The container runtime CRI-O limits the number of PIDs to 1024 by
default. When many PVCs are requested at the same time, it is possible
for the provisioner to start too many threads (or go routines) and
executing 'rbd' commands can start to fail. In case a go routine can not
get started, the process panics.
The PID limit can be changed by passing an argument to kubelet, but this
will affect all pids running on a host. Changing the parameters to
kubelet is also not a very elegant solution.
Instead, the provisioner pod can change the configuration itself. The
pod is running in privileged mode and can write to /sys/fs/cgroup where
the limit is configured.
With this change, the limit is configured to 'max', just as if there is
no limit at all. The logs of the csi-rbdplugin in the provisioner pod
will reflect the change it makes when starting the service:
$ oc -n rook-ceph logs -c csi-rbdplugin csi-rbdplugin-provisioner-0
..
I0726 13:59:19.737678 1 cephcsi.go:127] Initial PID limit is set to 1024
I0726 13:59:19.737746 1 cephcsi.go:136] Reconfigured PID limit to -1 (max)
..
It is possible to pass a different limit on the commandline of the
cephcsi executable. The following flag has been added:
--pidlimit=<int> the PID limit to configure through cgroups
This accepts special values -1 (max) and 0 (default, do not
reconfigure). Other integers will be the limit that gets configured in
cgroups.
Signed-off-by: Niels de Vos <ndevos@redhat.com>
This change also starts mapping nbd based access using ther rbd CLI
as, it is a prerequisite to get device listing for nbd as well.
Signed-off-by: ShyamsundarR <srangana@redhat.com>
Use Deployment with leader election instead of StatefulSet
Deployment behaves better when a node gets disconnected
from the rest of the cluster - new provisioner leader
is elected in ~15 seconds, while it may take up to
5 minutes for StatefulSet to start a new replica.
Refer: kubernetes-csi/external-provisioner@52d1fbc
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
Every Ceph CLI that is invoked at present passes the key via the
--key option, and hence is exposed to key being displayed on
the host using a ps command or such means.
This commit addresses this issue by stashing the key in a tmp
file, which is again created on a tmpfs (or empty dir backed by
memory). Further using such tmp files as arguments to the --keyfile
option for every CLI that is invoked.
This prevents the key from being visible as part of the argument list
of the invoked program on the system.
Fixes: #318
Signed-off-by: ShyamsundarR <srangana@redhat.com>
in NodeStage RPC call we have to map the
device to the node plugin and make sure the
the device will be mounted to the global path
in nodeUnstage request unmount the device from
global path and unmap the device
if the volume mode is block we will be creating
a file inside a stageTargetPath and it will be
considered as the global path
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
As detailed in issue #279, current lock scheme has hash
buckets that are count of CPUs. This causes a lot of contention
when parallel requests are made to the CSI plugin. To reduce
lock contention, this commit introduces granular locks per
identifier.
The commit also changes the timeout for gRPC requests to Create
and Delete volumes, as the current timeout is 10s (kubernetes
documentation says 15s but code defaults are 10s). A virtual
setup takes about 12-15s to complete a request at times, that leads
to unwanted retries of the same request, hence the increased
timeout to enable operation completion with minimal retries.
Tests to create PVCs before and after these changes look like so,
Before:
Default master code + sidecar provisioner --timeout option set
to 30 seconds
20 PVCs
Creation: 3 runs, 396/391/400 seconds
Deletion: 3 runs, 218/271/118 seconds
- Once was stalled for more than 8 minutes and cancelled the run
After:
Current commit + sidecar provisioner --timeout option set to 30 sec
20 PVCs
Creation: 3 runs, 42/59/65 seconds
Deletion: 3 runs, 32/32/31 seconds
Fixes: #279
Signed-off-by: ShyamsundarR <srangana@redhat.com>
Deployment behaves better when a node gets disconnected from the rest of
the cluster - new provisioner leader is elected in ~15 seconds, while
it may take up to 5 minutes for StatefulSet to start a new replica.
Refer: 52d1fbcf9dFixes: #335
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
This is a part of the stateless set of commits for CephCSI.
This commit removes the dependency on config maps to store cephFS provisioned
volumes, and instead relies on RADOS based objects and keys, and required
CSI VolumeID encoding to detect the provisioned volumes.
