Openstack, Wheezy and ZFS on Linux
A while ago, I started the install of an OpenStack cluster at Logilab, so our developers can play easily with any kind of environment. We are planning to improve our Apycot automatic testing platform so it can use "elastic power". And so on.
I first tried a Ubuntu Precise based setup, since at that time, Debian packages were not really usable. The setup never reached a point where it could be relased as production ready, due to the fact I tried a too complex and bleeding edge configuration (involving Quantum, openvswitch, sheepdog)...
Meanwhile, we went really short of storage capacity. For now, it mainly consists in hard drives distributed in our 19" Dell racks (generally with hardware RAID controllers). So I recently purchased a low-cost storage bay (SuperMicro SC937 with a 6Gb/s JBOD-only HBA) with 18 spinning hard drives and 4 SSDs. This storage bay being driven by ZFS on Linux (tip: the SSD-stored ZIL is a requirement to get decent performances). This storage setup is still under test for now.
I also went to the last Mini-DebConf in Paris, where Loic Dachary presented the status of the OpenStack packaging effort in Debian. This gave me the will to give a new try to OpenStack using Wheezy and a bit simpler setup. But I could not consider not to use my new ZFS-based storage as a nova volume provider. It is not available for now in OpenStack (there is a backend for Solaris, but not for ZFS on Linux). However, this is Python and in fact, the current ISCSIDriver backend needs very little to make it work with zfs instead of lvm as "elastics" block-volume provider and manager.
So, I wrote a custom nova volume driver to handle this. As I don't want the nova-volume daemon to run on my ZFS SAN, I wrote this backend mixing the SanISCSIDriver (which manages the storage system via SSH) and the standard ISCSIDriver (which uses standard Linux isci target tools). I'm not very fond of the API of the VolumeDriver (especially the fact that the ISCSIDriver is responsible for 2 roles: managing block-level volumes and exporting block-level volumes). This small design flaw (IMHO) is the reason I had to duplicate some code (not much but...) to implement my ZFSonLinuxISCSIDriver...
So here is the setup I made:
Infrastructure
My OpenStack Essex "cluster" consists for now in:
- one control node, running in a "normal" libvirt-controlled virtual
machine; it is a Wheezy that runs:
- nova-api
- nova-cert
- nova-network
- nova-scheduler
- nova-volume
- glance
- postgresql
- OpenStack dashboard
- one computing node (Dell R310, Xeon X3480, 32G, Wheezy), which runs:
- nova-api
- nova-network
- nova-compute
- ZFS-on-Linux SAN (3x raidz1 poools made of 6 1T drives, 2x (mirrored) 32G SLC SDDs, 2x 120G MLC SSDs for cache); for now, the storage is exported to the SAN via one 1G ethernet link.
OpensStack Essex setup
I mainly followed the Debian HOWTO to setup my private cloud. I mainly tuned the network settings to match my environement (and the fact my control node lives in a VM, with VLAN stuff handled by the host).
I easily got a working setup (I must admit that I think my previous experiment with OpenStack helped a lot when dealing with custom configurations... and vocabulary; I'm not sure I would have succeded "easily" following the HOWTO, but hey, it is a functionnal HOWTO, meaning if you do not follow the instructions because you want special tunings, don't blame the HOWTO).
Compared to the HOWTO, my nova.conf looks like (as of today):
[DEFAULT] logdir=/var/log/nova state_path=/var/lib/nova lock_path=/var/lock/nova root_helper=sudo nova-rootwrap auth_strategy=keystone dhcpbridge_flagfile=/etc/nova/nova.conf dhcpbridge=/usr/bin/nova-dhcpbridge sql_connection=postgresql://novacommon:XXX@control.openstack.logilab.fr/nova ## Network config # A nova-network on each compute node multi_host=true # VLan manger network_manager=nova.network.manager.VlanManager vlan_interface=eth1 # My ip my-ip=172.17.10.2 public_interface=eth0 # Dmz & metadata things dmz_cidr=169.254.169.254/32 ec2_dmz_host=169.254.169.254 metadata_host=169.254.169.254 ## More general things # The RabbitMQ host rabbit_host=control.openstack.logilab.fr ## Glance image_service=nova.image.glance.GlanceImageService glance_api_servers=control.openstack.logilab.fr:9292 use-syslog=true ec2_host=control.openstack.logilab.fr novncproxy_base_url=http://control.openstack.logilab.fr:6080/vnc_auto.html vncserver_listen=0.0.0.0 vncserver_proxyclient_address=127.0.0.1
Volume
I had a bit more work to do to make nova-volume work. First, I got hit by this nasty bug #695791 which is trivial to fix... when you know how to fix it (I noticed the bug report after I fixed it by myself).
