1MD(4) Kernel Interfaces Manual MD(4)
2
6 md - Multiple Device driver aka Linux Software RAID
7
9 /dev/mdn
10 /dev/md/n
11 /dev/md/name
12
14 The md driver provides virtual devices that are created from one or
15 more independent underlying devices. This array of devices often con‐
16 tains redundancy and the devices are often disk drives, hence the acro‐
17 nym RAID which stands for a Redundant Array of Independent Disks.
18 md supports RAID levels 1 (mirroring), 4 (striped array with parity
20 device), 5 (striped array with distributed parity information), 6
21 (striped array with distributed dual redundancy information), and 10
22 (striped and mirrored). If some number of underlying devices fails
23 while using one of these levels, the array will continue to function;
24 this number is one for RAID levels 4 and 5, two for RAID level 6, and
25 all but one (N-1) for RAID level 1, and dependent on configuration for
26 level 10.
27 md also supports a number of pseudo RAID (non-redundant) configurations
29 including RAID0 (striped array), LINEAR (catenated array), MULTIPATH (a
30 set of different interfaces to the same device), and FAULTY (a layer
31 over a single device into which errors can be injected).
32 MD METADATA
35 Each device in an array may have some metadata stored in the device.
36 This metadata is sometimes called a superblock. The metadata records
37 information about the structure and state of the array. This allows
38 the array to be reliably re-assembled after a shutdown.
39 From Linux kernel version 2.6.10, md provides support for two different
41 formats of metadata, and other formats can be added. Prior to this
42 release, only one format is supported.
43 The common format — known as version 0.90 — has a superblock that is 4K
45 long and is written into a 64K aligned block that starts at least 64K
46 and less than 128K from the end of the device (i.e. to get the address
47 of the superblock round the size of the device down to a multiple of
48 64K and then subtract 64K). The available size of each device is the
49 amount of space before the super block, so between 64K and 128K is lost
50 when a device in incorporated into an MD array. This superblock stores
51 multi-byte fields in a processor-dependent manner, so arrays cannot
52 easily be moved between computers with different processors.
53 The new format — known as version 1 — has a superblock that is normally
55 1K long, but can be longer. It is normally stored between 8K and 12K
56 from the end of the device, on a 4K boundary, though variations can be
57 stored at the start of the device (version 1.1) or 4K from the start of
58 the device (version 1.2). This metadata format stores multibyte data
59 in a processor-independent format and supports up to hundreds of compo‐
60 nent devices (version 0.90 only supports 28).
61 The metadata contains, among other things:
63 LEVEL The manner in which the devices are arranged into the array
65 (LINEAR, RAID0, RAID1, RAID4, RAID5, RAID10, MULTIPATH).
66 UUID a 128 bit Universally Unique Identifier that identifies the
68 array that contains this device.
69 When a version 0.90 array is being reshaped (e.g. adding extra devices
72 to a RAID5), the version number is temporarily set to 0.91. This
73 ensures that if the reshape process is stopped in the middle (e.g. by a
74 system crash) and the machine boots into an older kernel that does not
75 support reshaping, then the array will not be assembled (which would
76 cause data corruption) but will be left untouched until a kernel that
77 can complete the reshape processes is used.
78 ARRAYS WITHOUT METADATA
81 While it is usually best to create arrays with superblocks so that they
82 can be assembled reliably, there are some circumstances when an array
83 without superblocks is preferred. These include:
84 LEGACY ARRAYS
86 Early versions of the md driver only supported LINEAR and RAID0
87 configurations and did not use a superblock (which is less crit‐
88 ical with these configurations). While such arrays should be
89 rebuilt with superblocks if possible, md continues to support
90 them.
91 FAULTY Being a largely transparent layer over a different device, the
93 FAULTY personality doesn't gain anything from having a
94 superblock.
95 MULTIPATH
97 It is often possible to detect devices which are different paths
98 to the same storage directly rather than having a distinctive
99 superblock written to the device and searched for on all paths.
100 In this case, a MULTIPATH array with no superblock makes sense.
101 RAID1 In some configurations it might be desired to create a RAID1
103 configuration that does not use a superblock, and to maintain
104 the state of the array elsewhere. While not encouraged for gen‐
105 eral use, it does have special-purpose uses and is supported.
106 ARRAYS WITH EXTERNAL METADATA
109 From release 2.6.28, the md driver supports arrays with externally man‐
110 aged metadata. That is, the metadata is not managed by the kernel but
111 rather by a user-space program which is external to the kernel. This
112 allows support for a variety of metadata formats without cluttering the
113 kernel with lots of details.
114 md is able to communicate with the user-space program through various
116 sysfs attributes so that it can make appropriate changes to the meta‐
117 data - for example to mark a device as faulty. When necessary, md will
118 wait for the program to acknowledge the event by writing to a sysfs
119 attribute. The manual page for mdmon(8) contains more detail about
120 this interaction.
121 CONTAINERS
124 Many metadata formats use a single block of metadata to describe a num‐
125 ber of different arrays which all use the same set of devices. In this
126 case it is helpful for the kernel to know about the full set of devices
127 as a whole. This set is known to md as a container. A container is an
128 md array with externally managed metadata and with device offset and
129 size so that it just covers the metadata part of the devices. The
130 remainder of each device is available to be incorporated into various
131 arrays.
