btrfs-balance(8)



BTRFS-BALANCE(8)                 Btrfs Manual                 BTRFS-BALANCE(8)

NAME
       btrfs-balance - balance block groups on a btrfs filesystem

SYNOPSIS
       btrfs balance <subcommand> <args>

DESCRIPTION
       The primary purpose of the balance feature is to spread block groups
       across all devices so they match constraints defined by the respective
       profiles. See mkfs.btrfs(8) section PROFILES for more details. The
       scope of the balancing process can be further tuned by use of filters
       that can select the block groups to process. Balance works only on a
       mounted filesystem. Extent sharing is preserved and reflinks are not
       broken. Files are not defragmented nor recompressed, file extents are
       preserved but the physical location on devices will change.

       The balance operation is cancellable by the user. The on-disk state of
       the filesystem is always consistent so an unexpected interruption (eg.
       system crash, reboot) does not corrupt the filesystem. The progress of
       the balance operation is temporarily stored as an internal state and
       will be resumed upon mount, unless the mount option skip_balance is
       specified.

           Warning
           running balance without filters will take a lot of time as it
           basically move data/metadata from the whol filesystem and needs to
           update all block pointers.

       The filters can be used to perform following actions:

       o   convert block group profiles (filter convert)

       o   make block group usage more compact (filter usage)

       o   perform actions only on a given device (filters devid, drange)

       The filters can be applied to a combination of block group types (data,
       metadata, system). Note that changing only the system type needs the
       force option. Otherwise system gets automatically converted whenver
       metadata profile is converted.

       When metadata redundancy is reduced (eg. from RAID1 to single) the
       force option is also required and it is noted in system log.

           Note
           the balance operation needs enough work space, ie. space that is
           completely unused in the filesystem, otherwise this may lead to
           ENOSPC reports. See the section ENOSPC for more details.

COMPATIBILITY
           Note
           The balance subcommand also exists under the btrfs filesystem
           namespace. This still works for backward compatibility but is
           deprecated and should not be used any more.

           Note
           A short syntax btrfs balance <path> works due to backward
           compatibility but is deprecated and should not be used any more.
           Use btrfs balance start command instead.

PERFORMANCE IMPLICATIONS
       Balancing operations are very IO intensive and can also be quite CPU
       intensive, impacting other ongoing filesystem operations. Typically
       large amounts of data are copied from one location to another, with
       corresponding metadata updates.

       Depending upon the block group layout, it can also be seek heavy.
       Performance on rotational devices is noticeably worse compared to SSDs
       or fast arrays.

SUBCOMMAND
       cancel <path>
           cancels a running or paused balance, the command will block and
           wait until the current blockgroup being processed completes

           Since kernel 5.7 the response time of the cancellation is
           significantly improved, on older kernels it might take a long time
           until currently processed chunk is completely finished.

       pause <path>
           pause running balance operation, this will store the state of the
           balance progress and used filters to the filesystem

       resume <path>
           resume interrupted balance, the balance status must be stored on
           the filesystem from previous run, eg. after it was paused or
           forcibly interrupted and mounted again with skip_balance

       start [options] <path>
           start the balance operation according to the specified filters,
           without any filters the data and metadata from the whole filesystem
           are moved. The process runs in the foreground.

               Note
               the balance command without filters will basically move
               everything in the filesystem to a new physical location on
               devices (ie. it does not affect the logical properties of file
               extents like offsets within files and extent sharing). The run
               time is potentially very long, depending on the filesystem
               size. To prevent starting a full balance by accident, the user
               is warned and has a few seconds to cancel the operation before
               it starts. The warning and delay can be skipped with
               --full-balance option.
           Please note that the filters must be written together with the -d,
           -m and -s options, because they're optional and bare -d and -m also
           work and mean no filters.

           Options

           -d[<filters>]
               act on data block groups, see FILTERS section for details about
               filters

           -m[<filters>]
               act on metadata chunks, see FILTERS section for details about
               filters

           -s[<filters>]
               act on system chunks (requires -f), see FILTERS section for
               details about filters.

           -f
               force a reduction of metadata integrity, eg. when going from
               raid1 to single

           --background|--bg
               run the balance operation asynchronously in the background,
               uses fork(2) to start the process that calls the kernel ioctl

           -v
               (deprecated) alias for global -v option

       status [-v] <path>
           Show status of running or paused balance.

           Options

           -v
               (deprecated) alias for global -v option

FILTERS
       From kernel 3.3 onwards, btrfs balance can limit its action to a subset
       of the whole filesystem, and can be used to change the replication
       configuration (e.g. moving data from single to RAID1). This
       functionality is accessed through the -d, -m or -s options to btrfs
       balance start, which filter on data, metadata and system blocks
       respectively.

