After a few hours running with InfluxDB configured, CloudStack hangs due to OutOfMemoryException raised. The exception happens at com.cloud.server.StatsCollector.writeBatches(StatsCollector.java:1510):
2020-08-12 21:19:00,972 ERROR [c.c.s.StatsCollector] (StatsCollector-6:ctx-0a4cfe6a) (logid:03a7ba48) Error trying to retrieve host stats
java.lang.OutOfMemoryError: unable to create new native thread
...
at org.influxdb.impl.BatchProcessor.<init>(BatchProcessor.java:294)
at org.influxdb.impl.BatchProcessor$Builder.build(BatchProcessor.java:201)
at org.influxdb.impl.InfluxDBImpl.enableBatch(InfluxDBImpl.java:311)
at com.cloud.server.StatsCollector.writeBatches(StatsCollector.java:1510)
at com.cloud.server.StatsCollector$AbstractStatsCollector.sendMetricsToInfluxdb(StatsCollector.java:1351)
at com.cloud.server.StatsCollector$HostCollector.runInContext(StatsCollector.java:522)
Context on InfluxDB Batch: Enabling batch on InfluxDB is great and speeds writing but it requires caution to avoid Zombie threads.
Solution: This happens because the batching feature creates an internal thread pool that needs to be shut down explicitly; therefore, it is important to add: influxDB.close().
When the static route service is not available on the VPC and a static route is created, the static route is created in a revoked state.
Currently, the UI doesn't distinguish between active or revoked static routes.
This PR adds the missing state filter to the list routes command and only lists active routes in the UI.
It also ignores revoked routes when the private gateway is being removed but clears out the inactive routes before the gateway is removed.
Fixes#2908
This will purge all the cookies on logout including multiple sessionkey
cookies if passed. On login, this will restrict sessionkey cookie
(httponly) to the / path.
Fixes#4136
Co-authored-by: Pearl Dsilva <pearl.dsilva@shapeblue.com>
While migrate a vm, in the popup, the host dedicated to other accounts/domains are also 'Suitable" for migration, which is obviously wrong.
The same issue happens with api findHostsForMigration
If we resize a volume of a vm running on a host which is not Up or not Enable, the job will be scheduled to another normal host. Then the volume will be resized by "qemu-img resize" instead of "virsh blockresize", the image might be corrupted after resize.
Adding missing fields in the following APIs
osdisplayname in listVirtualMachines
vpcofferingname in listVpcs
vpcname in listPublicIpAddresses
vpcname in listPrivateGateways
vpcname in listVpnGateways
templatename, podname in listRouters
templatename, podname in listSystemVms
Fixes: #4161
Fixes wrong count in listAffinityGroup API.
API was returning the count of AffinityGroupJoinVO records.
Signed-off-by: Abhishek Kumar <abhishek.mrt22@gmail.com>
This PR fixes an issue where an instance fails to deploy due to a null pointer when using an L2 Guest Network with DefaultL2NetworkOfferingConfigDrive on Xenserver. It also fixes migrating an instance to another host.
This has been tested by:
- Creating an L2 Guest network, using DefaultL2NetworkOfferingConfigDrive as the network offering.
- Deploying an instance using the L2 Guest network created.
- Migrating the instance away from the host and back
When a VPC is restarted, the networks in the VPC is not restarted, this PR will add the logic to restart the networks in the VPC that needs a restart when the VPC is restarted.
Fixes#3816
Root cause:
Even though dynamic scaling job is handled in vmworkjob queue which ensures serilizing multiple jobs but the database updating and generating usage events are out of the job queue.
Solution:
Moved all updations into the job queue
Firstly I have tested all the scenarios to check if nothing is broken:
Scaling on a running VM with normal compute offering
Scaling on a stopped VM with normal compute offering
Scaling on a running VM with custom compute offering
Scaling on stopped VM with custom compute offering
Scaling on stopped/running VM between custom compute offering and normal compute offering and combinations among these. Checked if the custom parameters have been populated or deleted accordingly based on the offering to which the VM is scaled
Since this is a corner scenario I could not test the exact point where two usage events are recorded at the same time for two different API calls on same VM.
Though VMware does not support security groups, but in a basic zone with VMware and no isolation VMs should be able to deploy.
Root cause:
In case of VMware and basic zone control nic is set to 0.0.0.0 assuming control network will be shared with guest network.
But to have access to VMware instances management/private needs to be assigned to it.
Solution:
Assing a private ip even in case of basic zone VMware.
The listZonesMetrics does not return same keys are listZones as the
default response view is restricted. This fixes that by ensuring that
for root admin full response view is used.
Signed-off-by: Rohit Yadav <rohit.yadav@shapeblue.com>
As previously described by PR #3929:
If vm has attached ISO, the migration fails with error message "org.libvirt.LibvirtException: Cannot access storage file /mnt/b33e5a1d-e4ea-3465-b6ac-c98dc8ff8af0/207-2-cc5fd717-2d57-3bb3-bcf6-2c930268db6c.iso"
in 4.13, list sshkeypairs with keyword will ignore the search by name if name is specifed
Fixes an issue in #3098
for example,
(local) > list sshkeypairs name=wei keyword=wei filter=name
{
"count": 3,
"sshkeypair": [
{
"name": "wei3"
},
{
"name": "wei2"
},
{
"name": "wei"
}
]
}
with this patch ,it gives correct result.
