Clearly show if a volume is found and if not, that the pool is being refreshed
and the fetch is tried again.
Due to my commit b53a9dcc9f the chance of a volume
not being found is slightly bigger, but the performance gain is enormous on larger
deployments.
This is why we clearly have to log that we are refreshing the pool information
when a volume is not found.
It could be that a volume is created on host A and a few seconds later host B tries
to access the volume. In that case host B's libvirt doesn't know about the volume
yet and has to refresh the pool before it does.
On larger (especially RBD) storage pools this can take a lot of
time slowing operations like creating volumes down.
The getStorageStats command will still ask a pool to be refreshed so
that the management server has accurate information about the storage pools.
On larger deployments, with thousands of volumes in one pool, this should
significantly improve storage related operations
For ResizeVolume API command -
1. If hypervisor resource throws an exception, handle the NPE thrown by the job framework.
2. Improve user error message in case of RuntimeException by throwing the exception instead of 'Unexpected Exception'.
We don't need an external script to investigate the format of the RBD volume,
we only have to ask Libvirt to resize the volume and that will ask librbd to
do so.
In situations where libvirt lost the storage pool the KVM Agent will re-create the
storage pool in libvirt.
This could be then libvirt is restarted for example.
The object returned internally was missing essential information like the sourceDir
aka the Ceph pool, the monitor IPs, cephx information and such.
In this case the first operation on this newly created pool would fail. All operations
afterwards would succeed.
We used to create the snapshot after the copy from Secondary Storage,
but it could be that we never use the snapshot.
Now we check if the snapshot exists prior to performing the cloning operation
Since we use qemu-img to copy from RBD to Secondary Storage we no
longer have to force to RAW images, but can stick with QCOW2
When the snapshot backups are QCOW2 format they can easily be deployed
again when restoring from a backup
The KVMStorageProcessor no longer has a hardcoded if-statement which sets
RBD volumes to RAW, this is now handled in the LibvirtStorageAdapter
The Management Server still sends QCOW2 as format. That's a fix for later.
fix mismatch of ovs-host-setup, ovs_host_setup used Libvirt resource and
scripts
plug the nic to OVS bridges created for the tunnel network.
Conflicts:
plugins/hypervisors/kvm/src/com/cloud/hypervisor/kvm/resource/OvsVifDriver.java
Added a new flag 'checkBeforeCleanup' to StopCommand based on which check is done to see if VM is running in HV host.
If VM is running then in this case it is not stopped and the operation bails out.
Also modified the MS code to call the StopCommand with appropriate value for the flag based on the context.
Currently it is only set to 'true' when called from the new vmsync logic based on powerstate of VM. For rest it
is set to 'false' meaning no change in behaviour.
This reduces the amount of time and storage it takes dramatically. We no longer
do a full copy, but a sparse copy. The destination image is still in RAW
format, but we only copy over used blocks.
Qemu is also better in doing this then us doing it in Java code.
Otherwise a RBDException will be thrown with the message that the snapshot
isn't protected.
modified: plugins/hypervisors/kvm/src/com/cloud/hypervisor/kvm/storage/LibvirtStorageAdaptor.java
This saves the step of writing to a temporary image in /tmp first before
writing to RBD.
This is possible due to a new version in librbd. With the rbd_default_format
setting we can now force qemu-img to create format 2 RBD images.
This is available since Ceph version 0.67.5 (Dumpling).
Add executeInVR() with timeout interface to VirtualRouterDeployer
AggregationControlCommand with Action.Finish may take longer than normal command
since it would execute all the commands in one execution, and it may result in
SSH timeout for SshHelper or other mechanism communicate with VR.
Introduce an new executeInVR() interface with added timeout period for waiting
FinishAggregationCommand to complete execution.