       A filter has the following structure: type[=params][,type=...]

       The available types are:

       profiles=<profiles>
           Balances only block groups with the given profiles. Parameters are
           a list of profile names separated by "|" (pipe).

       usage=<percent>, usage=<range>
           Balances only block groups with usage under the given percentage.
           The value of 0 is allowed and will clean up completely unused block
           groups, this should not require any new work space allocated. You
           may want to use usage=0 in case balance is returning ENOSPC and
           your filesystem is not too full.

           The argument may be a single value or a range. The single value N
           means at most N percent used, equivalent to ..N range syntax.
           Kernels prior to 4.4 accept only the single value format. The
           minimum range boundary is inclusive, maximum is exclusive.

       devid=<id>
           Balances only block groups which have at least one chunk on the
           given device. To list devices with ids use btrfs filesystem show.

       drange=<range>
           Balance only block groups which overlap with the given byte range
           on any device. Use in conjunction with devid to filter on a
           specific device. The parameter is a range specified as start..end.

       vrange=<range>
           Balance only block groups which overlap with the given byte range
           in the filesystem's internal virtual address space. This is the
           address space that most reports from btrfs in the kernel log use.
           The parameter is a range specified as start..end.

       convert=<profile>
           Convert each selected block group to the given profile name
           identified by parameters.

               Note
               starting with kernel 4.5, the data chunks can be converted
               to/from the DUP profile on a single device.

               Note
               starting with kernel 4.6, all profiles can be converted to/from
               DUP on multi-device filesystems.

       limit=<number>, limit=<range>
           Process only given number of chunks, after all filters are applied.
           This can be used to specifically target a chunk in connection with
           other filters (drange, vrange) or just simply limit the amount of
           work done by a single balance run.

           The argument may be a single value or a range. The single value N
           means at most N chunks, equivalent to ..N range syntax. Kernels
           prior to 4.4 accept only the single value format. The range minimum
           and maximum are inclusive.

       stripes=<range>
           Balance only block groups which have the given number of stripes.
           The parameter is a range specified as start..end. Makes sense for
           block group profiles that utilize striping, ie. RAID0/10/5/6. The
           range minimum and maximum are inclusive.

       soft
           Takes no parameters. Only has meaning when converting between
           profiles. When doing convert from one profile to another and soft
           mode is on, chunks that already have the target profile are left
           untouched. This is useful e.g. when half of the filesystem was
           converted earlier but got cancelled.

           The soft mode switch is (like every other filter) per-type. For
           example, this means that we can convert metadata chunks the "hard"
           way while converting data chunks selectively with soft switch.

       Profile names, used in profiles and convert are one of: raid0, raid1,
       raid10, raid5, raid6, dup, single. The mixed data/metadata profiles can
       be converted in the same way, but it's conversion between mixed and
       non-mixed is not implemented. For the constraints of the profiles
       please refer to mkfs.btrfs(8), section PROFILES.

ENOSPC
       The way balance operates, it usually needs to temporarily create a new
       block group and move the old data there, before the old block group can
       be removed. For that it needs the work space, otherwise it fails for
       ENOSPC reasons. This is not the same ENOSPC as if the free space is
       exhausted. This refers to the space on the level of block groups, which
       are bigger parts of the filesystem that contain many file extents.

       The free work space can be calculated from the output of the btrfs
       filesystem show command:

              Label: 'BTRFS'  uuid: 8a9d72cd-ead3-469d-b371-9c7203276265
                      Total devices 2 FS bytes used 77.03GiB
                      devid    1 size 53.90GiB used 51.90GiB path /dev/sdc2
                      devid    2 size 53.90GiB used 51.90GiB path /dev/sde1

       size - used = free work space 53.90GiB - 51.90GiB = 2.00GiB

       An example of a filter that does not require workspace is usage=0. This
       will scan through all unused block groups of a given type and will
       reclaim the space. After that it might be possible to run other
       filters.

       CONVERSIONS ON MULTIPLE DEVICES

       Conversion to profiles based on striping (RAID0, RAID5/6) require the
       work space on each device. An interrupted balance may leave partially
       filled block groups that consume the work space.

EXAMPLES
       A more comprehensive example when going from one to multiple devices,
       and back, can be found in section TYPICAL USECASES of btrfs-device(8).

   MAKING BLOCK GROUP LAYOUT MORE COMPACT
       The layout of block groups is not normally visible; most tools report
       only summarized numbers of free or used space, but there are still some
       hints provided.