132 LINEAR
135 A LINEAR array simply catenates the available space on each drive to
136 form one large virtual drive.
137 One advantage of this arrangement over the more common RAID0 arrange‐
139 ment is that the array may be reconfigured at a later time with an
140 extra drive, so the array is made bigger without disturbing the data
141 that is on the array. This can even be done on a live array.
142 If a chunksize is given with a LINEAR array, the usable space on each
144 device is rounded down to a multiple of this chunksize.
145 RAID0
148 A RAID0 array (which has zero redundancy) is also known as a striped
149 array. A RAID0 array is configured at creation with a Chunk Size which
150 must be a power of two (prior to Linux 2.6.31), and at least 4
151 kibibytes.
152 The RAID0 driver assigns the first chunk of the array to the first
154 device, the second chunk to the second device, and so on until all
155 drives have been assigned one chunk. This collection of chunks forms a
156 stripe. Further chunks are gathered into stripes in the same way, and
157 are assigned to the remaining space in the drives.
158 If devices in the array are not all the same size, then once the small‐
160 est device has been exhausted, the RAID0 driver starts collecting
161 chunks into smaller stripes that only span the drives which still have
162 remaining space.
163 RAID1
167 A RAID1 array is also known as a mirrored set (though mirrors tend to
168 provide reflected images, which RAID1 does not) or a plex.
169 Once initialised, each device in a RAID1 array contains exactly the
171 same data. Changes are written to all devices in parallel. Data is
172 read from any one device. The driver attempts to distribute read
173 requests across all devices to maximise performance.
174 All devices in a RAID1 array should be the same size. If they are not,
176 then only the amount of space available on the smallest device is used
177 (any extra space on other devices is wasted).
178 Note that the read balancing done by the driver does not make the RAID1
180 performance profile be the same as for RAID0; a single stream of
181 sequential input will not be accelerated (e.g. a single dd), but multi‐
182 ple sequential streams or a random workload will use more than one
183 spindle. In theory, having an N-disk RAID1 will allow N sequential
184 threads to read from all disks.
185 Individual devices in a RAID1 can be marked as "write-mostly". These
187 drives are excluded from the normal read balancing and will only be
188 read from when there is no other option. This can be useful for
189 devices connected over a slow link.
190 RAID4
193 A RAID4 array is like a RAID0 array with an extra device for storing
194 parity. This device is the last of the active devices in the array.
195 Unlike RAID0, RAID4 also requires that all stripes span all drives, so
196 extra space on devices that are larger than the smallest is wasted.
197 When any block in a RAID4 array is modified, the parity block for that
199 stripe (i.e. the block in the parity device at the same device offset
200 as the stripe) is also modified so that the parity block always con‐
201 tains the "parity" for the whole stripe. I.e. its content is equiva‐
202 lent to the result of performing an exclusive-or operation between all
203 the data blocks in the stripe.
204 This allows the array to continue to function if one device fails. The
206 data that was on that device can be calculated as needed from the par‐
207 ity block and the other data blocks.
208 RAID5
211 RAID5 is very similar to RAID4. The difference is that the parity
212 blocks for each stripe, instead of being on a single device, are dis‐
213 tributed across all devices. This allows more parallelism when writ‐
214 ing, as two different block updates will quite possibly affect parity
215 blocks on different devices so there is less contention.
216 This also allows more parallelism when reading, as read requests are
218 distributed over all the devices in the array instead of all but one.
219 RAID6
222 RAID6 is similar to RAID5, but can handle the loss of any two devices
223 without data loss. Accordingly, it requires N+2 drives to store N
224 drives worth of data.
225 The performance for RAID6 is slightly lower but comparable to RAID5 in
227 normal mode and single disk failure mode. It is very slow in dual disk
228 failure mode, however.
229 RAID10
232 RAID10 provides a combination of RAID1 and RAID0, and is sometimes
233 known as RAID1+0. Every datablock is duplicated some number of times,
234 and the resulting collection of datablocks are distributed over multi‐
235 ple drives.
236 When configuring a RAID10 array, it is necessary to specify the number
238 of replicas of each data block that are required (this will usually
239 be 2) and whether their layout should be "near", "far" or "offset"
240 (with "offset" being available since Linux 2.6.18).
241 About the RAID10 Layout Examples:
243 The examples below visualise the chunk distribution on the underlying
244 devices for the respective layout.
245 For simplicity it is assumed that the size of the chunks equals the
247 size of the blocks of the underlying devices as well as those of the
248 RAID10 device exported by the kernel (for example /dev/md/name).
249 Therefore the chunks / chunk numbers map directly to the blocks /block
250 addresses of the exported RAID10 device.
251 Decimal numbers (0, 1, 2, ...) are the chunks of the RAID10 and due to
253 the above assumption also the blocks and block addresses of the
254 exported RAID10 device.
255 Repeated numbers mean copies of a chunk / block (obviously on different
256 underlying devices).
257 Hexadecimal numbers (0x00, 0x01, 0x02, ...) are the block addresses of
258 the underlying devices.
259 "near" Layout
262 When "near" replicas are chosen, the multiple copies of a given
263 chunk are laid out consecutively ("as close to each other as
264 possible") across the stripes of the array.
