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_active_tree
));
96 * Avoid inlining the function to keep vdev_mirror_io_start(), which
97 * is this functions only caller, as small as possible on the stack.
99 noinline
static mirror_map_t
*
100 vdev_mirror_map_alloc(zio_t
*zio
)
102 mirror_map_t
*mm
= NULL
;
104 vdev_t
*vd
= zio
->io_vd
;
108 dva_t
*dva
= zio
->io_bp
->blk_dva
;
109 spa_t
*spa
= zio
->io_spa
;
111 c
= BP_GET_NDVAS(zio
->io_bp
);
113 mm
= kmem_zalloc(offsetof(mirror_map_t
, mm_child
[c
]),
116 mm
->mm_replacing
= B_FALSE
;
117 mm
->mm_preferred
= spa_get_random(c
);
118 mm
->mm_root
= B_TRUE
;
121 * Check the other, lower-index DVAs to see if they're on
122 * the same vdev as the child we picked. If they are, use
123 * them since they are likely to have been allocated from
124 * the primary metaslab in use at the time, and hence are
125 * more likely to have locality with single-copy data.
127 for (c
= mm
->mm_preferred
, d
= c
- 1; d
>= 0; d
--) {
128 if (DVA_GET_VDEV(&dva
[d
]) == DVA_GET_VDEV(&dva
[c
]))
129 mm
->mm_preferred
= d
;
132 for (c
= 0; c
< mm
->mm_children
; c
++) {
133 mc
= &mm
->mm_child
[c
];
135 mc
->mc_vd
= vdev_lookup_top(spa
, DVA_GET_VDEV(&dva
[c
]));
136 mc
->mc_offset
= DVA_GET_OFFSET(&dva
[c
]);
139 int lowest_pending
= INT_MAX
;
142 c
= vd
->vdev_children
;
144 mm
= kmem_zalloc(offsetof(mirror_map_t
, mm_child
[c
]),
147 mm
->mm_replacing
= (vd
->vdev_ops
== &vdev_replacing_ops
||
148 vd
->vdev_ops
== &vdev_spare_ops
);
149 mm
->mm_preferred
= 0;
150 mm
->mm_root
= B_FALSE
;
152 for (c
= 0; c
< mm
->mm_children
; c
++) {
153 mc
= &mm
->mm_child
[c
];
154 mc
->mc_vd
= vd
->vdev_child
[c
];
155 mc
->mc_offset
= zio
->io_offset
;
157 if (mm
->mm_replacing
)
160 if (!vdev_readable(mc
->mc_vd
)) {
161 mc
->mc_error
= SET_ERROR(ENXIO
);
164 mc
->mc_pending
= INT_MAX
;
168 mc
->mc_pending
= vdev_mirror_pending(mc
->mc_vd
);
169 if (mc
->mc_pending
< lowest_pending
) {
170 lowest_pending
= mc
->mc_pending
;
172 } else if (mc
->mc_pending
== lowest_pending
) {
177 d
= gethrtime() / (NSEC_PER_USEC
* zfs_vdev_mirror_switch_us
);
178 d
= (d
% lowest_nr
) + 1;
180 for (c
= 0; c
< mm
->mm_children
; c
++) {
181 mc
= &mm
->mm_child
[c
];
183 if (mm
->mm_child
[c
].mc_pending
== lowest_pending
) {
185 mm
->mm_preferred
= c
;
193 zio
->io_vsd_ops
= &vdev_mirror_vsd_ops
;
198 vdev_mirror_open(vdev_t
*vd
, uint64_t *asize
, uint64_t *max_asize
,
205 if (vd
->vdev_children
== 0) {
206 vd
->vdev_stat
.vs_aux
= VDEV_AUX_BAD_LABEL
;
207 return (SET_ERROR(EINVAL
));
210 vdev_open_children(vd
);
212 for (c
= 0; c
< vd
->vdev_children
; c
++) {
213 vdev_t
*cvd
= vd
->vdev_child
[c
];
215 if (cvd
->vdev_open_error
) {
216 lasterror
= cvd
->vdev_open_error
;
221 *asize
= MIN(*asize
- 1, cvd
->vdev_asize
- 1) + 1;
222 *max_asize
= MIN(*max_asize
- 1, cvd
->vdev_max_asize
- 1) + 1;
223 *ashift
= MAX(*ashift
, cvd
->vdev_ashift
);
226 if (numerrors
== vd
->vdev_children
) {
227 vd
->vdev_stat
.vs_aux
= VDEV_AUX_NO_REPLICAS
;
235 vdev_mirror_close(vdev_t
*vd
)
239 for (c
= 0; c
< vd
->vdev_children
; c
++)
240 vdev_close(vd
->vdev_child
[c
]);
244 vdev_mirror_child_done(zio_t
*zio
)
246 mirror_child_t
*mc
= zio
->io_private
;
248 mc
->mc_error
= zio
->io_error
;
254 vdev_mirror_scrub_done(zio_t
*zio
)
256 mirror_child_t
*mc
= zio
->io_private
;
258 if (zio
->io_error
== 0) {
261 mutex_enter(&zio
->io_lock
);
262 while ((pio
= zio_walk_parents(zio
)) != NULL
) {
263 mutex_enter(&pio
->io_lock
);
264 ASSERT3U(zio
->io_size
, >=, pio
->io_size
);
265 bcopy(zio
->io_data
, pio
->io_data
, pio
->io_size
);
266 mutex_exit(&pio
->io_lock
);
268 mutex_exit(&zio
->io_lock
);
271 zio_buf_free(zio
->io_data
, zio
->io_size
);
273 mc
->mc_error
= zio
->io_error
;
279 * Try to find a child whose DTL doesn't contain the block we want to read.
