This document descibes a collection of device-mapper targets that
between them implement thin-provisioning and snapshots.
The main highlight of this implementation, compared to the previous
implementation of snapshots, is that it allows many virtual devices to
be stored on the same data volume. This simplifies administration and
allows the sharing of data between volumes, thus reducing disk usage.
Another significant feature is support for an arbitrary depth of
recursive snapshots (snapshots of snapshots of snapshots ...). The
previous implementation of snapshots did this by chaining together
lookup tables, and so performance was O(depth). This new
implementation uses a single data structure to avoid this degradation
with depth. Fragmentation may still be an issue, however, in some
Metadata is stored on a separate device from data, giving the
administrator some freedom, for example to:
- Improve metadata resilience by storing metadata on a mirrored volume
but data on a non-mirrored one.
- Improve performance by storing the metadata on SSD.
These targets are very much still in the EXPERIMENTAL state. Please
do not yet rely on them in production. But do experiment and offer us
feedback. Different use cases will have different performance
characteristics, for example due to fragmentation of the data volume.
If you find this software is not performing as expected please mail
firstname.lastname@example.org with details and we'll try our best to improve
things for you.
Userspace tools for checking and repairing the metadata are under
This section describes some quick recipes for using thin provisioning.
They use the dmsetup program to control the device-mapper driver
directly. End users will be advised to use a higher-level volume
manager such as LVM2 once support has been added.
The pool device ties together the metadata volume and the data volume.
It maps I/O linearly to the data volume and updates the metadata via
- Function calls from the thin targets
- Device-mapper 'messages' from userspace which control the creation of new
virtual devices amongst other things.
Setting up a fresh pool device
Setting up a pool device requires a valid metadata device, and a
data device. If you do not have an existing metadata device you can
make one by zeroing the first 4k to indicate empty metadata.
dd if=/dev/zero of=$metadata_dev bs=4096 count=1
The amount of metadata you need will vary according to how many blocks
are shared between thin devices (i.e. through snapshots). If you have
less sharing than average you'll need a larger-than-average metadata device.
As a guide, we suggest you calculate the number of bytes to use in the
metadata device as 48 * $data_dev_size / $data_block_size but round it up
to 2MB if the answer is smaller. The largest size supported is 16GB.
If you're creating large numbers of snapshots which are recording large
amounts of change, you may need find you need to increase this.
Reloading a pool table
You may reload a pool's table, indeed this is how the pool is resized
if it runs out of space. (N.B. While specifying a different metadata
device when reloading is not forbidden at the moment, things will go
wrong if it does not route I/O to exactly the same on-disk location as
Using an existing pool device
dmsetup create pool \
--table "0 20971520 thin-pool $metadata_dev $data_dev \
$data_block_size gives the smallest unit of disk space that can be
allocated at a time expressed in units of 512-byte sectors. People
primarily interested in thin provisioning may want to use a value such
as 1024 (512KB). People doing lots of snapshotting may want a smaller value
such as 128 (64KB). If you are not zeroing newly-allocated data,
a larger $data_block_size in the region of 256000 (128MB) is suggested.
$data_block_size must be the same for the lifetime of the
$low_water_mark is expressed in blocks of size $data_block_size. If
free space on the data device drops below this level then a dm event
will be triggered which a userspace daemon should catch allowing it to
extend the pool device. Only one such event will be sent.
Resuming a device with a new table itself triggers an event so the
userspace daemon can use this to detect a situation where a new table
already exceeds the threshold.
i) Creating a new thinly-provisioned volume.
To create a new thinly- provisioned volume you must send a message to an
active pool device, /dev/mapper/pool in this example.
dmsetup message /dev/mapper/pool 0 "create_thin 0"
Here '0' is an identifier for the volume, a 24-bit number. It's up
to the caller to allocate and manage these identifiers. If the
identifier is already in use, the message will fail with -EEXIST.
ii) Using a thinly-provisioned volume.