265 With an even number of devices, they will likely (unless some
267 misalignment is present) lay at the very same offset on the dif‐
268 ferent devices.
269 This is as the "classic" RAID1+0; that is two groups of mirrored
270 devices (in the example below the groups Device #1 / #2 and
271 Device #3 / #4 are each a RAID1) both in turn forming a striped
272 RAID0.
273 Example with 2 copies per chunk and an even number (4) of
275 devices:
276 ┌───────────┌───────────┌───────────┌───────────┐
278 │ Device #1 │ Device #2 │ Device #3 │ Device #4 │
279 ┌─────├───────────├───────────├───────────├───────────┤
280 │0x00 │ 0 │ 0 │ 1 │ 1 │
281 │0x01 │ 2 │ 2 │ 3 │ 3 │
282 │... │ ... │ ... │ ... │ ... │
283 │ : │ : │ : │ : │ : │
284 │... │ ... │ ... │ ... │ ... │
285 │0x80 │ 254 │ 254 │ 255 │ 255 │
286 └─────└───────────└───────────└───────────└───────────┘
287 \---------v---------/ \---------v---------/
288 RAID1 RAID1
289 \---------------------v---------------------/
290 RAID0
291 Example with 2 copies per chunk and an odd number (5) of
293 devices:
294 ┌────────┌────────┌────────┌────────┌────────┐
296 │ Dev #1 │ Dev #2 │ Dev #3 │ Dev #4 │ Dev #5 │
297 ┌─────├────────├────────├────────├────────├────────┤
298 │0x00 │ 0 │ 0 │ 1 │ 1 │ 2 │
299 │0x01 │ 2 │ 3 │ 3 │ 4 │ 4 │
300 │... │ ... │ ... │ ... │ ... │ ... │
301 │ : │ : │ : │ : │ : │ : │
302 │... │ ... │ ... │ ... │ ... │ ... │
303 │0x80 │ 317 │ 318 │ 318 │ 319 │ 319 │
304 └─────└────────└────────└────────└────────└────────┘
305 "far" Layout
309 When "far" replicas are chosen, the multiple copies of a given
310 chunk are laid out quite distant ("as far as reasonably possi‐
311 ble") from each other.
312 First a complete sequence of all data blocks (that is all the
314 data one sees on the exported RAID10 block device) is striped
315 over the devices. Then another (though "shifted") complete
316 sequence of all data blocks; and so on (in the case of more than
317 2 copies per chunk).
318 The "shift" needed to prevent placing copies of the same chunks
320 on the same devices is actually a cyclic permutation with off‐
321 set 1 of each of the stripes within a complete sequence of
322 chunks.
323 The offset 1 is relative to the previous complete sequence of
324 chunks, so in case of more than 2 copies per chunk one gets the
325 following offsets:
326 1. complete sequence of chunks: offset = 0
327 2. complete sequence of chunks: offset = 1
328 3. complete sequence of chunks: offset = 2
329 :
330 n. complete sequence of chunks: offset = n-1
331 Example with 2 copies per chunk and an even number (4) of
333 devices:
334 ┌───────────┌───────────┌───────────┌───────────┐
336 │ Device #1 │ Device #2 │ Device #3 │ Device #4 │
337 ┌─────├───────────├───────────├───────────├───────────┤
338 │0x00 │ 0 │ 1 │ 2 │ 3 │ \
339 │0x01 │ 4 │ 5 │ 6 │ 7 │ > [#]
340 │... │ ... │ ... │ ... │ ... │ :
341 │ : │ : │ : │ : │ : │ :
342 │... │ ... │ ... │ ... │ ... │ :
343 │0x40 │ 252 │ 253 │ 254 │ 255 │ /
344 │0x41 │ 3 │ 0 │ 1 │ 2 │ \
345 │0x42 │ 7 │ 4 │ 5 │ 6 │ > [#]~
346 │... │ ... │ ... │ ... │ ... │ :
347 │ : │ : │ : │ : │ : │ :
348 │... │ ... │ ... │ ... │ ... │ :
349 │0x80 │ 255 │ 252 │ 253 │ 254 │ /
350 └─────└───────────└───────────└───────────└───────────┘
351 Example with 2 copies per chunk and an odd number (5) of
353 devices:
354 ┌────────┌────────┌────────┌────────┌────────┐
356 │ Dev #1 │ Dev #2 │ Dev #3 │ Dev #4 │ Dev #5 │
357 ┌─────├────────├────────├────────├────────├────────┤
358 │0x00 │ 0 │ 1 │ 2 │ 3 │ 4 │ \
359 │0x01 │ 5 │ 6 │ 7 │ 8 │ 9 │ > [#]
360 │... │ ... │ ... │ ... │ ... │ ... │ :
361 │ : │ : │ : │ : │ : │ : │ :
362 │... │ ... │ ... │ ... │ ... │ ... │ :
363 │0x40 │ 315 │ 316 │ 317 │ 318 │ 319 │ /
364 │0x41 │ 4 │ 0 │ 1 │ 2 │ 3 │ \
365 │0x42 │ 9 │ 5 │ 6 │ 7 │ 8 │ > [#]~
366 │... │ ... │ ... │ ... │ ... │ ... │ :
367 │ : │ : │ : │ : │ : │ : │ :
368 │... │ ... │ ... │ ... │ ... │ ... │ :
369 │0x80 │ 319 │ 315 │ 316 │ 317 │ 318 │ /
370 └─────└────────└────────└────────└────────└────────┘
371 With [#] being the complete sequence of chunks and [#]~ the
373 cyclic permutation with offset 1 thereof (in the case of more
374 than 2 copies per chunk there would be
375 ([#]~)~, (([#]~)~)~, ...).