280 * If we can't, try the read on any vdev we haven't already tried.
283 vdev_mirror_child_select(zio_t
*zio
)
285 mirror_map_t
*mm
= zio
->io_vsd
;
287 uint64_t txg
= zio
->io_txg
;
290 ASSERT(zio
->io_bp
== NULL
|| BP_PHYSICAL_BIRTH(zio
->io_bp
) == txg
);
293 * Try to find a child whose DTL doesn't contain the block to read.
294 * If a child is known to be completely inaccessible (indicated by
295 * vdev_readable() returning B_FALSE), don't even try.
297 for (i
= 0, c
= mm
->mm_preferred
; i
< mm
->mm_children
; i
++, c
++) {
298 if (c
>= mm
->mm_children
)
300 mc
= &mm
->mm_child
[c
];
301 if (mc
->mc_tried
|| mc
->mc_skipped
)
303 if (!vdev_readable(mc
->mc_vd
)) {
304 mc
->mc_error
= SET_ERROR(ENXIO
);
305 mc
->mc_tried
= 1; /* don't even try */
309 if (!vdev_dtl_contains(mc
->mc_vd
, DTL_MISSING
, txg
, 1))
311 mc
->mc_error
= SET_ERROR(ESTALE
);
313 mc
->mc_speculative
= 1;
317 * Every device is either missing or has this txg in its DTL.
318 * Look for any child we haven't already tried before giving up.
320 for (c
= 0; c
< mm
->mm_children
; c
++)
321 if (!mm
->mm_child
[c
].mc_tried
)
325 * Every child failed. There's no place left to look.
331 vdev_mirror_io_start(zio_t
*zio
)
337 mm
= vdev_mirror_map_alloc(zio
);
339 if (zio
->io_type
== ZIO_TYPE_READ
) {
340 if ((zio
->io_flags
& ZIO_FLAG_SCRUB
) && !mm
->mm_replacing
) {
342 * For scrubbing reads we need to allocate a read
343 * buffer for each child and issue reads to all
344 * children. If any child succeeds, it will copy its
345 * data into zio->io_data in vdev_mirror_scrub_done.
347 for (c
= 0; c
< mm
->mm_children
; c
++) {
348 mc
= &mm
->mm_child
[c
];
349 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
350 mc
->mc_vd
, mc
->mc_offset
,
351 zio_buf_alloc(zio
->io_size
), zio
->io_size
,
352 zio
->io_type
, zio
->io_priority
, 0,
353 vdev_mirror_scrub_done
, mc
));
355 return (ZIO_PIPELINE_CONTINUE
);
358 * For normal reads just pick one child.
360 c
= vdev_mirror_child_select(zio
);
363 ASSERT(zio
->io_type
== ZIO_TYPE_WRITE
);
366 * Writes go to all children.
369 children
= mm
->mm_children
;
373 mc
= &mm
->mm_child
[c
];
374 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
375 mc
->mc_vd
, mc
->mc_offset
, zio
->io_data
, zio
->io_size
,
376 zio
->io_type
, zio
->io_priority
, 0,
377 vdev_mirror_child_done
, mc
));
381 return (ZIO_PIPELINE_CONTINUE
);
385 vdev_mirror_worst_error(mirror_map_t
*mm
)
387 int c
, error
[2] = { 0, 0 };
389 for (c
= 0; c
< mm
->mm_children
; c
++) {
390 mirror_child_t
*mc
= &mm
->mm_child
[c
];
391 int s
= mc
->mc_speculative
;
392 error
[s
] = zio_worst_error(error
[s
], mc
->mc_error
);
395 return (error
[0] ? error
[0] : error
[1]);
399 vdev_mirror_io_done(zio_t
*zio
)
401 mirror_map_t
*mm
= zio
->io_vsd
;
405 int unexpected_errors
= 0;
407 for (c
= 0; c
< mm
->mm_children
; c
++) {
408 mc
= &mm
->mm_child
[c
];
413 } else if (mc
->mc_tried
) {
418 if (zio
->io_type
== ZIO_TYPE_WRITE
) {
420 * XXX -- for now, treat partial writes as success.
