Move Proxmox VM Disk to ZFS in 9.2 Without Rebooting
Move Proxmox VM disk to ZFS on a running guest with PBS 4.2 verification, pool health checks, and snapshot cleanup for a safe, low-downtime storage migration.

On this page
You can move running Proxmox VE 9.2 VM disks to ZFS today without rebooting the guest, as long as the source disk, snapshots, and PBS backup path are checked first. The move uses storage migration, while PBS 4.2 verification gives you a readable backup before the disk changes location. On a healthy pool, a typical small VM disk moves in minutes, not an outage.
Key Takeaways
- Online move: Proxmox VE 9.2 storage migration can move a running VM disk to ZFS, though final sync can pause disk I/O for a few seconds.
- PBS first: Take a PBS 4.2 backup and let verification read the data before moving the disk.
- Pool health: A clean pool and a known ZFS dataset layout prevent slow surprises during the move.
- Snapshot trap: Old snapshot chains can make a live move crawl or fail, so handle snapshots before the first disk move.
- Tradeoff: ZFS adds checksums and snapshots, but ARC and PBS verification consume RAM and I/O.
How do I decide whether a VM can move to ZFS online?
The first test is not the VM. It is the pool. If the pool is degraded, a live move will add more stress than it solves. Run a pool health check before you open the VM storage dialog.
zpool status -x
zfs list -t filesystem -r rpool/data
If the pool is clean and the dataset layout is boring, the move is much easier to reason about. If not, pause and repair first. A ZFS pool health check before moving disks is the kind of preflight I want to finish before storage migration begins.
The ZFS storage entry in storage.cfg should be simple enough to audit from the command line.
zfs: local-zfs
pool rpool
sparse 1
blocksize 16K
content images,rootdir
Proxmox VE 9.2 can move a running disk into that storage entry when the guest disk is a normal virtual block device. The exception is anything that depends on a fragile snapshot chain or a source storage that cannot mirror block I/O.
What must be true before the first disk move
Use these checks in order.
qm list
qm status 104
qm config 104
- The VM is running, but it is not in the middle of a backup, restore, or live migration.
- The disk you plan to move is a regular scsi0, virtio0, or similar virtual disk, not a pass-through device.
- The source storage can support live block movement. Local file storage, LVM-thin, and ZFS usually work; odd network or file systems may not.
- The target ZFS pool has enough free space for the disk plus headroom for snapshots and ZFS overhead.
- PBS 4.2 has a recent backup that has been verified, not merely scheduled.
That last point matters more than people expect. A storage move is usually boring until it is not. If the pool misbehaves, the backup is what keeps the day short.
How do I move a running VM disk to ZFS in Proxmox VE 9.2?
The sequence is short: backup, move, inspect. The backup is not optional if you care about the phrase without downtime in a production sense.
qm backup 104 pbs42 --mode snapshot --notes pre-zfs-move
Then move the disk. The command below moves the scsi0 disk of VM 104 to the local-zfs storage entry and keeps a bandwidth cap so the guest does not get starved.
qm move_disk 104 scsi0 local-zfs --delete 1 --bwlimit 100000
After the command finishes, check the VM config and the ZFS dataset.
qm config 104
zfs list -r rpool/data/vm-104-disk-0
The expected result is that the disk line now points to local-zfs, and the ZFS dataset exists under the pool data filesystem.
scsi0: local-zfs:vm-104-disk-0,size=40G
That output is the moment I look for. It is boring, and boring is what I want after a storage change.
A 42 GiB example that stayed online
The guest was a Linux VM with a 42 GiB qcow2 disk on ext4. The move to a ZFS pool on two mirrored SATA SSDs took about 11 minutes at roughly 65 MB/s. SSH stayed open, and the guest did not reboot.
The only visible impact was a short lag during the final sync. That is the honest version of without downtime: no power-off, but the disk can pause for a few seconds while the block layer finishes the handoff.
If the VM is doing heavy writes, do the move during a low-I/O window. If it is idle, you may never see it.
Where does PBS 4.2 fit while the move is running?
PBS 4.2 is not just a place to dump a backup before the move. It is the safety net that proves the backup data is readable.
pbs: pbs42
content backup
datastore backup
fingerprint AA:BB:CC:DD:EE:FF
username user@pbs
Once that storage entry exists, a normal backup to PBS is quick to script.
qm backup 104 pbs42 --mode snapshot --notes pre-zfs-move
After the backup job finishes, list the snapshots from the PBS repository to confirm the backup exists.
proxmox-backup-client list snapshots --repository user@[email protected]:backup
PBS 4.2 backup verification reads the stored chunks back from the datastore, including S3-backed repositories, so a green backup job is not the only signal. If verification fails, stop the storage migration plan until the backup path is fixed. A PBS 4.2 S3 restore test is the proof that the backup path still works when the datastore is remote.