376 The advantage of this layout is that MD can easily spread
378 sequential reads over the devices, making them similar to RAID0
379 in terms of speed.
380 The cost is more seeking for writes, making them substantially
381 slower.
382 "offset" Layout
385 When "offset" replicas are chosen, all the copies of a given
386 chunk are striped consecutively ("offset by the stripe length
387 after each other") over the devices.
388 Explained in detail, <number of devices> consecutive chunks are
390 striped over the devices, immediately followed by a "shifted"
391 copy of these chunks (and by further such "shifted" copies in
392 the case of more than 2 copies per chunk).
393 This pattern repeats for all further consecutive chunks of the
394 exported RAID10 device (in other words: all further data
395 blocks).
396 The "shift" needed to prevent placing copies of the same chunks
398 on the same devices is actually a cyclic permutation with off‐
399 set 1 of each of the striped copies of <number of devices> con‐
400 secutive chunks.
401 The offset 1 is relative to the previous striped copy of <number
402 of devices> consecutive chunks, so in case of more than 2 copies
403 per chunk one gets the following offsets:
404 1. <number of devices> consecutive chunks: offset = 0
405 2. <number of devices> consecutive chunks: offset = 1
406 3. <number of devices> consecutive chunks: offset = 2
407 :
408 n. <number of devices> consecutive chunks: offset = n-1
409 Example with 2 copies per chunk and an even number (4) of
411 devices:
412 ┌───────────┌───────────┌───────────┌───────────┐
414 │ Device #1 │ Device #2 │ Device #3 │ Device #4 │
415 ┌─────├───────────├───────────├───────────├───────────┤
416 │0x00 │ 0 │ 1 │ 2 │ 3 │ ) AA
417 │0x01 │ 3 │ 0 │ 1 │ 2 │ ) AA~
418 │0x02 │ 4 │ 5 │ 6 │ 7 │ ) AB
419 │0x03 │ 7 │ 4 │ 5 │ 6 │ ) AB~
420 │... │ ... │ ... │ ... │ ... │ ) ...
421 │ : │ : │ : │ : │ : │ :
422 │... │ ... │ ... │ ... │ ... │ ) ...
423 │0x79 │ 251 │ 252 │ 253 │ 254 │ ) EX
424 │0x80 │ 254 │ 251 │ 252 │ 253 │ ) EX~
425 └─────└───────────└───────────└───────────└───────────┘
426 Example with 2 copies per chunk and an odd number (5) of
428 devices:
429 ┌────────┌────────┌────────┌────────┌────────┐
431 │ Dev #1 │ Dev #2 │ Dev #3 │ Dev #4 │ Dev #5 │
432 ┌─────├────────├────────├────────├────────├────────┤
433 │0x00 │ 0 │ 1 │ 2 │ 3 │ 4 │ ) AA
434 │0x01 │ 4 │ 0 │ 1 │ 2 │ 3 │ ) AA~
435 │0x02 │ 5 │ 6 │ 7 │ 8 │ 9 │ ) AB
436 │0x03 │ 9 │ 5 │ 6 │ 7 │ 8 │ ) AB~
437 │... │ ... │ ... │ ... │ ... │ ... │ ) ...
438 │ : │ : │ : │ : │ : │ : │ :
439 │... │ ... │ ... │ ... │ ... │ ... │ ) ...
440 │0x79 │ 314 │ 315 │ 316 │ 317 │ 318 │ ) EX
441 │0x80 │ 318 │ 314 │ 315 │ 316 │ 317 │ ) EX~
442 └─────└────────└────────└────────└────────└────────┘
443 With AA, AB, ..., AZ, BA, ... being the sets of <number of
445 devices> consecutive chunks and AA~, AB~, ..., AZ~, BA~, ... the
446 cyclic permutations with offset 1 thereof (in the case of more
447 than 2 copies per chunk there would be (AA~)~, ... as well as
448 ((AA~)~)~, ... and so on).
449 This should give similar read characteristics to "far" if a
451 suitably large chunk size is used, but without as much seeking
452 for writes.
453 It should be noted that the number of devices in a RAID10 array need
455 not be a multiple of the number of replica of each data block; however,
456 there must be at least as many devices as replicas.
457 If, for example, an array is created with 5 devices and 2 replicas,
459 then space equivalent to 2.5 of the devices will be available, and
460 every block will be stored on two different devices.
461 Finally, it is possible to have an array with both "near" and "far"
463 copies. If an array is configured with 2 near copies and 2 far copies,
464 then there will be a total of 4 copies of each block, each on a differ‐
465 ent drive. This is an artifact of the implementation and is unlikely
466 to be of real value.
467 MULTIPATH
470 MULTIPATH is not really a RAID at all as there is only one real device
471 in a MULTIPATH md array. However there are multiple access points
472 (paths) to this device, and one of these paths might fail, so there are
473 some similarities.
