, we want
367 * to be able to detach it -- which requires all
368 * writes to the old device to have succeeded.
370 if (good_copies
== 0 || zio
->io_vd
== NULL
)
371 zio
->io_error
= vdev_mirror_worst_error(mm
);
376 ASSERT(zio
->io_type
== ZIO_TYPE_READ
);
379 * If we don't have a good copy yet, keep trying other children.
382 if (good_copies
== 0 && (c
= vdev_mirror_child_select(zio
)) != -1) {
383 ASSERT(c
>= 0 && c
< mm
->mm_children
);
384 mc
= &mm
->mm_child
[c
];
385 zio_vdev_io_redone(zio
);
386 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
387 mc
->mc_vd
, mc
->mc_offset
, zio
->io_data
, zio
->io_size
,
388 ZIO_TYPE_READ
, zio
->io_priority
, 0,
389 vdev_mirror_child_done
, mc
));
394 if (good_copies
== 0) {
395 zio
->io_error
= vdev_mirror_worst_error(mm
);
396 ASSERT(zio
->io_error
!= 0);
399 if (good_copies
&& spa_writeable(zio
->io_spa
) &&
400 (unexpected_errors
||
401 (zio
->io_flags
& ZIO_FLAG_RESILVER
) ||
402 ((zio
->io_flags
& ZIO_FLAG_SCRUB
) && mm
->mm_replacing
))) {
404 * Use the good data we have in hand to repair damaged children.
406 for (c
= 0; c
< mm
->mm_children
; c
++) {
408 * Don't rewrite known good children.
409 * Not only is it unnecessary, it could
410 * actually be harmful: if the system lost
411 * power while rewriting the only good copy,
412 * there would be no good copies left!
414 mc
= &mm
->mm_child
[c
];
416 if (mc
->mc_error
== 0) {
419 if (!(zio
->io_flags
& ZIO_FLAG_SCRUB
) &&
420 !vdev_dtl_contains(mc
->mc_vd
, DTL_PARTIAL
,
423 mc
->mc_error
= ESTALE
;
426 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
427 mc
->mc_vd
, mc
->mc_offset
,
428 zio
->io_data
, zio
->io_size
,
429 ZIO_TYPE_WRITE
, zio
->io_priority
,
430 ZIO_FLAG_IO_REPAIR
| (unexpected_errors
?
431 ZIO_FLAG_SELF_HEAL
: 0), NULL
, NULL
));
437 vdev_mirror_state_change(vdev_t
*vd
, int faulted
, int degraded
)
439 if (faulted
== vd
->vdev_children
)
440 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_CANT_OPEN
,
441 VDEV_AUX_NO_REPLICAS
);
442 else if (degraded
+ faulted
!= 0)
443 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_DEGRADED
, VDEV_AUX_NONE
);
445 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_HEALTHY
, VDEV_AUX_NONE
);
448 vdev_ops_t vdev_mirror_ops
= {
452 vdev_mirror_io_start
,
454 vdev_mirror_state_change
,
457 VDEV_TYPE_MIRROR
, /* name of this vdev type */
458 B_FALSE
/* not a leaf vdev */
461 vdev_ops_t vdev_replacing_ops
= {
465 vdev_mirror_io_start
,
467 vdev_mirror_state_change
,
470 VDEV_TYPE_REPLACING
, /* name of this vdev type */
471 B_FALSE
/* not a leaf vdev */
474 vdev_ops_t vdev_spare_ops
= {
478 vdev_mirror_io_start
,
480 vdev_mirror_state_change
,
483 VDEV_TYPE_SPARE
, /* name of this vdev type */
484 B_FALSE
/* not a leaf vdev */