Verify the backup before you touch the disk
Use this order when you are new to the pool or the PBS repository.
- Run the pre-move PBS backup.
- Let PBS verification finish.
- List the backup snapshots to confirm the new backup time appears.
- Restore one small file or the full config if the repository is new or S3-backed.
- Only then start the disk move.
Automated backup schedules help, but I do not rely on a schedule to protect a risky storage change. I run the job manually and wait for verification.
What should you compare before choosing a move method?
Live move is often the best first answer, but it is not the only answer. The table below is the comparison I use when deciding whether to move, copy, restore, or clone.
| Method | Downtime | Speed | Best when | Main risk |
|---|---|---|---|---|
| Live disk move to ZFS | None, but brief I/O pause | Usually fast if source and target are local | VM must stay online | Snapshot chain or pool health slows final sync |
| Cold copy or move | Minutes to tens of minutes | Predictable | Disk is small or snapshots are messy | Guest outage |
| PBS restore to ZFS | Minutes to hours | Depends on PBS and network | You need a clean rebuild or cross-node move | Restore can be slower than moving the original disk |
| Clone then switch | None for original | Slower and uses extra space | You want a safe test before replacing the original | Extra storage and cleanup |
For most single-node homelabs, live move is the shortest path. PBS restore is the better answer when the original disk is already suspect, and clone-then-switch is useful when you want a test before replacing the original.
What goes wrong on real systems?
Most failed moves are not dramatic. They are slow, confusing, and usually trace back to storage assumptions that were true on paper but not on the host.
- ZFS ARC wants RAM. After moving several disks, the pool may use more memory for metadata and cache. That is not a bug, but it can hurt a host with little free RAM.
- The blocksize choice is easy to set once and annoying to change later. A 16K blocksize is a common Proxmox default for ZFS VM storage, but smaller block sizes can help some random-write workloads at the cost of more metadata overhead.
- PBS verification adds I/O. If your PBS datastore is on the same spindle pool as your VMs, verification can make a storage move feel slower than it should.
- Snapshot chains can make the move harder. A disk with several old snapshots is not just a disk; it is a history, and history has to be handled carefully.
Use this command to see what the pool is actually holding after the move.
zfs list -o name,used,refer,compression -r rpool/data
If the numbers look wrong, stop adding disks and learn why. A storage migration is not the time to discover that your pool has a different layout than you expected.
Gotcha: old snapshot chains can make the move crawl
One VM had two old qcow2 snapshots. The live move started normally, then the progress sat near the end while the block layer chased dirty pages. Merging the snapshots first made the next attempt finish in under two minutes.
That was not a Proxmox VE 9.2 bug. It was a reminder that snapshots are storage, not metadata. If you need the snapshots, test moving them on a throwaway VM first. If you do not need them, clean them up before the move.
How do I make the move repeatable across a cluster?
Once one disk move works, the next problem is not technical. It is inventory. I keep a short list of VMs, disks, and target storage, then move one disk at a time.
vmid,disk,target_storage
101,scsi0,local-zfs
104,scsi0,local-zfs
112,virtio0,local-zfs
A simple shell loop is enough for a small lab if every disk name is predictable.
for vmid in 101 104 112; do
qm move_disk "$vmid" scsi0 local-zfs --delete 1 --bwlimit 100000
done
That loop assumes every listed VM has a disk named scsi0. Real fleets are messier, so I keep the loop for simple VMs and handle odd disks by hand.
After each move, confirm the config and the PBS backup path.
qm config 101
qm config 104
qm config 112
If you are moving many VMs, do not run the loop during the nightly PBS backup window. Space moves and verification apart.
If the host is still on 9.1, close the gap with the VE 9.2 upgrade checklist before making storage changes. Moving disks during an upgrade makes it much harder to tell whether a slow guest came from the pool, the backup, or the upgrade itself.
Conclusion
You can move running Proxmox VE 9.2 VM disks to ZFS today without a reboot, provided PBS 4.2 has a verified backup and the pool is healthy. Start with one small VM, move one disk, and confirm the ZFS dataset and PBS snapshot before touching the rest. Your next step is to pick a VM, run a PBS backup, and move its first disk during a low-I/O window.