474 A MULTIPATH array is composed of a number of logically different
476 devices, often fibre channel interfaces, that all refer the the same
477 real device. If one of these interfaces fails (e.g. due to cable prob‐
478 lems), the MULTIPATH driver will attempt to redirect requests to
479 another interface.
480 The MULTIPATH drive is not receiving any ongoing development and should
482 be considered a legacy driver. The device-mapper based multipath driv‐
483 ers should be preferred for new installations.
484 FAULTY
487 The FAULTY md module is provided for testing purposes. A FAULTY array
488 has exactly one component device and is normally assembled without a
489 superblock, so the md array created provides direct access to all of
490 the data in the component device.
491 The FAULTY module may be requested to simulate faults to allow testing
493 of other md levels or of filesystems. Faults can be chosen to trigger
494 on read requests or write requests, and can be transient (a subsequent
495 read/write at the address will probably succeed) or persistent (subse‐
496 quent read/write of the same address will fail). Further, read faults
497 can be "fixable" meaning that they persist until a write request at the
498 same address.
499 Fault types can be requested with a period. In this case, the fault
501 will recur repeatedly after the given number of requests of the rele‐
502 vant type. For example if persistent read faults have a period of 100,
503 then every 100th read request would generate a fault, and the faulty
504 sector would be recorded so that subsequent reads on that sector would
505 also fail.
506 There is a limit to the number of faulty sectors that are remembered.
508 Faults generated after this limit is exhausted are treated as tran‐
509 sient.
510 The list of faulty sectors can be flushed, and the active list of fail‐
512 ure modes can be cleared.
513 UNCLEAN SHUTDOWN
516 When changes are made to a RAID1, RAID4, RAID5, RAID6, or RAID10 array
517 there is a possibility of inconsistency for short periods of time as
518 each update requires at least two block to be written to different
519 devices, and these writes probably won't happen at exactly the same
520 time. Thus if a system with one of these arrays is shutdown in the
521 middle of a write operation (e.g. due to power failure), the array may
522 not be consistent.
523 To handle this situation, the md driver marks an array as "dirty"
525 before writing any data to it, and marks it as "clean" when the array
526 is being disabled, e.g. at shutdown. If the md driver finds an array
527 to be dirty at startup, it proceeds to correct any possibly inconsis‐
528 tency. For RAID1, this involves copying the contents of the first
529 drive onto all other drives. For RAID4, RAID5 and RAID6 this involves
530 recalculating the parity for each stripe and making sure that the par‐
531 ity block has the correct data. For RAID10 it involves copying one of
532 the replicas of each block onto all the others. This process, known as
533 "resynchronising" or "resync" is performed in the background. The
534 array can still be used, though possibly with reduced performance.
535 If a RAID4, RAID5 or RAID6 array is degraded (missing at least one
537 drive, two for RAID6) when it is restarted after an unclean shutdown,
538 it cannot recalculate parity, and so it is possible that data might be
539 undetectably corrupted. The 2.4 md driver does not alert the operator
540 to this condition. The 2.6 md driver will fail to start an array in
541 this condition without manual intervention, though this behaviour can
542 be overridden by a kernel parameter.
543 RECOVERY
546 If the md driver detects a write error on a device in a RAID1, RAID4,
547 RAID5, RAID6, or RAID10 array, it immediately disables that device
548 (marking it as faulty) and continues operation on the remaining
549 devices. If there are spare drives, the driver will start recreating
550 on one of the spare drives the data which was on that failed drive,
551 either by copying a working drive in a RAID1 configuration, or by doing
552 calculations with the parity block on RAID4, RAID5 or RAID6, or by
553 finding and copying originals for RAID10.
554 In kernels prior to about 2.6.15, a read error would cause the same
556 effect as a write error. In later kernels, a read-error will instead
557 cause md to attempt a recovery by overwriting the bad block. i.e. it
558 will find the correct data from elsewhere, write it over the block that
559 failed, and then try to read it back again. If either the write or the
560 re-read fail, md will treat the error the same way that a write error
561 is treated, and will fail the whole device.
562 While this recovery process is happening, the md driver will monitor
564 accesses to the array and will slow down the rate of recovery if other
565 activity is happening, so that normal access to the array will not be
566 unduly affected. When no other activity is happening, the recovery
567 process proceeds at full speed. The actual speed targets for the two
568 different situations can be controlled by the speed_limit_min and
569 speed_limit_max control files mentioned below.
570 SCRUBBING AND MISMATCHES
573 As storage devices can develop bad blocks at any time it is valuable to
574 regularly read all blocks on all devices in an array so as to catch
575 such bad blocks early. This process is called scrubbing.
576 md arrays can be scrubbed by writing either check or repair to the file
578 md/sync_action in the sysfs directory for the device.
579 Requesting a scrub will cause md to read every block on every device in
581 the array, and check that the data is consistent. For RAID1 and
582 RAID10, this means checking that the copies are identical. For RAID4,
583 RAID5, RAID6 this means checking that the parity block is (or blocks
584 are) correct.
585 If a read error is detected during this process, the normal read-error
587 handling causes correct data to be found from other devices and to be
588 written back to the faulty device. In many case this will effectively
589 fix the bad block.
