The md driver provides virtual devices that are created from one or
more independent underlying devices. This array of devices often con-
tains redundancy and the devices are often disk drives, hence the acro-
nym RAID which stands for a Redundant Array of Independent Disks.
md supports RAID levels 1 (mirroring), 4 (striped array with parity
device), 5 (striped array with distributed parity information), 6
(striped array with distributed dual redundancy information), and 10
(striped and mirrored). If some number of underlying devices fails
while using one of these levels, the array will continue to function;
this number is one for RAID levels 4 and 5, two for RAID level 6, and
all but one (N-1) for RAID level 1, and dependent on configuration for
md also supports a number of pseudo RAID (non-redundant) configurations
including RAID0 (striped array), LINEAR (catenated array), MULTIPATH (a
set of different interfaces to the same device), and FAULTY (a layer
over a single device into which errors can be injected).
Each device in an array may have some metadata stored in the device.
This metadata is sometimes called a superblock. The metadata records
information about the structure and state of the array. This allows
the array to be reliably re-assembled after a shutdown.
From Linux kernel version 2.6.10, md provides support for two different
formats of metadata, and other formats can be added. Prior to this
release, only one format is supported.
The common format -- known as version 0.90 -- has a superblock that is
4K long and is written into a 64K aligned block that starts at least
64K and less than 128K from the end of the device (i.e. to get the
address of the superblock round the size of the device down to a multi-
ple of 64K and then subtract 64K). The available size of each device
is the amount of space before the super block, so between 64K and 128K
is lost when a device in incorporated into an MD array. This
superblock stores multi-byte fields in a processor-dependent manner, so
arrays cannot easily be moved between computers with different proces-
The new format -- known as version 1 -- has a superblock that is nor-
mally 1K long, but can be longer. It is normally stored between 8K and
12K from the end of the device, on a 4K boundary, though variations can
be stored at the start of the device (version 1.1) or 4K from the start
of the device (version 1.2). This metadata format stores multibyte
data in a processor-independent format and supports up to hundreds of
to a RAID5), the version number is temporarily set to 0.91. This
ensures that if the reshape process is stopped in the middle (e.g. by a
system crash) and the machine boots into an older kernel that does not
support reshaping, then the array will not be assembled (which would
cause data corruption) but will be left untouched until a kernel that
can complete the reshape processes is used.
ARRAYS WITHOUT METADATA
While it is usually best to create arrays with superblocks so that they
can be assembled reliably, there are some circumstances when an array
without superblocks is preferred. These include:
Early versions of the md driver only supported Linear and Raid0
configurations and did not use a superblock (which is less crit-
ical with these configurations). While such arrays should be
rebuilt with superblocks if possible, md continues to support
FAULTY Being a largely transparent layer over a different device, the
FAULTY personality doesn't gain anything from having a
It is often possible to detect devices which are different paths
to the same storage directly rather than having a distinctive
superblock written to the device and searched for on all paths.
In this case, a MULTIPATH array with no superblock makes sense.
RAID1 In some configurations it might be desired to create a raid1
configuration that does not use a superblock, and to maintain
the state of the array elsewhere. While not encouraged for gen-
eral use, it does have special-purpose uses and is supported.
ARRAYS WITH EXTERNAL METADATA
From release 2.6.28, the md driver supports arrays with externally man-
aged metadata. That is, the metadata is not managed by the kernel but
rather by a user-space program which is external to the kernel. This
allows support for a variety of metadata formats without cluttering the
kernel with lots of details.
md is able to communicate with the user-space program through various
sysfs attributes so that it can make appropriate changes to the meta-
data - for example to mark a device as faulty. When necessary, md will
wait for the program to acknowledge the event by writing to a sysfs
attribute. The manual page for mdmon(8) contains more detail about
Many metadata formats use a single block of metadata to describe a num-
ber of different arrays which all use the same set of devices. In this
One advantage of this arrangement over the more common RAID0 arrange-
ment is that the array may be reconfigured at a later time with an
extra drive, so the array is made bigger without disturbing the data
that is on the array. This can even be done on a live array.