Thinly-provisioned volumes are activated using the 'thin' target:
dmsetup create thin --table "0 2097152 thin /dev/mapper/pool 0"
The last parameter is the identifier for the thinp device.
i) Creating an internal snapshot.
Snapshots are created with another message to the pool.
N.B. If the origin device that you wish to snapshot is active, you
must suspend it before creating the snapshot to avoid corruption.
This is NOT enforced at the moment, so please be careful!
dmsetup suspend /dev/mapper/thin
dmsetup message /dev/mapper/pool 0 "create_snap 1 0"
dmsetup resume /dev/mapper/thin
Here '1' is the identifier for the volume, a 24-bit number. '0' is the
identifier for the origin device.
ii) Using an internal snapshot.
Once created, the user doesn't have to worry about any connection
between the origin and the snapshot. Indeed the snapshot is no
different from any other thinly-provisioned device and can be
snapshotted itself via the same method. It's perfectly legal to
have only one of them active, and there's no ordering requirement on
activating or removing them both. (This differs from conventional
Activate it exactly the same way as any other thinly-provisioned volume:
dmsetup create snap --table "0 2097152 thin /dev/mapper/pool 1"
All devices using a pool must be deactivated before the pool itself
dmsetup remove thin
dmsetup remove snap
dmsetup remove pool
thin-pool <metadata dev> <data dev> <data block size (sectors)> \
<low water mark (blocks)> [<number of feature args> [<arg>]*]
Optional feature arguments:
- 'skip_block_zeroing': skips the zeroing of newly-provisioned blocks.
Data block size must be between 64KB (128 sectors) and 1GB
(2097152 sectors) inclusive.
<transaction id> <used metadata blocks>/<total metadata blocks>
<used data blocks>/<total data blocks> <held metadata root>
A 64-bit number used by userspace to help synchronise with metadata
from volume managers.
used data blocks / total data blocks
If the number of free blocks drops below the pool's low water mark a
dm event will be sent to userspace. This event is edge-triggered and
it will occur only once after each resume so volume manager writers
should register for the event and then check the target's status.
held metadata root:
The location, in sectors, of the metadata root that has been
'held' for userspace read access. '-' indicates there is no
held root. This feature is not yet implemented so '-' is
create_thin <dev id>
Create a new thinly-provisioned device.
<dev id> is an arbitrary unique 24-bit identifier chosen by
create_snap <dev id> <origin id>
Create a new snapshot of another thinly-provisioned device.
<dev id> is an arbitrary unique 24-bit identifier chosen by
<origin id> is the identifier of the thinly-provisioned device
of which the new device will be a snapshot.
delete <dev id>
Deletes a thin device. Irreversible.
trim <dev id> <new size in sectors>
Delete mappings from the end of a thin device. Irreversible.
You might want to use this if you're reducing the size of
your thinly-provisioned device. In many cases, due to the
sharing of blocks between devices, it is not possible to
determine in advance how much space 'trim' will release. (In
future a userspace tool might be able to perform this
set_transaction_id <current id> <new id>
Userland volume managers, such as LVM, need a way to
synchronise their external metadata with the internal metadata of the
pool target. The thin-pool target offers to store an
arbitrary 64-bit transaction id and return it on the target's
status line. To avoid races you must provide what you think
the current transaction id is when you change it with this
thin <pool dev> <dev id>
the thin-pool device, e.g. /dev/mapper/my_pool or 253:0
the internal device identifier of the device to be
The pool doesn't store any size against the thin devices. If you
load a thin target that is smaller than you've been using previously,
then you'll have no access to blocks mapped beyond the end. If you
load a target that is bigger than before, then extra blocks will be
provisioned as and when needed.
If you wish to reduce the size of your thin device and potentially
regain some space then send the 'trim' message to the pool.
<nr mapped sectors> <highest mapped sector>
to post comments)