Roy's notes

Linux' Logical Volume Manager (LVM)

LVM in general

LVM is designed to be an abstraction layer on top of physical drives or RAID, typically mdraid or fakeraid. Keep in mind that fakeraid should be avoided unless you really need it, like in conjunction with dual-booting linux and windows on fakeraid. LVM broadly consists of three elements, the "physical" devices (PV), the volume group (VG) and the logical volume (LV). There can be multiple PVs, VGs and LVs, depending on requirement. More about this below. All commands are given as examples, and all of them can be fine-tuned using extra flags in need of so. In my experience, the defaults work well for most usecases.

I'm mentioning filesystem below too. Where I write ext4, the samme applies to ext2 and ext3.

Create a PV

A PV is the "physical" parts. This does not need to be a physical disk, but can also be another RAID, be it an mdraid, fakeraid, hardware raid, a virtual disk on a SAN or otherwisee or a partition on one of those.

#  These add three PVs, one on a drive (or hardware raid) and another two on mdraids.
pvcreate /dev/sdb
pvcreate /dev/md1 /dev/md2

For more information about pvcreate, see the manual.

Create a VG

The volume group consists of one or more PVs grouped together on which LVs can be placed. If several PVs are grouped in a VG, it's generally a good idea to make sure these PVs have some sort of redundancy, as in mdraid/fakeraid or hwraid. Otherwise it will be like using a RAID-0 with a single point of failure on each of the independant drives. LVM has RAID code in it as well, so you can use that. I haven't done so myself, as I generally stick to mraid. The reason is mdraid is, in my opinion older and more stable and has more users (meaning bugs are reported and fixed faster whenever they are found). That said, I beleive the actual RAID code used in LVM RAID are the same function calls as for mdraid, so it may not be much of a difference. I still stick with mdraid. To create a VG, run

# Create volume group my_vg
vgcreate my_vg /dev/md1

Note that if vgcreate is run with a PV (as /dev/md1 above) that is not defined as a PV (like above), this is done implicitly, so if you don't need any special flags to pvcreate, you can simply skip it and let vgcreate do that for you.

Create an LV

LVs can be compared to partitions, somehow, since they are bounderies of a fraction or all of that of a VG. The difference between them and a partition, however, is that they can be grown or shrunk easily and also moved around between PVs without downtime. This flexibility makes them superiour to partitions as your system can be changed without users noticing it. By default, an LV is alloated "thickly", meaning all the data given to it, is allocated from the VG and thus the PV. The following makes a 100GB LV named "thicklv". When making an LV, I usually allocate what's needed plus some more, but not everything, just to make sure it's space available for growth on any of the LVs on the VG, or new LVs.

# Create a thick provisioned LV named thicklv on the VG my_vg
lvcreate -n thicklv -L 100G my_vg

After the LV is created, a filesystem can be placed on it unless it is meant to be used directly. The application for direct use include swap space, VM storage and certain database systems. Most of these will, however, work on filesystems too, although my testing has shown that on swap space, there is a significant performance gain for using dedicated storage without a filesystem. As for filesystems, most Linux users use either ext4 or xfs. Personally, I generallly use XFS these days. The only thing left that can't be done on XFS is shrinking a filesystem.

# Create a filesystem on the LV - this could have been mkfs -t ext4 or whatever filesystem you choose
mkfs -t xfs /dev/my_vg/thicklv

Then just edit /etc/fstab with correct data and run mount -a, and you should be all set.

Create LVM cache

lvcreate -L 1G -n _cache_meta data /dev/md1

lvcreate -l100%FREE -n _cache data /dev/md1

# Some would want --cachemode writeback, but seriously, I wouldn't recommend it. Metadata or data can be easily corrupted in case of failure.
lvconvert --type cache-pool --cachemode writethrough --poolmetadata data/_cache_meta data/_cache

lvconvert --type cache --cachepool data/_cache data/data

Growing devices

LVM objects can be grown and shrunk. If a PV resides on a RAID where a new drive has been added or otherwise grown, or on a partition or virtual disk that has been extended, the PV must be updated to reflect these changes. The following command will grow the PV to the maximum available on the underlying storage.

# Resize the PV /dev/md (not the RAID, only the PV on the RAID) to its full size
pvresize /dev/md1

If a new PV is added, the VG can be grown to add the space on that in addition to what's there already.