       Let's use the following real life example and start with the output:

           $ btrfs filesystem df /path
           Data, single: total=75.81GiB, used=64.44GiB
           System, RAID1: total=32.00MiB, used=20.00KiB
           Metadata, RAID1: total=15.87GiB, used=8.84GiB
           GlobalReserve, single: total=512.00MiB, used=0.00B

       Roughly calculating for data, 75G - 64G = 11G, the used/total ratio is
       about 85%. How can we can interpret that:

       o   chunks are filled by 85% on average, ie. the usage filter with
           anything smaller than 85 will likely not affect anything

       o   in a more realistic scenario, the space is distributed unevenly, we
           can assume there are completely used chunks and the remaining are
           partially filled

       Compacting the layout could be used on both. In the former case it
       would spread data of a given chunk to the others and removing it. Here
       we can estimate that roughly 850 MiB of data have to be moved (85% of a
       1 GiB chunk).

       In the latter case, targeting the partially used chunks will have to
       move less data and thus will be faster. A typical filter command would
       look like:

           # btrfs balance start -dusage=50 /path
           Done, had to relocate 2 out of 97 chunks

           $ btrfs filesystem df /path
           Data, single: total=74.03GiB, used=64.43GiB
           System, RAID1: total=32.00MiB, used=20.00KiB
           Metadata, RAID1: total=15.87GiB, used=8.84GiB
           GlobalReserve, single: total=512.00MiB, used=0.00B

       As you can see, the total amount of data is decreased by just 1 GiB,
       which is an expected result. Let's see what will happen when we
       increase the estimated usage filter.

           # btrfs balance start -dusage=85 /path
           Done, had to relocate 13 out of 95 chunks

           $ btrfs filesystem df /path
           Data, single: total=68.03GiB, used=64.43GiB
           System, RAID1: total=32.00MiB, used=20.00KiB
           Metadata, RAID1: total=15.87GiB, used=8.85GiB
           GlobalReserve, single: total=512.00MiB, used=0.00B

       Now the used/total ratio is about 94% and we moved about 74G - 68G = 6G
       of data to the remaining blockgroups, ie. the 6GiB are now free of
       filesystem structures, and can be reused for new data or metadata block
       groups.

       We can do a similar exercise with the metadata block groups, but this
       should not typically be necessary, unless the used/total ratio is
       really off. Here the ratio is roughly 50% but the difference as an
       absolute number is "a few gigabytes", which can be considered normal
       for a workload with snapshots or reflinks updated frequently.

           # btrfs balance start -musage=50 /path
           Done, had to relocate 4 out of 89 chunks

           $ btrfs filesystem df /path
           Data, single: total=68.03GiB, used=64.43GiB
           System, RAID1: total=32.00MiB, used=20.00KiB
           Metadata, RAID1: total=14.87GiB, used=8.85GiB
           GlobalReserve, single: total=512.00MiB, used=0.00B

       Just 1 GiB decrease, which possibly means there are block groups with
       good utilization. Making the metadata layout more compact would in turn
       require updating more metadata structures, ie. lots of IO. As running
       out of metadata space is a more severe problem, it's not necessary to
       keep the utilization ratio too high. For the purpose of this example,
       let's see the effects of further compaction:

           # btrfs balance start -musage=70 /path
           Done, had to relocate 13 out of 88 chunks

           $ btrfs filesystem df .
           Data, single: total=68.03GiB, used=64.43GiB
           System, RAID1: total=32.00MiB, used=20.00KiB
           Metadata, RAID1: total=11.97GiB, used=8.83GiB
           GlobalReserve, single: total=512.00MiB, used=0.00B

   GETTING RID OF COMPLETELY UNUSED BLOCK GROUPS
       Normally the balance operation needs a work space, to temporarily move
       the data before the old block groups gets removed. If there's no work
       space, it ends with no space left.

       There's a special case when the block groups are completely unused,
       possibly left after removing lots of files or deleting snapshots.
       Removing empty block groups is automatic since 3.18. The same can be
       achieved manually with a notable exception that this operation does not
       require the work space. Thus it can be used to reclaim unused block
       groups to make it available.

           # btrfs balance start -dusage=0 /path

       This should lead to decrease in the total numbers in the btrfs
       filesystem df output.

EXIT STATUS
       Unless indicated otherwise below, all btrfs balance subcommands return
       a zero exit status if they succeed, and non zero in case of failure.

       The pause, cancel, and resume subcommands exit with a status of 2 if
       they fail because a balance operation was not running.

       The status subcommand exits with a status of 0 if a balance operation
       is not running, 1 if the command-line usage is incorrect or a balance
       operation is still running, and 2 on other errors.

AVAILABILITY
       btrfs is part of btrfs-progs. Please refer to the btrfs wiki
       http://btrfs.wiki.kernel.org for further details.

SEE ALSO
       mkfs.btrfs(8), btrfs-device(8)

Btrfs v5.7                        07/02/2020                  BTRFS-BALANCE(8)

Man(1) output converted with man2html
list of all man pages