Introduction
Want to take advantage of the efficiency and improved snapshotting of
thin LVs on an existing system? It will take a little work but it is
possible. The following steps will show how to convert a CentOS 6.4
basic installation to use thin logical volumes for the root device
(containing the root filesystem).
Preparation
To kick things off there are few preparation steps we need that seem a
bit unreleated but will prove useful. First I enabled LVM to issue
discards to underlying block devices (if you are interested in why this
is needed you can check out my post
here. )
\
[root@Cent64 ~]# cat /etc/lvm/lvm.conf | grep issue_discards
issue_discards = 0
[root@Cent64 ~]# sed -i -e 's/issue_discards = 0/issue_discards = 1/' /etc/lvm/lvm.conf
[root@Cent64 ~]# cat /etc/lvm/lvm.conf | grep issue_discards
issue_discards = 1
Next, since we are converting the whole system to use thin LVs we need
to enable our initramfs to mount and switch root to a thin LV. By
default dracut does not include the utilities that are needed to do
this (see
BZ#921235 ). This
means we need to tell dracut to add thin_dump, thin_restore, and
thin_check (provided by the device-mapper-persistent-data rpm) to the
initramfs. We also want to make sure they get added for any future
initramfs building so we will add it to a file within
/usr/share/dracut/modules.d/.
\
[root@Cent64 ~]# mkdir /usr/share/dracut/modules.d/99thinlvm
[root@Cent64 ~]# cat << EOF > /usr/share/dracut/modules.d/99thinlvm/install
> #!/bin/bash
> dracut_install -o thin_dump thin_restore thin_check
> EOF
[root@Cent64 ~]# chmod +x /usr/share/dracut/modules.d/99thinlvm/install
[root@Cent64 ~]# dracut --force
[root@Cent64 ~]# lsinitrd /boot/initramfs-2.6.32-358.el6.x86_64.img | grep thin_
-rwxr-xr-x 1 root root 351816 Sep 3 23:11 usr/sbin/thin_dump
-rwxr-xr-x 1 root root 238072 Sep 3 23:11 usr/sbin/thin_check
-rwxr-xr-x 1 root root 355968 Sep 3 23:11 usr/sbin/thin_restore
OK, so now that we have an adequate initramfs the final step before the
conversion is to make sure there is enough free space in the VG to move
our data around (in the worst case scenario we will need twice the space
we are currently using). On my system I just added a 2nd disk (sdb)
and added that disk to the VG:
\
[root@Cent64 ~]# lsblk
NAME MAJ:MIN RM SIZE RO TYPE MOUNTPOINT
sr0 11:0 1 1024M 0 rom
sdb 8:16 0 31G 0 disk
sda 8:0 0 30G 0 disk
├─sda1 8:1 0 500M 0 part /boot
└─sda2 8:2 0 29.5G 0 part
├─vg_cent64-lv_root (dm-0) 253:0 0 25.6G 0 lvm /
└─vg_cent64-lv_swap (dm-1) 253:1 0 4G 0 lvm [SWAP]
[root@Cent64 ~]#
[root@Cent64 ~]# vgextend vg_cent64 /dev/sdb
Volume group "vg_cent64" successfully extended
[root@Cent64 ~]#
[root@Cent64 ~]# vgs
VG #PV #LV #SN Attr VSize VFree
vg_cent64 2 2 0 wz--n- 60.50g 31.00g
Conversion
Now comes the main event! We need to create a thin LV pool and then move
the root LV over to the pool. Since thin pools currently cannot be
reduced in size (
BZ#812731) I
decided to make my thin pool be exactly the size of the LV I wanted to
put in the pool. Below I show creating the thin pool as well as the
thin_root that will be our new “thin” root logical volume.