590 If all blocks read successfully but are found to not be consistent,
592 then this is regarded as a mismatch.
593 If check was used, then no action is taken to handle the mismatch, it
595 is simply recorded. If repair was used, then a mismatch will be
596 repaired in the same way that resync repairs arrays. For RAID5/RAID6
597 new parity blocks are written. For RAID1/RAID10, all but one block are
598 overwritten with the content of that one block.
599 A count of mismatches is recorded in the sysfs file md/mismatch_cnt.
601 This is set to zero when a scrub starts and is incremented whenever a
602 sector is found that is a mismatch. md normally works in units much
603 larger than a single sector and when it finds a mismatch, it does not
604 determine exactly how many actual sectors were affected but simply adds
605 the number of sectors in the IO unit that was used. So a value of 128
606 could simply mean that a single 64KB check found an error (128 x
607 512bytes = 64KB).
608 If an array is created by mdadm with --assume-clean then a subsequent
610 check could be expected to find some mismatches.
611 On a truly clean RAID5 or RAID6 array, any mismatches should indicate a
613 hardware problem at some level - software issues should never cause
614 such a mismatch.
615 However on RAID1 and RAID10 it is possible for software issues to cause
617 a mismatch to be reported. This does not necessarily mean that the
618 data on the array is corrupted. It could simply be that the system
619 does not care what is stored on that part of the array - it is unused
620 space.
621 The most likely cause for an unexpected mismatch on RAID1 or RAID10
623 occurs if a swap partition or swap file is stored on the array.
624 When the swap subsystem wants to write a page of memory out, it flags
626 the page as 'clean' in the memory manager and requests the swap device
627 to write it out. It is quite possible that the memory will be changed
628 while the write-out is happening. In that case the 'clean' flag will
629 be found to be clear when the write completes and so the swap subsystem
630 will simply forget that the swapout had been attempted, and will possi‐
631 bly choose a different page to write out.
632 If the swap device was on RAID1 (or RAID10), then the data is sent from
634 memory to a device twice (or more depending on the number of devices in
635 the array). Thus it is possible that the memory gets changed between
636 the times it is sent, so different data can be written to the different
637 devices in the array. This will be detected by check as a mismatch.
638 However it does not reflect any corruption as the block where this mis‐
639 match occurs is being treated by the swap system as being empty, and
640 the data will never be read from that block.
641 It is conceivable for a similar situation to occur on non-swap files,
643 though it is less likely.
644 Thus the mismatch_cnt value can not be interpreted very reliably on
646 RAID1 or RAID10, especially when the device is used for swap.
647 BITMAP WRITE-INTENT LOGGING
651 From Linux 2.6.13, md supports a bitmap based write-intent log. If
652 configured, the bitmap is used to record which blocks of the array may
653 be out of sync. Before any write request is honoured, md will make
654 sure that the corresponding bit in the log is set. After a period of
655 time with no writes to an area of the array, the corresponding bit will
656 be cleared.
657 This bitmap is used for two optimisations.
659 Firstly, after an unclean shutdown, the resync process will consult the
661 bitmap and only resync those blocks that correspond to bits in the bit‐
662 map that are set. This can dramatically reduce resync time.
663 Secondly, when a drive fails and is removed from the array, md stops
665 clearing bits in the intent log. If that same drive is re-added to the
666 array, md will notice and will only recover the sections of the drive
667 that are covered by bits in the intent log that are set. This can
668 allow a device to be temporarily removed and reinserted without causing
669 an enormous recovery cost.
670 The intent log can be stored in a file on a separate device, or it can
672 be stored near the superblocks of an array which has superblocks.
673 It is possible to add an intent log to an active array, or remove an
675 intent log if one is present.
676 In 2.6.13, intent bitmaps are only supported with RAID1. Other levels
678 with redundancy are supported from 2.6.15.
679 BAD BLOCK LIST
682 From Linux 3.5 each device in an md array can store a list of known-
683 bad-blocks. This list is 4K in size and usually positioned at the end
684 of the space between the superblock and the data.
685 When a block cannot be read and cannot be repaired by writing data
687 recovered from other devices, the address of the block is stored in the
688 bad block list. Similarly if an attempt to write a block fails, the
689 address will be recorded as a bad block. If attempting to record the
690 bad block fails, the whole device will be marked faulty.
691 Attempting to read from a known bad block will cause a read error.
693 Attempting to write to a known bad block will be ignored if any write
694 errors have been reported by the device. If there have been no write
695 errors then the data will be written to the known bad block and if that
696 succeeds, the address will be removed from the list.
697 This allows an array to fail more gracefully - a few blocks on differ‐
699 ent devices can be faulty without taking the whole array out of action.
700 The list is particularly useful when recovering to a spare. If a few
702 blocks cannot be read from the other devices, the bulk of the recovery
703 can complete and those few bad blocks will be recorded in the bad block
704 list.
705 RAID456 WRITE JOURNAL
708 Due to non-atomicity nature of RAID write operations, interruption of
709 write operations (system crash, etc.) to RAID456 array can lead to
710 inconsistent parity and data loss (so called RAID-5 write hole).