Then, as I wanted the volumes to be stored and exported by my shiny new ZFS-on-Linux setup, I had to write my own volume driver, which was quite easy, since it is Python, and the logic to implement was already provided by the ISCSIDriver class on the one hand, and by the SanISCSIDrvier on the other hand. So I ended with this firt implementation. This file should be copied to nova volumes package directory (nova/volume/zol.py):
# vim: tabstop=4 shiftwidth=4 softtabstop=4
# Copyright 2010 United States Government as represented by the
# Administrator of the National Aeronautics and Space Administration.
# Copyright 2011 Justin Santa Barbara
# Copyright 2012 David DOUARD, LOGILAB S.A.
# All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"); you may
# not use this file except in compliance with the License. You may obtain
# a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
# License for the specific language governing permissions and limitations
# under the License.
"""
Driver for ZFS-on-Linux-stored volumes.
This is mainly a custom version of the ISCSIDriver that uses ZFS as
volume provider, generally accessed over SSH.
"""
import os
from nova import exception
from nova import flags
from nova import utils
from nova import log as logging
from nova.openstack.common import cfg
from nova.volume.driver import _iscsi_location
from nova.volume import iscsi
from nova.volume.san import SanISCSIDriver
LOG = logging.getLogger(__name__)
san_opts = [
cfg.StrOpt('san_zfs_command',
default='/sbin/zfs',
help='The ZFS command.'),
]
FLAGS = flags.FLAGS
FLAGS.register_opts(san_opts)
class ZFSonLinuxISCSIDriver(SanISCSIDriver):
"""Executes commands relating to ZFS-on-Linux-hosted ISCSI volumes.
Basic setup for a ZoL iSCSI server:
XXX
Note that current implementation of ZFS on Linux does not handle:
zfs allow/unallow
For now, needs to have root access to the ZFS host. The best is to
use a ssh key with ssh authorized_keys restriction mechanisms to
limit root access.
Make sure you can login using san_login & san_password/san_private_key
"""
ZFSCMD = FLAGS.san_zfs_command
_local_execute = utils.execute
def _getrl(self):
return self._runlocal
def _setrl(self, v):
if isinstance(v, basestring):
v = v.lower() in ('true', 't', '1', 'y', 'yes')
self._runlocal = v
run_local = property(_getrl, _setrl)
def __init__(self):
super(ZFSonLinuxISCSIDriver, self).__init__()
self.tgtadm.set_execute(self._execute)
LOG.info("run local = %s (%s)" % (self.run_local, FLAGS.san_is_local))
def set_execute(self, execute):
LOG.debug("override local execute cmd with %s (%s)" % (repr(execute), execute.__module__))
self._local_execute = execute
def _execute(self, *cmd, **kwargs):
if self.run_local:
LOG.debug("LOCAL execute cmd %s (%s)" % (cmd, kwargs))
return self._local_execute(*cmd, **kwargs)
else:
LOG.debug("SSH execute cmd %s (%s)" % (cmd, kwargs))
check_exit_code = kwargs.pop('check_exit_code', None)
command = ' '.join(cmd)
return self._run_ssh(command, check_exit_code)
def _create_volume(self, volume_name, sizestr):
zfs_poolname = self._build_zfs_poolname(volume_name)
# Create a zfs volume
cmd = [self.ZFSCMD, 'create']
if FLAGS.san_thin_provision:
cmd.append('-s')
cmd.extend(['-V', sizestr])
cmd.append(zfs_poolname)
self._execute(*cmd)
def _volume_not_present(self, volume_name):
zfs_poolname = self._build_zfs_poolname(volume_name)
try:
out, err = self._execute(self.ZFSCMD, 'list', '-H', zfs_poolname)
if out.startswith(zfs_poolname):
return False
except Exception as e:
# If the volume isn't present
return True
return False
def create_volume_from_snapshot(self, volume, snapshot):
"""Creates a volume from a snapshot."""
zfs_snap = self._build_zfs_poolname(snapshot['name'])
zfs_vol = self._build_zfs_poolname(snapshot['name'])
self._execute(self.ZFSCMD, 'clone', zfs_snap, zfs_vol)
self._execute(self.ZFSCMD, 'promote', zfs_vol)
def delete_volume(self, volume):
"""Deletes a volume."""
if self._volume_not_present(volume['name']):
# If the volume isn't present, then don't attempt to delete
return True
zfs_poolname = self._build_zfs_poolname(volume['name'])
self._execute(self.ZFSCMD, 'destroy', zfs_poolname)
def create_export(self, context, volume):
"""Creates an export for a logical volume."""
