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
{
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 return avl_numnodes(&vd
->vdev_queue
.vq_pending_tree
);
96 vdev_mirror_map_alloc(zio_t
*zio
)
98 mirror_map_t
*mm
= NULL
;
100 vdev_t
*vd
= zio
->io_vd
;
104 dva_t
*dva
= zio
->io_bp
->blk_dva
;
105 spa_t
*spa
= zio
->io_spa
;
107 c
= BP_GET_NDVAS(zio
->io_bp
);
109 mm
= kmem_zalloc(offsetof(mirror_map_t
, mm_child
[c
]), KM_PUSHPAGE
);
111 mm
->mm_replacing
= B_FALSE
;
112 mm
->mm_preferred
= spa_get_random(c
);
113 mm
->mm_root
= B_TRUE
;
116 * Check the other, lower-index DVAs to see if they're on
117 * the same vdev as the child we picked. If they are, use
118 * them since they are likely to have been allocated from
119 * the primary metaslab in use at the time, and hence are
120 * more likely to have locality with single-copy data.
122 for (c
= mm
->mm_preferred
, d
= c
- 1; d
>= 0; d
--) {
123 if (DVA_GET_VDEV(&dva
[d
]) == DVA_GET_VDEV(&dva
[c
]))
124 mm
->mm_preferred
= d
;
127 for (c
= 0; c
< mm
->mm_children
; c
++) {
128 mc
= &mm
->mm_child
[c
];
130 mc
->mc_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[c
]));
131 mc
->mc_offset
= DVA_GET_OFFSET(&dva
[c
]);
134 int lowest_pending
= INT_MAX
;
137 c
= vd
->vdev_children
;
139 mm
= kmem_zalloc(offsetof(mirror_map_t
, mm_child
[c
]), KM_PUSHPAGE
);
141 mm
->mm_replacing
= (vd
->vdev_ops
== &vdev_replacing_ops
||
142 vd
->vdev_ops
== &vdev_spare_ops
);
143 mm
->mm_preferred
= 0;
144 mm
->mm_root
= B_FALSE
;
146 for (c
= 0; c
< mm
->mm_children
; c
++) {
147 mc
= &mm
->mm_child
[c
];
148 mc
->mc_vd
= vd
->vdev_child
[c
];
149 mc
->mc_offset
= zio
->io_offset
;
151 if (mm
->mm_replacing
)
154 if (!vdev_readable(mc
->mc_vd
)) {
155 mc
->mc_error
= SET_ERROR(ENXIO
);
158 mc
->mc_pending
= INT_MAX
;
162 mc
->mc_pending
= vdev_mirror_pending(mc
->mc_vd
);
163 if (mc
->mc_pending
< lowest_pending
) {
164 lowest_pending
= mc
->mc_pending
;
166 } else if (mc
->mc_pending
== lowest_pending
) {
171 d
= gethrtime() / (NSEC_PER_USEC
* zfs_vdev_mirror_switch_us
);
172 d
= (d
% lowest_nr
) + 1;
174 for (c
= 0; c
< mm
->mm_children
; c
++) {
175 mc
= &mm
->mm_child
[c
];
177 if (mm
->mm_child
[c
].mc_pending
== lowest_pending
) {
179 mm
->mm_preferred
= c
;
187 zio
->io_vsd_ops
= &vdev_mirror_vsd_ops
;
192 vdev_mirror_open(vdev_t
*vd
, uint64_t *asize
, uint64_t *max_asize
,
199 if (vd
->vdev_children
== 0) {
200 vd
->vdev_stat
.vs_aux
= VDEV_AUX_BAD_LABEL
;
201 return (SET_ERROR(EINVAL
));
204 vdev_open_children(vd
);
206 for (c
= 0; c
< vd
->vdev_children
; c
++) {
207 vdev_t
*cvd
= vd
->vdev_child
[c
];
209 if (cvd
->vdev_open_error
) {
210 lasterror
= cvd
->vdev_open_error
;
215 *asize
= MIN(*asize
- 1, cvd
->vdev_asize
- 1) + 1;
216 *max_asize
= MIN(*max_asize
- 1, cvd
->vdev_max_asize
- 1) + 1;
217 *ashift
= MAX(*ashift
, cvd
->vdev_ashift
);
220 if (numerrors
== vd
->vdev_children
) {
221 vd
->vdev_stat
.vs_aux
= VDEV_AUX_NO_REPLICAS
;
229 vdev_mirror_close(vdev_t
*vd
)
233 for (c
= 0; c
< vd
->vdev_children
; c
++)
234 vdev_close(vd
->vdev_child
[c
]);
238 vdev_mirror_child_done(zio_t
*zio
)
240 mirror_child_t
*mc
= zio
->io_private
;
242 mc
->mc_error
= zio
->io_error
;
248 vdev_mirror_scrub_done(zio_t
*zio
)
250 mirror_child_t
*mc
= zio
->io_private
;
252 if (zio
->io_error
== 0) {
255 mutex_enter(&zio
->io_lock
);
256 while ((pio
= zio_walk_parents(zio
)) != NULL
) {
257 mutex_enter(&pio
->io_lock
);
258 ASSERT3U(zio
->io_size
, >=, pio
->io_size
);
259 bcopy(zio
->io_data
, pio
->io_data
, pio
->io_size
);
260 mutex_exit(&pio
->io_lock
);
262 mutex_exit(&zio
->io_lock
);
265 zio_buf_free(zio
->io_data
, zio
->io_size
);
267 mc
->mc_error
= zio
->io_error
;
273 * Try to find a child whose DTL doesn't contain the block we want to read.