If a chunksize is given with a LINEAR array, the usable space on each
device is rounded down to a multiple of this chunksize.
A RAID0 array (which has zero redundancy) is also known as a striped
array. A RAID0 array is configured at creation with a Chunk Size which
must be a power of two (prior to Linux 2.6.31), and at least 4
The RAID0 driver assigns the first chunk of the array to the first
device, the second chunk to the second device, and so on until all
drives have been assigned one chunk. This collection of chunks forms a
stripe. Further chunks are gathered into stripes in the same way, and
are assigned to the remaining space in the drives.
If devices in the array are not all the same size, then once the small-
est device has been exhausted, the RAID0 driver starts collecting
chunks into smaller stripes that only span the drives which still have
A RAID1 array is also known as a mirrored set (though mirrors tend to
provide reflected images, which RAID1 does not) or a plex.
Once initialised, each device in a RAID1 array contains exactly the
same data. Changes are written to all devices in parallel. Data is
read from any one device. The driver attempts to distribute read
requests across all devices to maximise performance.
All devices in a RAID1 array should be the same size. If they are not,
then only the amount of space available on the smallest device is used
(any extra space on other devices is wasted).
Note that the read balancing done by the driver does not make the RAID1
performance profile be the same as for RAID0; a single stream of
sequential input will not be accelerated (e.g. a single dd), but multi-
ple sequential streams or a random workload will use more than one
spindle. In theory, having an N-disk RAID1 will allow N sequential
threads to read from all disks.
Individual devices in a RAID1 can be marked as "write-mostly". These
drives are excluded from the normal read balancing and will only be
read from when there is no other option. This can be useful for
devices connected over a slow link.
This allows the array to continue to function if one device fails. The
data that was on that device can be calculated as needed from the par-
ity block and the other data blocks.
RAID5 is very similar to RAID4. The difference is that the parity
blocks for each stripe, instead of being on a single device, are dis-
tributed across all devices. This allows more parallelism when writ-
ing, as two different block updates will quite possibly affect parity
blocks on different devices so there is less contention.
This also allows more parallelism when reading, as read requests are
distributed over all the devices in the array instead of all but one.
RAID6 is similar to RAID5, but can handle the loss of any two devices
without data loss. Accordingly, it requires N+2 drives to store N
drives worth of data.
The performance for RAID6 is slightly lower but comparable to RAID5 in
normal mode and single disk failure mode. It is very slow in dual disk
failure mode, however.
RAID10 provides a combination of RAID1 and RAID0, and is sometimes
known as RAID1+0. Every datablock is duplicated some number of times,
and the resulting collection of datablocks are distributed over multi-
When configuring a RAID10 array, it is necessary to specify the number
of replicas of each data block that are required (this will normally be
2) and whether the replicas should be 'near', 'offset' or 'far'. (Note
that the 'offset' layout is only available from 2.6.18).
When 'near' replicas are chosen, the multiple copies of a given chunk
are laid out consecutively across the stripes of the array, so the two
copies of a datablock will likely be at the same offset on two adjacent
When 'far' replicas are chosen, the multiple copies of a given chunk
are laid out quite distant from each other. The first copy of all data
blocks will be striped across the early part of all drives in RAID0
fashion, and then the next copy of all blocks will be striped across a
later section of all drives, always ensuring that all copies of any
given block are on different drives.
The 'far' arrangement can give sequential read performance equal to
that of a RAID0 array, but at the cost of reduced write performance.
When 'offset' replicas are chosen, the multiple copies of a given chunk
every block will be stored on two different devices.
Finally, it is possible to have an array with both 'near' and 'far'
copies. If an array is configured with 2 near copies and 2 far copies,
then there will be a total of 4 copies of each block, each on a differ-
ent drive. This is an artifact of the implementation and is unlikely
to be of real value.