\
[root@Cent64 ~]# lvs --units=b /dev/vg_cent64/lv_root
LV VG Attr LSize Pool Origin Data% Move Log Cpy%Sync Convert
lv_root vg_cent64 -wi-ao--- 27455913984B
[root@Cent64 ~]#
[root@Cent64 ~]# lvcreate -T vg_cent64/thinp --size=27455913984B
Logical volume "thinp" created
[root@Cent64 ~]#
[root@Cent64 ~]# lvcreate -T vg_cent64/thinp -n thin_root -V 27455913984B
Logical volume "thin_root" created
[root@Cent64 ~]#
[root@Cent64 ~]# lvs
LV VG Attr LSize Pool Origin Data% Move Log Cpy%Sync Convert
lv_root vg_cent64 -wi-ao--- 25.57g
lv_swap vg_cent64 -wi-ao--- 3.94g
thin_root vg_cent64 Vwi-a-tz- 25.57g thinp 0.00
thinp vg_cent64 twi-a-tz- 25.57g 0.00
Now we need to get all of the data from lv_root and into thin_root.
My original thought is just to dd all of the content from one to the
other, but there is one problem: we are still mounted on lv_root. For
safety I would probably recommend booting into a rescue mode from a cd
and then doing the dd without either filesystem mounted. However,
today I just decided to make an LVM snapshot of the root LV which gives
us a consistent view of the block device for the duration of the copy.
\
[root@Cent64 ~]# lvcreate --snapshot -n snap_root --size=2g vg_cent64/lv_root
Logical volume "snap_root" created
[root@Cent64 ~]#
[root@Cent64 ~]# dd if=/dev/vg_cent64/snap_root of=/dev/vg_cent64/thin_root
53624832+0 records in
53624832+0 records out
27455913984 bytes (27 GB) copied, 597.854 s, 45.9 MB/s
[root@Cent64 ~]#
[root@Cent64 ~]# lvs
LV VG Attr LSize Pool Origin Data% Move Log Cpy%Sync Convert
lv_root vg_cent64 owi-aos-- 25.57g
lv_swap vg_cent64 -wi-ao--- 3.94g
snap_root vg_cent64 swi-a-s-- 2.00g lv_root 0.07
thin_root vg_cent64 Vwi-a-tz- 25.57g thinp 100.00
thinp vg_cent64 twi-a-tz- 25.57g 100.00
[root@Cent64 ~]#
[root@Cent64 ~]# lvremove /dev/vg_cent64/snap_root
Do you really want to remove active logical volume snap_root? [y/n]: y
Logical volume "snap_root" successfully removed
So there we have it. All of the data has been copied to the thin_root
LV. You can see from the output of lvs that the thin LV and the thin
pool are both 100% full. 100% full? really? I thought these were
“thin” LVs. :)
Let’s recover that space! I’ll do this by mounting thin_root and then
running fstrim to release the unused blocks back to the pool. First I
check the fs and clean up any dirt by running fsck.
\
[root@Cent64 ~]# fsck /dev/vg_cent64/thin_root
fsck from util-linux-ng 2.17.2
e2fsck 1.41.12 (17-May-2010)
Clearing orphaned inode 1047627 (uid=0, gid=0, mode=0100700, size=0)
Clearing orphaned inode 1182865 (uid=0, gid=0, mode=0100755, size=15296)
Clearing orphaned inode 1182869 (uid=0, gid=0, mode=0100755, size=24744)
Clearing orphaned inode 1444589 (uid=0, gid=0, mode=0100755, size=15256)
...
/dev/mapper/vg_cent64-thin_root: clean, 30776/1676080 files, 340024/6703104 blocks
[root@Cent64 ~]#
[root@Cent64 ~]# mount /dev/vg_cent64/thin_root /mnt/
[root@Cent64 ~]#
[root@Cent64 ~]# fstrim -v /mnt/
/mnt/: 26058436608 bytes were trimmed
[root@Cent64 ~]#
[root@Cent64 ~]# lvs
LV VG Attr LSize Pool Origin Data% Move Log Cpy%Sync Convert
lv_root vg_cent64 -wi-ao--- 25.57g
lv_swap vg_cent64 -wi-ao--- 3.94g
thin_root vg_cent64 Vwi-aotz- 25.57g thinp 5.13
thinp vg_cent64 twi-a-tz- 25.57g 5.13
Success! All the way from 100% back down to 5%.