Changes:
- Provide backward compatibility to provisioned volumes by older plugin versions (1.0.0 or older)
- Remove Create/Delete support for statically provisioned volumes (fixes#382)
- Added namespace support to RADOS OMaps and used the same to store RADOS CSI objects and keys in the CephFS metadata pool
- Added support to mention fsname for CephFS provisioning (fixes#359)
- Changed field name in CSI Identifier to 'location', to denote a pool or fscid
- Updated mounter cache to use new scheme
- Required Helm manifests are updated
- Required documentation and other manifests are updated
- Made driver option 'metadatastorage' as optional, as fresh installs do not need to specify the same
Testing done:
- Create/Mount/Delete PVC
- Create/Delete 5 PVCs
- Mount version 1.0.0 PVC
- Delete version 1.0.0 PV
- Mount Statically defined PV/PVC/Pod
- Mount Statically defined version 1.0.0 PV/PVC/Pod
- Delete Statically defined version 1.0.0 PV/PVC/Pod
- Node restart when mounted to test mountcache
- Use InstanceID other than 'default'
- RBD basic round of tests, as namespace is added to OMaps
- csitest against ceph-fs plugin
- NOTE: CephFS plugin still does not detect and address already created
volumes but of a different size
- Test not providing any value to the metadata storage parameter
Signed-off-by: ShyamsundarR <srangana@redhat.com>
currently, we have 3 docker files(cephcsi,rbd,cephfs) in the ceph-csi repo.
[commit ](85e121ebfe)
added by John to build a single image which can act as rbd or
cephfs based on the input configuration.
This PR updates the makefile and kubernetes templates to use
the unified image and also its deletes the other two dockerfiles.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
Existing config maps are now replaced with rados omaps that help
store information regarding the requested volume names and the rbd
image names backing the same.
Further to detect cluster, pool and which image a volume ID refers
to, changes to volume ID encoding has been done as per provided
design specification in the stateless ceph-csi proposal.
Additional changes and updates,
- Updated documentation
- Updated manifests
- Updated Helm chart
- Addressed a few csi-test failures
Signed-off-by: ShyamsundarR <srangana@redhat.com>
The kubernetes manifests and Helm templates have been updated to use
aggregated ClusterRoles. The same change has been done in Rook as well.
Refer rook/rook#2634 and rook/rook#2975
Signed-off-by: Kaushal M <kshlmster@gmail.com>
PR #290 missed the update permission to persistentvolumes.
Without that permission, you will get the following error when attaching a RBD volume to a pod:
```
Warning FailedAttachVolume 100s (x11 over 7m52s) attachdetach-controller AttachVolume.Attach failed for volume "pvc-d23f8745-60bb-11e9-bd35-5254001c78d6" : could not add PersistentVolume finalizer: persistentvolumes "pvc-d23f8745-60bb-11e9-bd35-5254001c78d6" is forbidden: User "system:serviceaccount:kube-system:rbd-csi-provisioner" cannot update resource "persistentvolumes" in API group "" at the cluster scope
```
if attacher is not enabled, we need to
create the csidriver CRD with spec
to make attachRequired as false to
skip volume attach check in kube.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
adding the condition will help us
to easily remove the attacher later.
or even we can add else condition
if we have an alternate to attacher.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
currently we are deploying external-attacher
as a seperate statefulset, which leads to
attacher communicating with the node provisoner
daemonset, This PR deploys external-attacher
as a sidecar container inside provisioner
statefulset, so that external-provisioner
always communicates with the plugin responsible
for the provision controller capcabilities.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
Based on the review comments addressed the following,
- Moved away from having to update the pod with volumes
when a new Ceph cluster is added for provisioning via the
CSI driver
- The above now used k8s APIs to fetch secrets
- TBD: Need to add a watch mechanisim such that these
secrets can be cached and updated when changed
- Folded the Cephc configuration and ID/key config map
and secrets into a single secret
- Provided the ability to read the same config via mapped
or created files within the pod
Tests:
- Ran PV creation/deletion/attach/use using new scheme
StorageClass
- Ran PV creation/deletion/attach/use using older scheme
to ensure nothing is broken
- Did not execute snapshot related tests
Signed-off-by: ShyamsundarR <srangana@redhat.com>
as the socket directory will be created
inside the container no need to follow
the plugin name in for the directory
creation, this will also reduce the code
changes if we want to change driver name.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
remove unwanted rules and update
rbac to have permission to modify
endpoints and configmaps in the
current namespace.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
remove unwanted rules and update
rbac to have permission to modify
endpoints and configmaps in the
current namespace.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