- get the hosts on which VPC spans given vpc id
- get the VM's in the VPC
- get the hosts on which a network spans
- get the VPC's to which a hosts is part of
- get VM's of a VPC on a hosts
introduces capability to build a physical toplogy representation of a
VPC. This json file is encapsulated in
OvsVpcPhysicalTopologyConfigCommand, and is used to send full topology
to hypervisor hosts. On hypervisor this json config can be used to setup
tunnels, configure bridge, add flow rules etc
Ovs GURU, to use different broasdcast scheme VS://vpcid.gerkey for the
networks in VPC that use distributed routing
each VIF and tunnel interface to carry the network UUID in other/options
config
2) Corrected some logging in MidoNetPublicNetworkGuru - removed .toString method call on the objects in the log body as toString is called on the object by default when use log4j
With VirtIO enabled on KVM. FreeBSD 10 supports VirtIO for both the
network and the disks. This frees us from IDE and E1000 which should
also improve performance.
By default all network disks are in RAW format. Gluster works fine with
QCOW2 which has some advantages.
Disks are by default in QCOW2 format. It is possible to run into
a mismatch, where the disk is in QCOW2 format, but QEMU gets started
with format=raw. This causes the virtual machines to lockup on boot.
Failures to start a virtual machine can be verified by checking the log
of the virtual machine, and compare the output of 'qemu-img info'.
In /var/log/libvirt/qemu/<VM>.log find the URL for the drive:
-drive file=gluster+tcp://...,format=raw,..
Compare this with the 'qemu-img info' output of the same file, mounted
under /mnt/<pool-uuid>/<img-uuid>:
# qemu-img info /mnt/<pool-uuid>/<img-uuid>
...
file format: qcow2
...
This change makes passes the format when creating a disk located on RBD
(RAW only) and Gluster (QCOW2).
Signed-off-by: Niels de Vos <ndevos@redhat.com>
The support for Gluster as Primary Storage is mostly based on the
implementation for NFS. Like NFS, libvirt can address a Gluster environment
through the 'netfs' pool-type.
PrepareForMigrationCommand, so that destination hypervisor can
mount pool. This further exposed an issue for KVM where iso
was not getting cleaned up upon successful migration, fixed as well.
By default only the Integers between -128..127 are cached (unless overridden by java.lang.Integer.IntegerCache.high system property)
If the inbound or outbound values are higher, the reference comparison won't work.
Signed-off-by: Laszlo Hornyak <laszlo.hornyak@gmail.com>
- minor resource leak cleaned up
- cpu-speed reading method extracted
- test added
- logging added in case of exception
Signed-off-by: Laszlo Hornyak <laszlo.hornyak@gmail.com>
This saves us a lot of code and libvirt is probably a better
place to do this.
libvirt-java now has the support we want, so we can now resize volumes
with libvirt.
(C)LVM volumes can't be resized using libvirt, so we have to
invoke a resize script for that.
By default the client_mount_timeout setting in librados is 300 seconds,
but that causes the connect to the Ceph cluster to block for 5 minutes
if the Ceph cluster is not available.
This patch is not ideal, but it mitigates the problem for now.
At a later point all this librados/librbd code should go back to libvirt
again, but the current versions of libvirt in the distributions are
to old for all the features we require.
For now this should prevent the CloudStack agent blocking for 5 minutes
when the Ceph cluster isn't available.
This is also tracked at the Ceph tracker: http://tracker.ceph.com/issues/6507
brought back up after being down for few hours,snapshot jobs do not get
triggered with reason "there is other active snapshot tasks on the
instance to which the volume is attached".
- the result of dividing long with long resulted in loss of precision both for network and IO
- unit tests included
Signed-off-by: Laszlo Hornyak <laszlo.hornyak@gmail.com>
replace vlanid wih broadcast uri to support vxlan to identify whether id is VLAN ID or VNI
Signed-off-by: ynojima <mail@ynojima.net>
Signed-off-by: Hugo Trippaers <htrippaers@schubergphilis.com>
Detail: getPhysicalDisk() was not matching on volumes with .raw, so
instead setting disk format to QCOW2.
BUG-ID: CLOUDSTACK-5018
Bugfix-for:
Reviewed-by:
Reported-by:
Signed-off-by: John Kinsella <jlk@stratosec.co> 1383287538 -0700
1) vxlan will use bridge scheme 'brvx-<vni>'. Multiple physical networks can host guest
traffic type with vxlan isolation, so long as they don't use the same VNI range.