274 * If we can't, try the read on any vdev we haven't already tried.
277 vdev_mirror_child_select(zio_t
*zio
)
279 mirror_map_t
*mm
= zio
->io_vsd
;
281 uint64_t txg
= zio
->io_txg
;
284 ASSERT(zio
->io_bp
== NULL
|| BP_PHYSICAL_BIRTH(zio
->io_bp
) == txg
);
287 * Try to find a child whose DTL doesn't contain the block to read.
288 * If a child is known to be completely inaccessible (indicated by
289 * vdev_readable() returning B_FALSE), don't even try.
291 for (i
= 0, c
= mm
->mm_preferred
; i
< mm
->mm_children
; i
++, c
++) {
292 if (c
>= mm
->mm_children
)
294 mc
= &mm
->mm_child
[c
];
295 if (mc
->mc_tried
|| mc
->mc_skipped
)
297 if (!vdev_readable(mc
->mc_vd
)) {
298 mc
->mc_error
= SET_ERROR(ENXIO
);
299 mc
->mc_tried
= 1; /* don't even try */
303 if (!vdev_dtl_contains(mc
->mc_vd
, DTL_MISSING
, txg
, 1))
305 mc
->mc_error
= SET_ERROR(ESTALE
);
307 mc
->mc_speculative
= 1;
311 * Every device is either missing or has this txg in its DTL.
312 * Look for any child we haven't already tried before giving up.
314 for (c
= 0; c
< mm
->mm_children
; c
++)
315 if (!mm
->mm_child
[c
].mc_tried
)
319 * Every child failed. There's no place left to look.
325 vdev_mirror_io_start(zio_t
*zio
)
331 mm
= vdev_mirror_map_alloc(zio
);
333 if (zio
->io_type
== ZIO_TYPE_READ
) {
334 if ((zio
->io_flags
& ZIO_FLAG_SCRUB
) && !mm
->mm_replacing
) {
336 * For scrubbing reads we need to allocate a read
337 * buffer for each child and issue reads to all
338 * children. If any child succeeds, it will copy its
339 * data into zio->io_data in vdev_mirror_scrub_done.
341 for (c
= 0; c
< mm
->mm_children
; c
++) {
342 mc
= &mm
->mm_child
[c
];
343 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
344 mc
->mc_vd
, mc
->mc_offset
,
345 zio_buf_alloc(zio
->io_size
), zio
->io_size
,
346 zio
->io_type
, zio
->io_priority
, 0,
347 vdev_mirror_scrub_done
, mc
));
349 return (ZIO_PIPELINE_CONTINUE
);
352 * For normal reads just pick one child.
354 c
= vdev_mirror_child_select(zio
);
357 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
360 * Writes go to all children.
363 children
= mm
->mm_children
;
367 mc
= &mm
->mm_child
[c
];
368 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
369 mc
->mc_vd
, mc
->mc_offset
, zio
->io_data
, zio
->io_size
,
370 zio
->io_type
, zio
->io_priority
, 0,
371 vdev_mirror_child_done
, mc
));
375 return (ZIO_PIPELINE_CONTINUE
);
379 vdev_mirror_worst_error(mirror_map_t
*mm
)
381 int c
, error
[2] = { 0, 0 };
383 for (c
= 0; c
< mm
->mm_children
; c
++) {
384 mirror_child_t
*mc
= &mm
->mm_child
[c
];
385 int s
= mc
->mc_speculative
;
386 error
[s
] = zio_worst_error(error
[s
], mc
->mc_error
);
389 return (error
[0] ? error
[0] : error
[1]);
393 vdev_mirror_io_done(zio_t
*zio
)
395 mirror_map_t
*mm
= zio
->io_vsd
;
399 int unexpected_errors
= 0;
401 for (c
= 0; c
< mm
->mm_children
; c
++) {
402 mc
= &mm
->mm_child
[c
];
407 } else if (mc
->mc_tried
) {
412 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
414 * XXX -- for now, treat partial writes as success.
