Introduction

About oVirt

• Virtualization platform built on top of KVM, QEMU, Libvirt
• Manages virtual machines, storage, networks

Ceph in a nutshell

• Ceph is a massively scalable and highly reliable software defined storage
• Ceph provides object storage, block storage, and file system
• oVirt only uses Ceph's block storage (RBD)

What is Cinder?

• A Block Storage service for the OpenStack compute project (Nova)
• Supports many storage backends
• oVirt only uses the Ceph backend

OpenStack Java SDK

• oVirt Engine communicates with Cinder via the OpenStack Java SDK
• Support for Cinder added by Daniel Erez
• https://github.com/woorea/openstack-java-sdk

Focusing on Ceph integration

For more information on the Cinder integration, see Maor Lipchuk's talk from FOSDEM 2016 Managing Ceph through Cinder using oVirt.

oVirt traditional storage

traditional storage from bird's view

The SPM

• Only the SPM can do metadata operations
• Complex, single point of failure
• Will be removed in 4.x

Block storage

• iSCSI
• Fibre Channel

Block storage domain

A VG, special LVs for metadata, locking, messaging.

# lvs -o name,attr,size 6acaac02-7e7d-...-70e07d96581d
  LV                                   Attr       LSize
  56b25c4a-085f-406e-8a2e-acb06a03d0cc -wi-a----- 128.00m
  78f22775-916c-4e72-8c5b-9917734b26da -wi-a-----   4.50g
  c00fd591-f85e-4572-b82e-35c542e383a7 -wi-a----- 128.00m
  ids                                  -wi-ao---- 128.00m
  inbox                                -wi-a----- 128.00m
  leases                               -wi-a-----   2.00g
  master                               -wi-ao----   1.00g
  metadata                             -wi-a----- 512.00m
  outbox                               -wi-a----- 128.00m

Block Volume

Block storage problems

• Scale (350 LVs per VG)
• Complex
• Hard to manage
• Timeouts (much better since EL 7.2)
• QOS (for storage operations)

Block storage thin provisioning

Block thin provisioning problems

• Scale (polling)
• Complex
• Slow (VM may pause)
• Single point of failure (SPM)

File storage

• NFS
• GlusterFS
• POSIX
• Local FS

File storage domain

A directory on a mounted filesystem

# ls -lh /rhev/data-center/mnt/example.com:_export_data/869...783
total 4.0K
drwxr-xr-x. 2 vdsm kvm   69 Jan 28 22:56 dom_md
drwxr-xr-x. 7 vdsm kvm 4.0K Jan 23 19:13 images
drwxr-xr-x. 2 vdsm kvm    6 Jan 28 22:56 master

File volume

Files in the image directory

# ls -lh /rhev/data-center/mnt/example.com:_export_data/869...783/images/24c...57c
total 1.7G
-rw-rw----. 1 vdsm kvm 1.7G Dec  4 15:52 1f6...e06
-rw-rw----. 1 vdsm kvm 1.0M Dec  4 15:49 1f6...e06.lease
-rw-r--r--. 1 vdsm kvm  309 Dec  4 15:52 1f6...e06.meta

File storage thin provisioning

(file system handles it)

File storage problems

• Overhead
• Timeouts (ioprocess)
• No QOS (for storage operations)
• Reliability

Snapshots

It's QCOW all the way down

Example: taking a snapshot

Example: taking a snapshot

Example: taking a snapshot

Example: taking a snapshot

Snapshots problems

• Performance
• Scale (on block storage)
• Difficult to manage

Live merge

Extremely complicated - we will not talk about it.

See Adam Litke's talk "Thanks for Live Snapshots, Where's Live Merge?" from KVM Forum 2014.

★ New: Ceph storage ★

Ceph storage from bird's view

Where is the SPM?