self._ensure_iscsi_targets(context, volume['host'])
iscsi_target = self.db.volume_allocate_iscsi_target(context,
volume['id'],
volume['host'])
iscsi_name = "%s%s" % (FLAGS.iscsi_target_prefix, volume['name'])
volume_path = self.local_path(volume)
# XXX (ddouard) this code is not robust: does not check for
# existing iscsi targets on the host (ie. not created by
# nova), but fixing it require a deep refactoring of the iscsi
# handling code (which is what have been done in cinder)
self.tgtadm.new_target(iscsi_name, iscsi_target)
self.tgtadm.new_logicalunit(iscsi_target, 0, volume_path)
if FLAGS.iscsi_helper == 'tgtadm':
lun = 1
else:
lun = 0
if self.run_local:
iscsi_ip_address = FLAGS.iscsi_ip_address
else:
iscsi_ip_address = FLAGS.san_ip
return {'provider_location': _iscsi_location(
iscsi_ip_address, iscsi_target, iscsi_name, lun)}
def remove_export(self, context, volume):
"""Removes an export for a logical volume."""
try:
iscsi_target = self.db.volume_get_iscsi_target_num(context,
volume['id'])
except exception.NotFound:
LOG.info(_("Skipping remove_export. No iscsi_target " +
"provisioned for volume: %d"), volume['id'])
return
try:
# ietadm show will exit with an error
# this export has already been removed
self.tgtadm.show_target(iscsi_target)
except Exception as e:
LOG.info(_("Skipping remove_export. No iscsi_target " +
"is presently exported for volume: %d"), volume['id'])
return
self.tgtadm.delete_logicalunit(iscsi_target, 0)
self.tgtadm.delete_target(iscsi_target)
def check_for_export(self, context, volume_id):
"""Make sure volume is exported."""
tid = self.db.volume_get_iscsi_target_num(context, volume_id)
try:
self.tgtadm.show_target(tid)
except exception.ProcessExecutionError, e:
# Instances remount read-only in this case.
# /etc/init.d/iscsitarget restart and rebooting nova-volume
# is better since ensure_export() works at boot time.
LOG.error(_("Cannot confirm exported volume "
"id:%(volume_id)s.") % locals())
raise
def local_path(self, volume):
zfs_poolname = self._build_zfs_poolname(volume['name'])
zvoldev = '/dev/zvol/%s' % zfs_poolname
return zvoldev
def _build_zfs_poolname(self, volume_name):
zfs_poolname = '%s%s' % (FLAGS.san_zfs_volume_base, volume_name)
return zfs_poolname
To configure my nova-volume instance (which runs on the control node, since it's only a manager), I added these to my nova.conf file:
# nove-volume config volume_driver=nova.volume.zol.ZFSonLinuxISCSIDriver iscsi_ip_address=172.17.1.7 iscsi_helper=tgtadm san_thin_provision=false san_ip=172.17.1.7 san_private_key=/etc/nova/sankey san_login=root san_zfs_volume_base=data/openstack/volume/ san_is_local=false verbose=true
Note that the private key (/etc/nova/sankey here) is stored in clear and that it must be readable by the nova user.
This key being stored in clear and giving root acces to my ZFS host, I have limited a bit this root access by using a custom command wrapper in the .ssh/authorized_keys file.
Something like (naive implementation):
[root@zfshost ~]$ cat /root/zfswrapper
#!/bin/sh
CMD=`echo $SSH_ORIGINAL_COMMAND | awk '{print $1}'`
if [ "$CMD" != "/sbin/zfs" && "$CMD" != "tgtadm" ]; then
echo "Can do only zfs/tgtadm stuff here"
exit 1
fi
echo "[`date`] $SSH_ORIGINAL_COMMAND" >> .zfsopenstack.log
exec $SSH_ORIGINAL_COMMAND
Using this in root's .ssh/authorized_keys file:
[root@zfshost ~]$ cat /root/.ssh/authorized_keys | grep control
from="control.openstack.logilab.fr",no-pty,no-port-forwarding,no-X11-forwarding,no-agent-forwarding,command="/root/zfswrapper" ssh-rsa AAAA[...] root@control
I had to set the iscsi_ip_address (the ip address of the ZFS host), but I think this is a result of something mistakenly implemented in my ZFSonLinux driver.
Using this config, I can boot an image, create a volume on my ZFS storage, and attach it to the running image.
I have to test things like snapshot, (live?) migration and so. This is a very first draft implementation which needs to be refined, improved and tested.
What's next
Besides the fact that it needs more tests, I plan to use salt for my OpenStack deployment (first to add more compute nodes in my cluster), and on the other side, I'd like to try the salt-cloud so I have a bunch of Debian images that "just work" (without the need of porting the cloud-init Ubuntu package).
On the side of my zol driver, I need to port it to Cinder, but I do not have a Folsom install to test it...