416 * Now that we support write reallocation, it would be better
417 * to treat partial failure as real failure unless there are
418 * no non-degraded top-level vdevs left, and not update DTLs
419 * if we intend to reallocate.
422 if (good_copies
!= mm
->mm_children
) {
424 * Always require at least one good copy.
426 * For ditto blocks (io_vd == NULL), require
427 * all copies to be good.
429 * XXX -- for replacing vdevs, there's no great answer.
430 * If the old device is really dead, we may not even
431 * be able to access it -- so we only want to
432 * require good writes to the new device. But if
433 * the new device turns out to be flaky, we want
434 * to be able to detach it -- which requires all
435 * writes to the old device to have succeeded.
437 if (good_copies
== 0 || zio
->io_vd
== NULL
)
438 zio
->io_error
= vdev_mirror_worst_error(mm
);
443 ASSERT(zio
->io_type
== ZIO_TYPE_READ
);
446 * If we don't have a good copy yet, keep trying other children.
449 if (good_copies
== 0 && (c
= vdev_mirror_child_select(zio
)) != -1) {
450 ASSERT(c
>= 0 && c
< mm
->mm_children
);
451 mc
= &mm
->mm_child
[c
];
452 zio_vdev_io_redone(zio
);
453 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
454 mc
->mc_vd
, mc
->mc_offset
, zio
->io_data
, zio
->io_size
,
455 ZIO_TYPE_READ
, zio
->io_priority
, 0,
456 vdev_mirror_child_done
, mc
));
461 if (good_copies
== 0) {
462 zio
->io_error
= vdev_mirror_worst_error(mm
);
463 ASSERT(zio
->io_error
!= 0);
466 if (good_copies
&& spa_writeable(zio
->io_spa
) &&
467 (unexpected_errors
||
468 (zio
->io_flags
& ZIO_FLAG_RESILVER
) ||
469 ((zio
->io_flags
& ZIO_FLAG_SCRUB
) && mm
->mm_replacing
))) {
471 * Use the good data we have in hand to repair damaged children.
473 for (c
= 0; c
< mm
->mm_children
; c
++) {
475 * Don't rewrite known good children.
476 * Not only is it unnecessary, it could
477 * actually be harmful: if the system lost
478 * power while rewriting the only good copy,
479 * there would be no good copies left!
481 mc
= &mm
->mm_child
[c
];
483 if (mc
->mc_error
== 0) {
486 if (!(zio
->io_flags
& ZIO_FLAG_SCRUB
) &&
487 !vdev_dtl_contains(mc
->mc_vd
, DTL_PARTIAL
,
490 mc
->mc_error
= SET_ERROR(ESTALE
);
493 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
494 mc
->mc_vd
, mc
->mc_offset
,
495 zio
->io_data
, zio
->io_size
,
496 ZIO_TYPE_WRITE
, zio
->io_priority
,
497 ZIO_FLAG_IO_REPAIR
| (unexpected_errors
?
498 ZIO_FLAG_SELF_HEAL
: 0), NULL
, NULL
));
504 vdev_mirror_state_change(vdev_t
*vd
, int faulted
, int degraded
)
506 if (faulted
== vd
->vdev_children
)
507 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_CANT_OPEN
,
508 VDEV_AUX_NO_REPLICAS
);
509 else if (degraded
+ faulted
!= 0)
510 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_DEGRADED
, VDEV_AUX_NONE
);
512 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_HEALTHY
, VDEV_AUX_NONE
);
515 vdev_ops_t vdev_mirror_ops
= {
519 vdev_mirror_io_start
,
521 vdev_mirror_state_change
,
524 VDEV_TYPE_MIRROR
, /* name of this vdev type */
525 B_FALSE
/* not a leaf vdev */
528 vdev_ops_t vdev_replacing_ops
= {
532 vdev_mirror_io_start
,
534 vdev_mirror_state_change
,
537 VDEV_TYPE_REPLACING
, /* name of this vdev type */
538 B_FALSE
/* not a leaf vdev */
541 vdev_ops_t vdev_spare_ops
= {
545 vdev_mirror_io_start
,
547 vdev_mirror_state_change
,
550 VDEV_TYPE_SPARE
, /* name of this vdev type */
551 B_FALSE
/* not a leaf vdev */
554 #if defined(_KERNEL) && defined(HAVE_SPL)
555 module_param(zfs_vdev_mirror_switch_us
, int, 0644);
556 MODULE_PARM_DESC(zfs_vdev_mirror_switch_us
, "Switch mirrors every N usecs");