beta version of LVM snapshot merging
Mikulas Patocka wrote:
I'd like announce beta version of snapshot merging. The patches can be
downloaded at http://people.redhat.com/mpatocka/patches/ The patches are
against kernel 2.6.26, device-mapper.1.02.27 and LVM2.2.02.39.
The snapshot merging allows you to merge snapshot content back into the
original device. The most useful use for this feature is the possibility
to rollback state of the whole computer after failed package upgrade,
administrator's error or so.
Merging of a snapshot is initiated with command "lvconvert -M
vg/lv_snapshot". lvconvert polls for the termination of merging and then
automatically removes the merging snapshot. If the computer crashes, the
merging resumes after a crash and background polling is restarted too
(from lvchange -ay or vgchange -ay command).
While the merging is active, any accesses to the origin device are
dicrected the snapshot that is being merged. When the merging finishes,
the origin target is seamlessly reloaded and the merging snapshot is
dropped. The filesystem can stay mounted during this time.
There are three types of merging:
--nameorigin (default) - the resulting logical volume will have name,
minor number and UUID of the original origin volume. When this mode is
used, neither snapshot nor the origin can be mounted when the merging
starts. (but you can mount the origin immediatelly after you start
merging, you don't have to wait until it finishes)
--namesnapshot - the resulting logical volume will have name, minor number
and UUID of the snapshot to be merged. When the merging starts, the
snapshot can be mounted and the origin cannot be mounted.
--onactivate - marks the snapshot for merge on next activate (when
lvchange -ay or vgchange -ay changes the state from inactiva to active),
but doesn't do actual merging. This option is useful if you need to merge
over a filesystem that cannot be unmounted (for example root) --- use
lvconvert -M --onactivate and reboot the computer. On next reboot, the
merge will start and the system will run with root filesystem
corresponding to the snapshot that is being merged.
You can have multiple snapshots before you start merging, any existing
snapshots are maintained stable (i.e. new exceptions to them are allocated
if the merging modifies the origin). The exception is --namesnapshot
option which requires that you have only one snapshot. While the merging
is in progress, new ssnapshots cannot be created.
- take snapshot before system-wide package update
- do update
- now suppose that you don't like the result of the update:
- take another snapshot
- initiate -M --onactivate with the first snapshot
- after reboot you are running with root filesystem in the state before
the update and you can examine the result of the update on the second
snapshot. If you decide that you like the update anyway, you can merge it
back again (with reboot).
(don't forget to regenerate lvm on your initramdisk if you are going to
try this on root filesystem).
This implementation of snapshot merging is meant to be stable, report any
possible bugs to me.
this is exactly what I want and need, I played once with the snapshots
but found that there is no way to merge them back, thus this will make
linux be able to do full rollback. Amazing.
Now the only big thing linux is missing now, is support for good packet
writing since both direct DVDRW and pktcdvd writing is very slow.
The solution here is to create a specialized implementation of UDF
filesystem that uses its own caches and is bypassing VFS completly.
Maybe do it with fuse?
And probably it would be great to fuse RAID and LVM systems so we can
have one standard volume manager.
This can be used together with hibernating, to snapshot full system?
Full integration with swsusp would be great.
Thanks a million for such a good work.
As soon as I have free time I will try this.
dm-devel mailing list