(No dependency on the SPM)

Ceph storage domain

• Nothing on host
• Cinder provider on Engine side
• Only Ceph supported for now

Ceph volume

<disk type='network' device='disk'>
    <source protocol='rbd' name='pool/volume'>
        <host name='10.0.1.1' port='6789'/>
        <host name='10.0.1.2' port='6789'/>
        <host name='10.0.1.3' port='6789'/>
    </source>
    <auth username="cinder">
        <secret type="ceph" uuid="ec5...236"/>
    </auth>
</disk>

Ceph volume thin provisioning

(Nothing to do)

Ceph snapshots

• Nothing on the host - transparent
• Using Cinder APIs on Engine side, via OpenStack Java SDK

Example: taking a Ceph snapshot

Example: taking a Ceph snapshot

Example: taking a Ceph snapshot

Example: taking a Ceph snapshot

Example: taking a Ceph snapshot

Live merge

(No such thing, just delete a snapshot)

Ceph storage benefits

• Highly available
• Scalable
• Cost effective
• Offloading
• Zero management (on host)

Ceph storage problems

• Feature incomplete
• Cinder and oVirt have different views on things
• Too complicated for small oVirt deployment
• Cinder is the new SPM

Keeping secrets

Cephx authentication

# cat /etc/ceph/ceph.client.admin.keyring
[client.admin]
   key = AQCj2YpRiAe6CxAA7/ETt7Hcl9IyxyYciVs47w==

We need a more secure way to deploy the key

Libvirt secrets

<secret ephemeral="yes" private="yes">
    <uuid>3a27b133-abb2-...-bd0a4032866f</uuid>
    <usage type="ceph">
        <name>ovirt/3a27b133-abb2-...-bd0a4032866f</name>
    </usage>
</secret>

What if Libvirt crashed or restarted?

Engine manages Ceph auth keys

• Store encrypted auth keys
• Register Livbirt secrets with all hosts
• Unregister secrets when unneeded

registerSecrets

##
# @Host.registerSecrets:
#
# Register secrets with libvirt
#
# @secrets:     List of secrets to register
#
# @clear:       #optional If true, unregister other
#               registered secrets
#
# Since: 4.17.0
##

unregisterSecrets

##
# @Host.unregisterSecrets:
#
# Unregister secrets registered for given uuids
#
# @uuids:       List of UUIDs
#
# Since: 4.17.0
##

Registering secrets

• When host is activated
• When storage domain is activated
• When reconnecting to a host

Unregistering secrets

• When host is put to maintenance
• When storage domain is put to maintenance
• Vdsm clears secrets on startup and shutdown

Live snapshot

Consistent snapshots

Requires qemu-guest-agent on Linux, VSS on Windows

Traditional snapshot flow

Freezing too late!

Engine must control the entire flow

Freeze guest file systems before creating Ceph disk snapshot

freeze

New verb using Libvirt's virDomainFSFreeze

##
# @VM.freeze:
#
# Freeze mounted filesystems within the guest (hence
# guest agent may be required depending on hypervisor
# used).
#
# @vmID:         The UUID of the VM
#
# Since: 4.17.0
##

snapshot

New frozen flag

##
# @VM.snapshot:
#
# Create a live snapshot of a VM's disks.
#
...
# @frozen:      #optional Boolean value set to true if
#               the vm is frozen and does not need to
#               be freezed during the snapshot.
#               (new in version 4.17.0)
#
# Since: 4.10.0
##

thaw

New verb using Libvirt's virDomainFSThaw

##
# @VM.thaw:
#
# Thaw mounted filesystems within the guest (hence
# guest agent may be required depending on hypervisor
# used).
#
# @vmID:         The UUID of the VM
#
# Since: 4.17.0
##

Live snapshot with Ceph disk

Engine now controls the entire flow

Live snapshot without Ceph disks

• Take snapshot via Vdsm (frozen=False)
• Vdsm will freeze and thaw guest file systems if needed

from __future__ import directions

Complete missing features

• Move disk
• Live storage migration
• Disaster recovery
• Monitoring

Stateless secret management

• No management is better than some
• Register temporary secret during VM startup
• Migrations will be a challenge

Cinder appliance

• Cinder does not provide the control we need
• Make our Cinder appliance, control it via an agent
• Hide Cinder behind engine UI

Drop Cinder?

• We don't need the middleman
• But we want the drivers...

Drop traditional storage?

Hard to modify, easy to break