MULTIPATH is not really a RAID at all as there is only one real device
in a MULTIPATH md array. However there are multiple access points
(paths) to this device, and one of these paths might fail, so there are
A MULTIPATH array is composed of a number of logically different
devices, often fibre channel interfaces, that all refer the the same
real device. If one of these interfaces fails (e.g. due to cable prob-
lems), the multipath driver will attempt to redirect requests to
The MULTIPATH drive is not receiving any ongoing development and should
be considered a legacy driver. The device-mapper based multipath driv-
ers should be preferred for new installations.
The FAULTY md module is provided for testing purposes. A faulty array
has exactly one component device and is normally assembled without a
superblock, so the md array created provides direct access to all of
the data in the component device.
The FAULTY module may be requested to simulate faults to allow testing
of other md levels or of filesystems. Faults can be chosen to trigger
on read requests or write requests, and can be transient (a subsequent
read/write at the address will probably succeed) or persistent (subse-
quent read/write of the same address will fail). Further, read faults
can be "fixable" meaning that they persist until a write request at the
Fault types can be requested with a period. In this case, the fault
will recur repeatedly after the given number of requests of the rele-
vant type. For example if persistent read faults have a period of 100,
then every 100th read request would generate a fault, and the faulty
sector would be recorded so that subsequent reads on that sector would
There is a limit to the number of faulty sectors that are remembered.
Faults generated after this limit is exhausted are treated as tran-
The list of faulty sectors can be flushed, and the active list of fail-
ure modes can be cleared.
is being disabled, e.g. at shutdown. If the md driver finds an array
to be dirty at startup, it proceeds to correct any possibly inconsis-
tency. For RAID1, this involves copying the contents of the first
drive onto all other drives. For RAID4, RAID5 and RAID6 this involves
recalculating the parity for each stripe and making sure that the par-
ity block has the correct data. For RAID10 it involves copying one of
the replicas of each block onto all the others. This process, known as
"resynchronising" or "resync" is performed in the background. The
array can still be used, though possibly with reduced performance.
If a RAID4, RAID5 or RAID6 array is degraded (missing at least one
drive, two for RAID6) when it is restarted after an unclean shutdown,
it cannot recalculate parity, and so it is possible that data might be
undetectably corrupted. The 2.4 md driver does not alert the operator
to this condition. The 2.6 md driver will fail to start an array in
this condition without manual intervention, though this behaviour can
be overridden by a kernel parameter.
If the md driver detects a write error on a device in a RAID1, RAID4,
RAID5, RAID6, or RAID10 array, it immediately disables that device
(marking it as faulty) and continues operation on the remaining
devices. If there are spare drives, the driver will start recreating
on one of the spare drives the data which was on that failed drive,
either by copying a working drive in a RAID1 configuration, or by doing
calculations with the parity block on RAID4, RAID5 or RAID6, or by
finding and copying originals for RAID10.
In kernels prior to about 2.6.15, a read error would cause the same
effect as a write error. In later kernels, a read-error will instead
cause md to attempt a recovery by overwriting the bad block. i.e. it
will find the correct data from elsewhere, write it over the block that
failed, and then try to read it back again. If either the write or the
re-read fail, md will treat the error the same way that a write error
is treated, and will fail the whole device.
While this recovery process is happening, the md driver will monitor
accesses to the array and will slow down the rate of recovery if other
activity is happening, so that normal access to the array will not be
unduly affected. When no other activity is happening, the recovery
process proceeds at full speed. The actual speed targets for the two
different situations can be controlled by the speed_limit_min and
speed_limit_max control files mentioned below.
SCRUBBING AND MISMATCHES
As storage devices can develop bad blocks at any time it is valuable to
regularly read all blocks on all devices in an array so as to catch
such bad blocks early. This process is called scrubbing.
md arrays can be scrubbed by writing either check or repair to the file
md/sync_action in the sysfs directory for the device.
then this is regarded as a mismatch.
If check was used, then no action is taken to handle the mismatch, it
is simply recorded. If repair was used, then a mismatch will be
repaired in the same way that resync repairs arrays. For RAID5/RAID6
new parity blocks are written. For RAID1/RAID10, all but one block are
overwritten with the content of that one block.