2) Guest traffic labels can be physical interface if bridge by given name is not found.
Normally we take traffic label name, find the matching bridge, then resolve that to a
physical interface. Then we create guest bridges on that interface. Now we can just
specify the interface.
xs 6.1/6.2 introduce the new virtual platform, so there are two virtual platforms, windows PV driver version must match virtual platforms,
this patch tracks PV driver versions in vm details and template details.
Anthony
Detail: Checks for other Ethernet interface names uses startsWith(),
whereas the p1p1 style interface uses a regex that doesn't allow for
tailing characters, and so blocks vlan IDs. Fixed.
BUG-ID: CLOUDSTACK-4884
Bugfix-for: 4.2.1
Reviewed-by:
Reported-by:
Signed-off-by: John Kinsella <jlk@stratosec.co> 1381965250 -0700
I don't think host kernel version has any bearing on it. Original code
was tested with CentOS 6.3 and 6.4, but it seems to succeed or fail per-host,
e.g. a fast host might work and a slow host might not. I was getting intermittent
failures with ubuntu 12.04.3 prior to this patch.
These changes are a joint effort between Edison and I to refactor some
of the code around snapshotting VM volumes and creating
templates/volumes from VM volume snapshots. In general, we were working
towards allowing PrimaryDataStoreDrivers to create snapshots on primary
storage and not requiring the snapshots to be transferred to secondary
storage.
High level changes:
-Added uuid to NfsTO, SwiftTO & S3TO to cut down on the requirement of
PrimaryDataStoreTO and ImageStoreTO which don't really serve much of a
purpose
-Initial work towards enable reverting VM volume from snapshots
-Added hypervisor commands for introducing and forgetting new hypervisor
objects (snapshots, templates & volumes)
Signed-off-by: Edison Su <sudison@gmail.com>
The managed context framework provides a simple way to add logic
to ACS at the various entry points of the system. As threads are
launched and ran listeners can be registered for onEntry or onLeave
of the managed context. This framework will be used specifically
to handle DB transaction checking and setting up the CallContext.
This framework is need to transition away from ACS custom AOP to
Spring AOP.
Initial patch for VXLAN support.
Fully functional, hopefully, for GuestNetwork - AdvancedZone.
Patch Note:
in cloudstack-server
- Add isolation method VXLAN
- Add VxlanGuestNetworkGuru as plugin for VXLAN isolation
- Modify NetworkServiceImpl to handle extended vNet range for VXLAN isolation
- Add VXLAN isolation option in zoneWizard UI
in cloudstack-agent (kvm)
- Add modifyvxlan.sh script that handle bridge/vxlan interface manipulation script
-- Usage is exactly same to modifyvlan.sh
- BridgeVifDriver will call modifyvxlan.sh instead of modifyvlan.sh when VXLAN is used for isolation
Database changes:
- No change in database structure.
- VXLAN isolation uses same tables that VLAN uses to store vNet allocation status.
Known Issue and/or TODO:
- Some resource still says 'VLAN' in log even if VXLAN is used
- in UI, "Network - GuestNetworks" dosen't display VNI
-- VLAN ID field displays "N/A"
- Documentation!
Signed-off-by : Toshiaki Hatano <haeena@haeena.net>
Libvirt reports:
org.libvirt.LibvirtException: Storage volume not found: no storage vol
with matching name
in some cases, if the volume is created on one kvm host, while accessed
from other host.
It's possible due to concurrent access(read/write) storage.
The current fix is to try serveral times, and wait for 30 seconds for
each retry.
If the issue still there, then need to sync the storage pool access
CLOUDSTACK-4457:
CLOUDSTACK-4459:
harden kvm getvolume. It's possible that one volume created on other kvm host, won't show up on another host, try more times to refresh storage pool if volume won't shown up
Conflicts:
engine/storage/integration-test/test/org/apache/cloudstack/storage/test/FakeDriverTestConfiguration.java
plugins/hypervisors/kvm/src/com/cloud/hypervisor/kvm/storage/KVMStorageProcessor.java
There still exist two issues after Edison's commits.