(local) > list sshkeypairs name=wei keyword=wei filter=name
{
"count": 1,
"sshkeypair": [
{
"name": "wei"
}
]
}
When both routers of VPC is in MASTER state
then multiple alerts are sent equally to the number of tiers in the VPC.
If the VPC has 3 tiers then 6 alerts will be sent. This is not good
if VPC has more than 10 networks in it.
Instead of checking the router status for all the tiers in the VPC,
just check the status of the router for one tier in a VPC so that
multiple duplicate alerts can be avoided
Currently, the cloudstack sends VM password only to the first
router in the network even if its the backup and return the result.
In some cases the first router will be back up and the second will be master.
Since password server is not running in backup, when the user resets the password,
it is sent to the first router which can be backup.
In that case, the new password is not stored in the password server and users cant log in with a new password.
This change ensures that we send the password to both the routers instead
of the first router so that a new password is stored in the master router.
When we add new guest os, sometimes we missed the records in guest_os_hypervisor.
However, the guest disk model (virtio/ide) is determined by record in the table.
It causes the issue that some new guest os(eg Debian 8/9) uses e1000 instead of virtio nic, and ide disk instead of virtio disk.
To fix the issue permanantly, pass the guest os name in guest_os if the record for kvm is not found in guest_os_hypervisor.
Related commit:7ac9f00eeeb4cd37ec39efeba066e799b581b1a0
When the state of the site to site vpn changes, the check
is done on all the virtual routers including the internal
load balancing vm as well. It is not needed to check the
state for internal load balancing vm
This implements the systemvm list API response creator to find and use
the host record for a ssvm/cpvm to get the agent status and other
details like last disconnected date and agent version.
Fixes 3875
Signed-off-by: Rohit Yadav <rohit.yadav@shapeblue.com>
This makes the listSystemVms API to return the host status (agent state),
version and last pinged information. This makes it possible for UIs
to call a single API to get this information.
After a local template is uploaded via browser, the generated usage event with type = "TEMPLATE.CREATE" is persisted with the data store ID instead of the zone ID on the zone_id column. The fix will refactor the upload monitor logic, as after the upload completes, it sets the datastore ID on the zone ID column for the created "TEMPLATE.CREATE" usage event. This refactor will query the DB for the data store and will set its associated zone ID in the usage field.
The fix produces the same behaviour as when registering a template from URL.
FIx is also for uploading VOLUME from local/via browser.
Fixes#3783
As reported in the issue, creating volumes from pure snapshot fails with NPE. This is due to order of calls where disk offering access is checked before checking disk offering value. This PR fixes the same.
Signed-off-by: Abhishek Kumar <abhishek.mrt22@gmail.com>
Fixes#3191
When a template is registered, code stores md5sum of the downloaded file in the vm_template table. However, this downloaded file could be deleted after template installation if it is not an actual (.qcow2, .ova, etc.) file. When the user copies a template using copyTemplate API, the actual template file will be copied across the image stores. Matching checksum for the copied templated file and the stored value from the vm_template table will result in a mismatch.
Changes will set an empty checksum value for the copied template while passing to download service which allows skipping wrong checksum check for the copied while install.
However, this results in a change in checksum value for concerned template entry in vm_template table post template install.
Co-authored-by: dahn <daan.hoogland@gmail.com>
The List Management Server api returns a list of all the management servers but fails when trying to list by id or name. This ensures that it fetches the details as per the parameters passed.
Fixes: #3833
The metrics API has few properties missing that are present in the corresponding resource.
Fixes#3831
Signed-off-by: Rohit Yadav <rohit.yadav@shapeblue.com>
Co-authored-by: Rohit Yadav <rohit@apache.org>
Steps to reproduce the issue
(1) create a custom service offering
(2) create a vm with the offering
(3) update vm with displayvm=false, returns an error
(local) > update virtualmachine id=f33fd06a-7643-40d1-833f-272845d9ba09 displayvm=false
Error 530: {"updatevirtualmachineresponse":{"uuidList":[],"errorcode":530,"cserrorcode":9999}}
When start a vm or migrate a vm (away from a host in host maintenance), cloudstack will check capacity of all hosts and choose one. If there are hundreds of hosts on the platform, it will take some seconds. When cloudstack choose a host and start/migrate vm to it, the resource consumption of the host might have been changed. This normally happens when we start/migrate multiple vms.
It would be better to double check the host capacity when start vm on a host.
This PR includes the fix for cpucore capacity when start/migrate a vm.
When we calculate a resource consumption of a host, we need to take the vms in following states into calculation: Running, Starting, Stopping, Migrating (to the host), and vms are Migrating from the host. Because, when stop a vm, the resource on host will be released when vm is stopped. When migrate a vm, the resource on destination host will be increased before migration starts, and resource on source host will be decreased after migraiton succeeds.
In cloudstack, there is a task named CapacityChecked which run every 5 minutes (capacity.check.period =300000 ms by default). It recalculates capacity of all hosts. However, it takes only vms in Running and Starting into consideration. We have faced some issues in host maintenance due to it.
Steps to reproduce the issue
(1) migrate N vms from host A to host B, cpu/ram resource increases before the migration.
(2) capacity check recalculate the capacity of hosts. used capacity of Host B will be reset to original value (not including the vms in Migrating).
(3) migrate some more vms from other host to host B, the migrations are allowed by cloudstack (because used capacity is incorrect). If the actual used memory exceed the physical memory on the host, there might be some critical issues (for example, libvirt dies)