422 * Now that we support write reallocation, it would be better
423 * to treat partial failure as real failure unless there are
424 * no non-degraded top-level vdevs left, and not update DTLs
425 * if we intend to reallocate.
428 if (good_copies
!= mm
->mm_children
) {
430 * Always require at least one good copy.
432 * For ditto blocks (io_vd == NULL), require
433 * all copies to be good.
435 * XXX -- for replacing vdevs, there's no great answer.
436 * If the old device is really dead, we may not even
437 * be able to access it -- so we only want to
438 * require good writes to the new device. But if
439 * the new device turns out to be flaky, we want
440 * to be able to detach it -- which requires all
441 * writes to the old device to have succeeded.
443 if (good_copies
== 0 || zio
->io_vd
== NULL
)
444 zio
->io_error
= vdev_mirror_worst_error(mm
);
449 ASSERT(zio
->io_type
== ZIO_TYPE_READ
);
452 * If we don't have a good copy yet, keep trying other children.
455 if (good_copies
== 0 && (c
= vdev_mirror_child_select(zio
)) != -1) {
456 ASSERT(c
>= 0 && c
< mm
->mm_children
);
457 mc
= &mm
->mm_child
[c
];
458 zio_vdev_io_redone(zio
);
459 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
460 mc
->mc_vd
, mc
->mc_offset
, zio
->io_data
, zio
->io_size
,
461 ZIO_TYPE_READ
, zio
->io_priority
, 0,
462 vdev_mirror_child_done
, mc
));
467 if (good_copies
== 0) {
468 zio
->io_error
= vdev_mirror_worst_error(mm
);
469 ASSERT(zio
->io_error
!= 0);
472 if (good_copies
&& spa_writeable(zio
->io_spa
) &&
473 (unexpected_errors
||
474 (zio
->io_flags
& ZIO_FLAG_RESILVER
) ||
475 ((zio
->io_flags
& ZIO_FLAG_SCRUB
) && mm
->mm_replacing
))) {
477 * Use the good data we have in hand to repair damaged children.
479 for (c
= 0; c
< mm
->mm_children
; c
++) {
481 * Don't rewrite known good children.
482 * Not only is it unnecessary, it could
483 * actually be harmful: if the system lost
484 * power while rewriting the only good copy,
485 * there would be no good copies left!
487 mc
= &mm
->mm_child
[c
];
489 if (mc
->mc_error
== 0) {
492 if (!(zio
->io_flags
& ZIO_FLAG_SCRUB
) &&
493 !vdev_dtl_contains(mc
->mc_vd
, DTL_PARTIAL
,
496 mc
->mc_error
= SET_ERROR(ESTALE
);
499 zio_nowait(zio_vdev_child_io(zio
, zio
->io_bp
,
500 mc
->mc_vd
, mc
->mc_offset
,
501 zio
->io_data
, zio
->io_size
,
502 ZIO_TYPE_WRITE
, ZIO_PRIORITY_ASYNC_WRITE
,
503 ZIO_FLAG_IO_REPAIR
| (unexpected_errors
?
504 ZIO_FLAG_SELF_HEAL
: 0), NULL
, NULL
));
510 vdev_mirror_state_change(vdev_t
*vd
, int faulted
, int degraded
)
512 if (faulted
== vd
->vdev_children
)
513 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_CANT_OPEN
,
514 VDEV_AUX_NO_REPLICAS
);
515 else if (degraded
+ faulted
!= 0)
516 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_DEGRADED
, VDEV_AUX_NONE
);
518 vdev_set_state(vd
, B_FALSE
, VDEV_STATE_HEALTHY
, VDEV_AUX_NONE
);
521 vdev_ops_t vdev_mirror_ops
= {
525 vdev_mirror_io_start
,
527 vdev_mirror_state_change
,
530 VDEV_TYPE_MIRROR
, /* name of this vdev type */
531 B_FALSE
/* not a leaf vdev */
534 vdev_ops_t vdev_replacing_ops
= {
538 vdev_mirror_io_start
,
540 vdev_mirror_state_change
,
543 VDEV_TYPE_REPLACING
, /* name of this vdev type */
544 B_FALSE
/* not a leaf vdev */
547 vdev_ops_t vdev_spare_ops
= {
551 vdev_mirror_io_start
,
553 vdev_mirror_state_change
,
556 VDEV_TYPE_SPARE
, /* name of this vdev type */
557 B_FALSE
/* not a leaf vdev */
560 #if defined(_KERNEL) && defined(HAVE_SPL)
561 module_param(zfs_vdev_mirror_switch_us
, int, 0644);
562 MODULE_PARM_DESC(zfs_vdev_mirror_switch_us
, "Switch mirrors every N usecs");