(1) Migration from new hosts to old hosts failed.
The bridge name on old host is set to cloudVirBr* if network.bridge.name.schema is set to 3.0 in /etc/cloudstack/agent/agent.properties, but the actual bridge name is breth*-* after running cloudstack-agent-upgrade.
(2) all ports of vms (Basic zone, or Advanced zone with security groups) on old hosts are open, because the iptables rules are binding to device (bridge) name which is changed by cloudstack-agent-upgrade.
After this, the KVM upgrade steps :
a. Install 4.2 cloudstack agent on each kvm host
b. Run "cloudstack-agent-upgrade". This script will upgrade all the existing bridge name to new bridge name, and update related firewall rules.
c. install a libvirt hook:
c1. mkdir /etc/libvirt/hooks
c2. cp /usr/share/cloudstack-agent/lib/libvirtqemuhook /etc/libvirt/hooks/qemu
c3. chmod +x /etc/libvirt/hooks/qemu
c4. service libvirtd restart
c5. service cloudstack-agent restart
Signed-off-by: Wei Zhou <w.zhou@leaseweb.com>
The migrate method from libvirt supports passing down a different XML for running
the instance of the target hypervisor.
This enables the VNC to bind to the private IP Address of the hypervisor and during
migration this will be changed to the private IP address of the target host.
This way VNC doesn't listen world wide and is much safer.
Initial patch for VXLAN support.
Fully functional, hopefully, for GuestNetwork - AdvancedZone.
Patch Note:
in cloudstack-server
- Add isolation method VXLAN
- Add VxlanGuestNetworkGuru as plugin for VXLAN isolation
- Modify NetworkServiceImpl to handle extended vNet range for VXLAN isolation
- Add VXLAN isolation option in zoneWizard UI
in cloudstack-agent (kvm)
- Add modifyvxlan.sh script that handle bridge/vxlan interface manipulation script
-- Usage is exactly same to modifyvlan.sh
- BridgeVifDriver will call modifyvxlan.sh instead of modifyvlan.sh when VXLAN is used for isolation
Database changes:
- No change in database structure.
- VXLAN isolation uses same tables that VLAN uses to store vNet allocation status.
Known Issue:
- Some resource still says 'VLAN' in log even if VXLAN is used
- in UI, "Network - GuestNetworks" dosen't display VNI
-- VLAN ID field displays "N/A"
This failed due to a RAW -> QCOW2 conversion (again).
The current code still makes to much assumptions about everything always
being QCOW2 while that is not always true.
KVM - Create template from volume
Vmware - Create template from volume / Create template from snapshot
send the physical size in the copycommand which accordingly will populate template store ref and the usage_event tables with the right physical size
Signed off by : nitin mehta<nitin.mehta@citrix.com>
Although libvirt supports resizing RBD volumes (and other formats) the
Java bindings (libvirt-java) don't.
Right now we use the Java bindings for librbd to handle the resizing for us,
but in the future this should be done by libvirt rather then these
Java bindings.
Logan Barfield
punith-cloudbyte
Hugo Trippaers
Mike Tutkowski
Edison Su
Santhosh Edukulla
Wido den Hollander
Rohit Yadav
Kishan Kavala
Rajani Karuturi
Kishan Kavala
Wido den Hollander
Likitha Shetty
Anthony Xu
amoghvk
Suresh Ramamurthy
Marcus Sorensen
ynojima
Murali Reddy
Koushik Das
Kishan Kavala
Sheng Yang
Alena Prokharchyk
Kelven Yang
Jayapal
Min Chen
Niels de Vos
Laszlo Hornyak
Jayapal
Alex Huang
tuna
Wei Zhou
Harikrishna Patnala
John Kinsella
Chris Suich
Darren Shepherd
Daan Hoogland
Toshiaki Hatano
Nitin Mehta
Edison Su
JijunLiu