711 To plug the write hole, from Linux 4.4 (to be confirmed), md supports
713 write ahead journal for RAID456. When the array is created, an addi‐
714 tional journal device can be added to the array through write-journal
715 option. The RAID write journal works similar to file system journals.
716 Before writing to the data disks, md persists data AND parity of the
717 stripe to the journal device. After crashes, md searches the journal
718 device for incomplete write operations, and replay them to the data
719 disks.
720 When the journal device fails, the RAID array is forced to run in read-
722 only mode.
723 WRITE-BEHIND
726 From Linux 2.6.14, md supports WRITE-BEHIND on RAID1 arrays.
727 This allows certain devices in the array to be flagged as write-mostly.
729 MD will only read from such devices if there is no other option.
730 If a write-intent bitmap is also provided, write requests to write-
732 mostly devices will be treated as write-behind requests and md will not
733 wait for writes to those requests to complete before reporting the
734 write as complete to the filesystem.
735 This allows for a RAID1 with WRITE-BEHIND to be used to mirror data
737 over a slow link to a remote computer (providing the link isn't too
738 slow). The extra latency of the remote link will not slow down normal
739 operations, but the remote system will still have a reasonably up-to-
740 date copy of all data.
741 FAILFAST
744 From Linux 4.10, md supports FAILFAST for RAID1 and RAID10 arrays.
745 This is a flag that can be set on individual drives, though it is usu‐
746 ally set on all drives, or no drives.
747 When md sends an I/O request to a drive that is marked as FAILFAST, and
749 when the array could survive the loss of that drive without losing
750 data, md will request that the underlying device does not perform any
751 retries. This means that a failure will be reported to md promptly,
752 and it can mark the device as faulty and continue using the other
753 device(s). md cannot control the timeout that the underlying devices
754 use to determine failure. Any changes desired to that timeout must be
755 set explictly on the underlying device, separately from using mdadm.
756 If a FAILFAST request does fail, and if it is still safe to mark the
758 device as faulty without data loss, that will be done and the array
759 will continue functioning on a reduced number of devices. If it is not
760 possible to safely mark the device as faulty, md will retry the request
761 without disabling retries in the underlying device. In any case, md
762 will not attempt to repair read errors on a device marked as FAILFAST
763 by writing out the correct. It will just mark the device as faulty.
764 FAILFAST is appropriate for storage arrays that have a low probability
766 of true failure, but will sometimes introduce unacceptable delays to
767 I/O requests while performing internal maintenance. The value of set‐
768 ting FAILFAST involves a trade-off. The gain is that the chance of
769 unacceptable delays is substantially reduced. The cost is that the
770 unlikely event of data-loss on one device is slightly more likely to
771 result in data-loss for the array.
772 When a device in an array using FAILFAST is marked as faulty, it will
774 usually become usable again in a short while. mdadm makes no attempt
775 to detect that possibility. Some separate mechanism, tuned to the spe‐
776 cific details of the expected failure modes, needs to be created to
777 monitor devices to see when they return to full functionality, and to
778 then re-add them to the array. In order of this "re-add" functionality
779 to be effective, an array using FAILFAST should always have a write-
780 intent bitmap.
781 RESTRIPING
784 Restriping, also known as Reshaping, is the processes of re-arranging
785 the data stored in each stripe into a new layout. This might involve
786 changing the number of devices in the array (so the stripes are wider),
787 changing the chunk size (so stripes are deeper or shallower), or chang‐
788 ing the arrangement of data and parity (possibly changing the RAID
789 level, e.g. 1 to 5 or 5 to 6).
790 As of Linux 2.6.35, md can reshape a RAID4, RAID5, or RAID6 array to
792 have a different number of devices (more or fewer) and to have a dif‐
793 ferent layout or chunk size. It can also convert between these differ‐
794 ent RAID levels. It can also convert between RAID0 and RAID10, and
795 between RAID0 and RAID4 or RAID5. Other possibilities may follow in
796 future kernels.
797 During any stripe process there is a 'critical section' during which
799 live data is being overwritten on disk. For the operation of increas‐
800 ing the number of drives in a RAID5, this critical section covers the
801 first few stripes (the number being the product of the old and new num‐
802 ber of devices). After this critical section is passed, data is only
803 written to areas of the array which no longer hold live data — the live
804 data has already been located away.
805 For a reshape which reduces the number of devices, the 'critical sec‐
807 tion' is at the end of the reshape process.
808 md is not able to ensure data preservation if there is a crash (e.g.
810 power failure) during the critical section. If md is asked to start an
811 array which failed during a critical section of restriping, it will
812 fail to start the array.
813 To deal with this possibility, a user-space program must
815 · Disable writes to that section of the array (using the sysfs inter‐
817 face),
818 · take a copy of the data somewhere (i.e. make a backup),
820 · allow the process to continue and invalidate the backup and restore
822 write access once the critical section is passed, and
823 · provide for restoring the critical data before restarting the array
825 after a system crash.
826 mdadm versions from 2.4 do this for growing a RAID5 array.
828 For operations that do not change the size of the array, like simply
830 increasing chunk size, or converting RAID5 to RAID6 with one extra
831 device, the entire process is the critical section. In this case, the
832 restripe will need to progress in stages, as a section is suspended,
833 backed up, restriped, and released.