# Extend my_vg to include md2 and its space
vgextend my_vg /dev/md2

With more space available in the VG, the LV can now be extended. Let's add another 50GB to it.

# Extend thicklv - add 50GB. If you know you won't need the space anywhere else, you may want to
# lvresize -l+100%FREE instead. Keep in mind the difference between -l (extents) and -L (bytes)
lvresize -L +50G my_vg/thicklv

After the LV has grown, run xfs_growfs (xfs) or resize2fs (ext4) to make use of the new data.

Migration tools

At times, storage regimes change, new storage is added and sometimes it's not easy to migrate with the current hardware or its support systems. Once I had to migrate a 45TiB fileserver from one storage system to another, preferably without downtime. The server originally had three 15TiB PVs of which two were full and the third half full. I resorted to using pvmove to just move the data on the PVs in use to a new PV. We started out with creating a new 50TiB PV, sde, and then to pvmove.

# Attach /dev/sde to the VG my_vg
vgextend my_vg /dev/sde

# Move the contents from /dev/sdb over to /dev/sde block by block. If a target is not given in pvmove,
# the contents of the source (sdb here) will be put somewhere else in the pool.
pvmove /dev/sdb /dev/sde

This took a while (as in a week or so) - the pvmove command uses old code and logic (or at least did that when I migrated this in November 2016), but it affected performance on the server very little, so users didn't notice. After the first PV was migrated, I continued with the other two, one after the other, and after a month or so, it was migrated. We used this on a number of large fileservers, and it worked flawlessly. Before doing the production servers I also tried interrupting pvmove in various ways, including hard resets, and it just kept on after the reboot.

NOTE: Even though it worked well for us, always keep a good backup before doing this. Things may go wrong, and without a good backup, you may be looking for a hard-to-find new job the next morning.

mdadm workarounds

Migrate from a mirror (raid-1) to raid-10

Create a mirror

mdadm --create /dev/md0 --level=1 --raid-devices=2 /dev/sdb /dev/sdc

LVM (pv/vg/lv) on md0 (see above), put a filesystem on the lv and fill it with some data.

Now, some months later, you want to change this to raid-10 to allow for the same amount of redundancy, but to allow more disks into the system.

mdadm --grow /dev/md0 --level=10
mdadm: Impossibly level change request for RAID1

So - no - doesn't work. But - we're on

Since we're on lvm already, this'll be easy. If you're using filesystems directly on partitions, you can do this the same way, but without the pvmove part, using rsync or whateever you like instead. I'd recommend using lvm for for the new raid, which should be rather obvious from this article. Now plug in a new drive and create a new raid10 on that one. If you two new drives, install both.

# change "missing" to the other device name if you installed two new drives
mdadm --create /dev/md1 --level=10 --raid-devices=2 /dev/vdd missing

Now, as described above, just vgextend the vg, adding the new raid and run pvmove to move the data from the old pv (residing on the old raid1) to the new pv (on raid10). Afte rpvmove is finished (which may take awile, see above), just

vgreduce raidtest /dev/md0
pvremove /dev/md0
mdadm --stop /dev/md0

…and your disk is free to be added to the new array. If the new raid is running in degraded mode (if you created it with just one drive), better don't wait too long, since you don't have redundancy. Just mdadm --add the devs.

If you now have /dev/mdstat telling you your raid10 is active and have three drives, of which one is a spare, it should look something like this

Personalities : [linear] [multipath] [raid0] [raid1] [raid6] [raid5] [raid4] [raid10]
md1 : active raid10 vdc[3](S) vdb[2] vdd[0]
      8380416 blocks super 1.2 2 near-copies [2/2] [UU]

Just mdadm --grow --raid-devices=3 /dev/md1

RAID section is not complete. I'll write more on migraing from raid10 to raid6 later. It's not straight forward, but easier than this one :)

Thin provisioned LVs

Thin provisioning on LVM is a method used for not allocating all the space given to an LV. For instance, if you have a 1TB VG and want to give an LV 500GB, although it currently only uses a fraction of that, a thin lv can be a good alternative. This will allow for adding a limit to how much it can use, but also to let it grow dynamically without manual work by the sysadmin. Thin provisioning adds another layer of abstaction by creating a special LV as a thin pool from which data is allocated to the thin volume(s)

Create a thin pool

Allocate 1TB to the thin pool to be used for thinly provisioned LVs. Keep in mind that the space allocated to the thin pool is in fact thick provisioned. Only the volumes put on thinpool are thin provisioned. This will create two hidden LVs, one for data and one for metadata. Normally the defaults will do, but check the manual if the data doesn't match the usual patterns (such as billions of files, resulting in huge amounts of metadata). I beleive the metadata part should be resizable at a later time if needed, but I have not tested it.