A count of mismatches is recorded in the sysfs file md/mismatch_cnt.
This is set to zero when a scrub starts and is incremented whenever a
sector is found that is a mismatch. md normally works in units much
larger than a single sector and when it finds a mismatch, it does not
determine exactly how many actual sectors were affected but simply adds
the number of sectors in the IO unit that was used. So a value of 128
could simply mean that a single 64KB check found an error (128 x
512bytes = 64KB).
If an array is created by mdadm with --assume-clean then a subsequent
check could be expected to find some mismatches.
On a truly clean RAID5 or RAID6 array, any mismatches should indicate a
hardware problem at some level - software issues should never cause
such a mismatch.
However on RAID1 and RAID10 it is possible for software issues to cause
a mismatch to be reported. This does not necessarily mean that the
data on the array is corrupted. It could simply be that the system
does not care what is stored on that part of the array - it is unused
The most likely cause for an unexpected mismatch on RAID1 or RAID10
occurs if a swap partition or swap file is stored on the array.
When the swap subsystem wants to write a page of memory out, it flags
the page as 'clean' in the memory manager and requests the swap device
to write it out. It is quite possible that the memory will be changed
while the write-out is happening. In that case the 'clean' flag will
be found to be clear when the write completes and so the swap subsystem
will simply forget that the swapout had been attempted, and will possi-
bly choose a different page to write out.
If the swap device was on RAID1 (or RAID10), then the data is sent from
memory to a device twice (or more depending on the number of devices in
the array). Thus it is possible that the memory gets changed between
the times it is sent, so different data can be written to the different
devices in the array. This will be detected by check as a mismatch.
However it does not reflect any corruption as the block where this mis-
match occurs is being treated by the swap system as being empty, and
the data will never be read from that block.
It is conceivable for a similar situation to occur on non-swap files,
though it is less likely.
Thus the mismatch_cnt value can not be interpreted very reliably on
This bitmap is used for two optimisations.
Firstly, after an unclean shutdown, the resync process will consult the
bitmap and only resync those blocks that correspond to bits in the bit-
map that are set. This can dramatically reduce resync time.
Secondly, when a drive fails and is removed from the array, md stops
clearing bits in the intent log. If that same drive is re-added to the
array, md will notice and will only recover the sections of the drive
that are covered by bits in the intent log that are set. This can
allow a device to be temporarily removed and reinserted without causing
an enormous recovery cost.
The intent log can be stored in a file on a separate device, or it can
be stored near the superblocks of an array which has superblocks.
It is possible to add an intent log to an active array, or remove an
intent log if one is present.
In 2.6.13, intent bitmaps are only supported with RAID1. Other levels
with redundancy are supported from 2.6.15.
From Linux 2.6.14, md supports WRITE-BEHIND on RAID1 arrays.
This allows certain devices in the array to be flagged as write-mostly.
MD will only read from such devices if there is no other option.
If a write-intent bitmap is also provided, write requests to write-
mostly devices will be treated as write-behind requests and md will not
wait for writes to those requests to complete before reporting the
write as complete to the filesystem.
This allows for a RAID1 with WRITE-BEHIND to be used to mirror data
over a slow link to a remote computer (providing the link isn't too
slow). The extra latency of the remote link will not slow down normal
operations, but the remote system will still have a reasonably up-to-
date copy of all data.
Restriping, also known as Reshaping, is the processes of re-arranging
the data stored in each stripe into a new layout. This might involve
changing the number of devices in the array (so the stripes are wider),
changing the chunk size (so stripes are deeper or shallower), or chang-
ing the arrangement of data and parity (possibly changing the raid
level, e.g. 1 to 5 or 5 to 6).
As of Linux 2.6.35, md can reshape a RAID4, RAID5, or RAID6 array to
have a different number of devices (more or fewer) and to have a dif-
ferent layout or chunk size. It can also convert between these differ-
ent RAID levels. It can also convert between RAID0 and RAID10, and
between RAID0 and RAID4 or RAID5. Other possibilities may follow in
md is not able to ensure data preservation if there is a crash (e.g.
power failure) during the critical section. If md is asked to start an
array which failed during a critical section of restriping, it will
fail to start the array.