834 SYSFS INTERFACE
837 Each block device appears as a directory in sysfs (which is usually
838 mounted at /sys). For MD devices, this directory will contain a subdi‐
839 rectory called md which contains various files for providing access to
840 information about the array.
841 This interface is documented more fully in the file Documenta‐
843 tion/md.txt which is distributed with the kernel sources. That file
844 should be consulted for full documentation. The following are just a
845 selection of attribute files that are available.
846 md/sync_speed_min
849 This value, if set, overrides the system-wide setting in
850 /proc/sys/dev/raid/speed_limit_min for this array only. Writing
851 the value system to this file will cause the system-wide setting
852 to have effect.
853 md/sync_speed_max
856 This is the partner of md/sync_speed_min and overrides
857 /proc/sys/dev/raid/speed_limit_max described below.
858 md/sync_action
861 This can be used to monitor and control the resync/recovery
862 process of MD. In particular, writing "check" here will cause
863 the array to read all data block and check that they are consis‐
864 tent (e.g. parity is correct, or all mirror replicas are the
865 same). Any discrepancies found are NOT corrected.
866 A count of problems found will be stored in md/mismatch_count.
868 Alternately, "repair" can be written which will cause the same
870 check to be performed, but any errors will be corrected.
871 Finally, "idle" can be written to stop the check/repair process.
873 md/stripe_cache_size
876 This is only available on RAID5 and RAID6. It records the size
877 (in pages per device) of the stripe cache which is used for
878 synchronising all write operations to the array and all read
879 operations if the array is degraded. The default is 256. Valid
880 values are 17 to 32768. Increasing this number can increase
881 performance in some situations, at some cost in system memory.
882 Note, setting this value too high can result in an "out of mem‐
883 ory" condition for the system.
884 memory_consumed = system_page_size * nr_disks *
886 stripe_cache_size
887 md/preread_bypass_threshold
890 This is only available on RAID5 and RAID6. This variable sets
891 the number of times MD will service a full-stripe-write before
892 servicing a stripe that requires some "prereading". For fair‐
893 ness this defaults to 1. Valid values are 0 to
894 stripe_cache_size. Setting this to 0 maximizes sequential-write
895 throughput at the cost of fairness to threads doing small or
896 random writes.
897 KERNEL PARAMETERS
900 The md driver recognised several different kernel parameters.
901 raid=noautodetect
903 This will disable the normal detection of md arrays that happens
904 at boot time. If a drive is partitioned with MS-DOS style par‐
905 titions, then if any of the 4 main partitions has a partition
906 type of 0xFD, then that partition will normally be inspected to
907 see if it is part of an MD array, and if any full arrays are
908 found, they are started. This kernel parameter disables this
909 behaviour.
910 raid=partitionable
913 raid=part
915 These are available in 2.6 and later kernels only. They indi‐
916 cate that autodetected MD arrays should be created as partition‐
917 able arrays, with a different major device number to the origi‐
918 nal non-partitionable md arrays. The device number is listed as
919 mdp in /proc/devices.
920 md_mod.start_ro=1
923 /sys/module/md_mod/parameters/start_ro
925 This tells md to start all arrays in read-only mode. This is a
926 soft read-only that will automatically switch to read-write on
927 the first write request. However until that write request,
928 nothing is written to any device by md, and in particular, no
929 resync or recovery operation is started.
930 md_mod.start_dirty_degraded=1
933 /sys/module/md_mod/parameters/start_dirty_degraded
935 As mentioned above, md will not normally start a RAID4, RAID5,
936 or RAID6 that is both dirty and degraded as this situation can
937 imply hidden data loss. This can be awkward if the root
938 filesystem is affected. Using this module parameter allows such
939 arrays to be started at boot time. It should be understood that
940 there is a real (though small) risk of data corruption in this
941 situation.
942 md=n,dev,dev,...
945 md=dn,dev,dev,...
947 This tells the md driver to assemble /dev/md n from the listed
948 devices. It is only necessary to start the device holding the
949 root filesystem this way. Other arrays are best started once
950 the system is booted.
951 In 2.6 kernels, the d immediately after the = indicates that a
953 partitionable device (e.g. /dev/md/d0) should be created rather
954 than the original non-partitionable device.
955 md=n,l,c,i,dev...
958 This tells the md driver to assemble a legacy RAID0 or LINEAR
959 array without a superblock. n gives the md device number, l
960 gives the level, 0 for RAID0 or -1 for LINEAR, c gives the chunk
961 size as a base-2 logarithm offset by twelve, so 0 means 4K, 1
962 means 8K. i is ignored (legacy support).
963
966 /proc/mdstat
967 Contains information about the status of currently running
968 array.
969 /proc/sys/dev/raid/speed_limit_min
971 A readable and writable file that reflects the current "goal"
972 rebuild speed for times when non-rebuild activity is current on
973 an array. The speed is in Kibibytes per second, and is a per-
974 device rate, not a per-array rate (which means that an array
975 with more disks will shuffle more data for a given speed). The
976 default is 1000.
977 /proc/sys/dev/raid/speed_limit_max
980 A readable and writable file that reflects the current "goal"
981 rebuild speed for times when no non-rebuild activity is current
982 on an array. The default is 200,000.
983
986 mdadm(8),
987 MD(4)