# Create a pool for thin LVs. Keep in mind that the pool itself is not thin provisioned, only the volumes residing on it
lvcreate --size 1T --type thin-pool --thinpool thinpool my_vg

Create a thin volume

You have the thinpool, now put a thin volume on it. It will allocate some space for metadata, but probably not much (a few megs, perhaps).

# Now, create a volume with a virtual (-V) size of half a terabyte named thin_pool, but using the thinpool (-T) named thin_pool for storage
lvcreate -V 500G -T mh_vg/thin_pool --name thinvol

The new volume's device name will be /dev/thinvol. Now, create a filesystem on it, add to fstab and mount it. The df command will return available space according to the virtual size (-V), while lvs will show how much data is actually used on each of the thinly provisioned volumes.

Low level disk handling

This section is for low-level handling of disks, regardless of storage system elsewhere. These apply to mdraid, lvmraid and zfs and possibly other solutions, with the exception of individual drives on hwraid (and maybe fakeraid), since there, the drives are hidden from Linux (or whatever OS).

"Unplug" drive from system

# Find the device unit's number with something like
find /sys/bus/scsi/devices/*/block/ -name sdd
# returning /sys/bus/scsi/devices/3:0:0:0/block/sdd here, 
# meaning 3:0:0:0 is the SCSI device we're looking for.

# Given this is the correct device, and you want to 'unplug' it, do so by
echo 1 > /sys/bus/scsi/devices/3:0:0:0/delete

Rescan disk controllers

If a drive is added and for some reason doesn't get detected, or is removed by the command above, it can be rediscovered with a sysfs command to the scsi host to which it is connected. Since there may be quite a few of these, an easy way is to just scan them all and check the output of dmesg -T after running it.

# Make a list of controllers and send a rescan message to each of them. It won't do anything 
# for those where nothing has changed, but it will show new drives where they have been added.
for host_scan in /sys/class/scsi_host/host*/scan
do
  echo '- - -' > $host_scan
done

Fail injection with debugfs

https://lxadm.com/Using_fault_injection

mdadm stuff

Spare pools

In the old itmes, mdadm supported only a dedicated spare per md device, so if working with a large set of drives (20? 40?), where you'd want to setup more raid sets to increase redundancy, you'd dedicate a spare to each of the raid sets. This has changed (some time ago?), so in these modern, heathen days, you can change mdadm.conf, usually placed under /etc/mdadm, and add 'spare-group=somegroup' where 'somegroup' is a string identifying the spare group. After doing this, run update-initramfs -u and reboot and add a spare to one of the raid sets in the spare group, and md will use that or those spares for all the raidsets in that group.

Example config

# Create two raidsets

mdadm --create /dev/md1 --level=6 --raid-devices=6 /dev/vd[efghij]
mdadm --create /dev/md2 --level=6 --raid-devices=6 /dev/vd[klmnop]

# get their UUID etc

ARRAY /dev/md/1 metadata=1.2 name=raidtest:1 UUID=1a8cbdcb:f4092350:348b6b80:c054d74c
ARRAY /dev/md/2 metadata=1.2 name=raidtest:2 UUID=894b1b7c:cb7eba70:917d6033:ea5afd2b

# Put those lines into /etc/mdadm/mdadm.conf and and add the spare-group

ARRAY /dev/md/1 metadata=1.2 name=raidtest:1 spare-group=raidtest UUID=1a8cbdcb:f4092350:348b6b80:c054d74c
ARRAY /dev/md/2 metadata=1.2 name=raidtest:2 spare-group=raidtest UUID=894b1b7c:cb7eba70:917d6033:ea5afd2b

# update the initramfs and reboot

update-initramfs -u

# add a spare drive to one of them

mdadm --add /dev/md1 /dev/vdq

# fail a drive on the other raid

mdadm --fail /dev/md2 /dev/vdn

# check /proc/mdstat to see md2 rebuilding with the spare from md1