Now let’s update the grub.conf and the fstab to use the new
thin_root LV.
NOTE: grub.conf is on the filesystem on sda1.
NOTE: fstab is on the filesystem on thin_root.
\
[root@Cent64 ~]# sed -i -e 's/lv_root/thin_root/g' /boot/grub/grub.conf
[root@Cent64 ~]# sed -i -e 's/lv_root/thin_root/g' /mnt/etc/fstab
[root@Cent64 ~]# umount /mnt/
Time for a reboot!
After the system comes back up we should now be able to delete the
original lv_root.
\
[root@Cent64 ~]# lvremove /dev/vg_cent64/lv_root
Do you really want to remove active logical volume lv_root? [y/n]: y
Logical volume "lv_root" successfully removed
Now we want to remove that extra disk (/dev/sdb) I added. However
there is a subtle difference between my system now and my system before.
There is metadata LV (thinp_tmeta) that is taking up a minute amount
of space that is preventing us from being able to fit completely on the
first disk (/dev/sda).
No biggie. We’ll just steal this amount of space from lv_swap. And
then run pvmove to move all data back to /dev/sda.
\
[root@Cent64 ~]# lvs -a --units=b
LV VG Attr LSize Pool Origin Data% Move Log Cpy%Sync Convert
lv_swap vg_cent64 -wi-ao--- 4227858432B
thin_root vg_cent64 Vwi-aotz- 27455913984B thinp 5.13
thinp vg_cent64 twi-a-tz- 27455913984B 5.13
[thinp_tdata] vg_cent64 Twi-aot-- 27455913984B
[thinp_tmeta] vg_cent64 ewi-aot-- 29360128B
[root@Cent64 ~]#
[root@Cent64 ~]# swapoff /dev/vg_cent64/lv_swap
[root@Cent64 ~]#
[root@Cent64 ~]# lvresize --size=-29360128B /dev/vg_cent64/lv_swap
WARNING: Reducing active logical volume to 3.91 GiB
THIS MAY DESTROY YOUR DATA (filesystem etc.)
Do you really want to reduce lv_swap? [y/n]: y
Reducing logical volume lv_swap to 3.91 GiB
Logical volume lv_swap successfully resized
[root@Cent64 ~]#
[root@Cent64 ~]# mkswap /dev/vg_cent64/lv_swap
mkswap: /dev/vg_cent64/lv_swap: warning: don't erase bootbits sectors
on whole disk. Use -f to force.
Setting up swapspace version 1, size = 4100092 KiB
no label, UUID=7b023342-a9a9-4676-8bc6-1e60541010e4
[root@Cent64 ~]#
[root@Cent64 ~]# swapon -v /dev/vg_cent64/lv_swap
swapon on /dev/vg_cent64/lv_swap
swapon: /dev/mapper/vg_cent64-lv_swap: found swap signature: version 1, page-size 4, same byte order
swapon: /dev/mapper/vg_cent64-lv_swap: pagesize=4096, swapsize=4198498304, devsize=4198498304
Now we can get rid of sdb by running pvmove and vgreduce.
\
[root@Cent64 ~]# pvmove /dev/sdb
/dev/sdb: Moved: 0.1%
/dev/sdb: Moved: 11.8%
/dev/sdb: Moved: 21.0%
/dev/sdb: Moved: 32.0%
/dev/sdb: Moved: 45.6%
/dev/sdb: Moved: 56.2%
/dev/sdb: Moved: 68.7%
/dev/sdb: Moved: 79.6%
/dev/sdb: Moved: 90.7%
/dev/sdb: Moved: 100.0%
[root@Cent64 ~]#
[root@Cent64 ~]# pvs
PV VG Fmt Attr PSize PFree
/dev/sda2 vg_cent64 lvm2 a-- 29.51g 0
/dev/sdb vg_cent64 lvm2 a-- 31.00g 31.00g
[root@Cent64 ~]#
[root@Cent64 ~]# vgreduce vg_cent64 /dev/sdb
Removed "/dev/sdb" from volume group "vg_cent64"
Boom! You’re done!
Dusty