To deal with this possibility, a user-space program must
o Disable writes to that section of the array (using the sysfs inter-
o take a copy of the data somewhere (i.e. make a backup),
o allow the process to continue and invalidate the backup and restore
write access once the critical section is passed, and
o provide for restoring the critical data before restarting the array
after a system crash.
mdadm versions from 2.4 do this for growing a RAID5 array.
For operations that do not change the size of the array, like simply
increasing chunk size, or converting RAID5 to RAID6 with one extra
device, the entire process is the critical section. In this case, the
restripe will need to progress in stages, as a section is suspended,
backed up, restriped, and released.
Each block device appears as a directory in sysfs (which is usually
mounted at /sys). For MD devices, this directory will contain a subdi-
rectory called md which contains various files for providing access to
information about the array.
This interface is documented more fully in the file Documenta-
tion/md.txt which is distributed with the kernel sources. That file
should be consulted for full documentation. The following are just a
selection of attribute files that are available.
This value, if set, overrides the system-wide setting in
/proc/sys/dev/raid/speed_limit_min for this array only. Writing
the value system to this file will cause the system-wide setting
to have effect.
This is the partner of md/sync_speed_min and overrides
/proc/sys/dev/raid/speed_limit_max described below.
This can be used to monitor and control the resync/recovery
This is only available on RAID5 and RAID6. It records the size
(in pages per device) of the stripe cache which is used for
synchronising all write operations to the array and all read
operations if the array is degraded. The default is 256. Valid
values are 17 to 32768. Increasing this number can increase
performance in some situations, at some cost in system memory.
Note, setting this value too high can result in an "out of mem-
ory" condition for the system.
memory_consumed = system_page_size * nr_disks *
This is only available on RAID5 and RAID6. This variable sets
the number of times MD will service a full-stripe-write before
servicing a stripe that requires some "prereading". For fair-
ness this defaults to 1. Valid values are 0 to
stripe_cache_size. Setting this to 0 maximizes sequential-write
throughput at the cost of fairness to threads doing small or
The md driver recognised several different kernel parameters.
This will disable the normal detection of md arrays that happens
at boot time. If a drive is partitioned with MS-DOS style par-
titions, then if any of the 4 main partitions has a partition
type of 0xFD, then that partition will normally be inspected to
see if it is part of an MD array, and if any full arrays are
found, they are started. This kernel parameter disables this
These are available in 2.6 and later kernels only. They indi-
cate that autodetected MD arrays should be created as partition-
able arrays, with a different major device number to the origi-
nal non-partitionable md arrays. The device number is listed as
mdp in /proc/devices.
This tells md to start all arrays in read-only mode. This is a
soft read-only that will automatically switch to read-write on
the first write request. However until that write request,
there is a real (though small) risk of data corruption in this
This tells the md driver to assemble /dev/md n from the listed
devices. It is only necessary to start the device holding the
root filesystem this way. Other arrays are best started once
the system is booted.
In 2.6 kernels, the d immediately after the = indicates that a
partitionable device (e.g. /dev/md/d0) should be created rather
than the original non-partitionable device.
This tells the md driver to assemble a legacy RAID0 or LINEAR
array without a superblock. n gives the md device number, l
gives the level, 0 for RAID0 or -1 for LINEAR, c gives the chunk
size as a base-2 logarithm offset by twelve, so 0 means 4K, 1
means 8K. i is ignored (legacy support).
Contains information about the status of currently running
A readable and writable file that reflects the current "goal"
rebuild speed for times when non-rebuild activity is current on
an array. The speed is in Kibibytes per second, and is a per-
device rate, not a per-array rate (which means that an array
with more disks will shuffle more data for a given speed). The
default is 1000.
A readable and writable file that reflects the current "goal"
rebuild speed for times when no non-rebuild activity is current
on an array. The default is 200,000.
Man Pages Copyright Respective Owners. Site Copyright (C) 1994 - 2017
All Rights Reserved.