"OSD on PVC" doesn't work for PV backed by LV. Fixing this problem
by the following changes.
- Rook accepts LVM disk type.
- If a LV-backed device is passed, Rook/Ceph invokes
"ceph-volume lvm prepare" with "--data vg/lv"
instead of "--data /path/to/device".
- If a LV-backed device is passed, Rook/Ceph suppresses
activation/deactivation of VG that owns this LV.
Fixes: https://github.com/rook/rook/issues/4185
Signed-off-by: dulltz <isrgnoe@gmail.com>
Co-authored-by: Satoru Takeuchi <satoru.takeuchi@gmail.com>
This patch removes the need of using both tini and rook binary to run
the main osd container.
By removing this, we know can upgrade the operator image without
restarting the OSD since the osd pod does not use the rook image
anymore.
This patch only works when not running OSD on PVC since they are certain
actions that we cannot achieve with an init container.
A new init container has been introduced to do the activation of the
OSD, the activate files are placed in a empty dir mount and shared with
the main osd process.
Now the main osd process runs the ceph-osd daemon binary directly.
This currently has a few limitation:
* only works with bluestore OSDs
* only works when not running OSD on PVC
Closes: https://github.com/rook/rook/issues/4363
Signed-off-by: Sébastien Han <seb@redhat.com>
We shouldn't need to compute the location from the main OSD pod, since
the prepare pod runs on the same node we can populate the crush location
information from that pod and pass it along to the main osd container.
This is an initial commit that aims to eventually remove the need of
using the rook binary when running the main osd container.
Closes: https://github.com/rook/rook/issues/4362
Signed-off-by: Sébastien Han <seb@redhat.com>
During an orchestration if there were failures to update any OSD deployment,
the orchestration would fail and continue trying again. If the OSD(s)
continued failing, the operator would be stuck in a loop and never
respond to other CR updates.
Now the operator will log the OSD errors, if any, and complete the
orchestration. We will rely on K8s to continue to attempt to start
the OSD pods instead of the operator being stuck in the loop.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
we need to set a ENV `CSI_RBD_ENABLE_SNAPSHOTTER`
variable to "false" in operator template to disable
deployment of snapshotter sidecar in ceph-csi rbd
provisioner.
Fixes: #4401
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
We recently discovered a race condition which happens not to update the
child CRDs when CR image changes.
The race happens under the following sequence:
* orchestration onAdd() is called when the operator image restarts
or its image updated
* the orchestration then goes into processing mon/mgr/osd
* during this the cluster CR is updated so onUpdate() is called,
however, onAdd() is not done yet for child CRDs (mds, rgw)
* onUpdate() runs against the cluster CR and updates the cluster images
* by the time the child CRDs reached onUpdate() from their respective
ParentClusterChanged() methods, the image is already up to date
so we were exiting since our check for updating
these resources are based on the cluster image version
To fix this, we changed the comparison, but using isUpgrade instead and
slightly changed the way isUpgrade operates. So now, isUpgrade is true
even if the Ceph version is identical which allows us to verify whether
or not the image changed.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1775624
Signed-off-by: Sébastien Han <seb@redhat.com>
The Ceph Mimic release introduced the monitor config store which allows
us to set configuration flag per daemon in a centralized fashion. Let's
use it for Mimic and newer and keep the CLI flag of daemon startup for
older versions.
The main benefit is that the config of a daemon can now be overridden.
Also, this commit refactors some code and add methods that are unit
tested.
Closes: https://github.com/rook/rook/issues/4307
Signed-off-by: Sébastien Han <seb@redhat.com>
The Ceph Mimic release introduced the monitor config store which allows
us to set configuration flag per daemon in a centralized fashion. Let's
use it for Mimic and newer and keep the CLI flag of daemon startup for
older versions.
The main benefit is that the config of a daemon can now be overridden.
Closes: https://github.com/rook/rook/issues/4307
Signed-off-by: Sébastien Han <seb@redhat.com>
The indexes were wrongly positionned and were pointing to the previous
one instead of the one that failed to be converted to an int.
Signed-off-by: Sébastien Han <seb@redhat.com>
Checking for isUpgrade is irrelevant since we need to restart the
deployment because the the spec has changed, the only thing that might
change is whether or not we perform upgrade checks.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1775624
Signed-off-by: Sébastien Han <seb@redhat.com>
Rook now supports a build versions with a format of:
ceph version 14.2.4-64
Where -64 represents the 64th version of that package.
This is needed in order to properly detect upgrades.
Signed-off-by: Sébastien Han <seb@redhat.com>
Because of this, the mgr would be restarted if the operator image
change. This must not happen.
Added unit tests for cephMgrOrchestratorModuleEnvs() as well.
Closes: https://github.com/rook/rook/issues/4361
Signed-off-by: Sébastien Han <seb@redhat.com>
Some log messages indicate when the upgrade checks will be
performed. If the Ceph image hasn't changed, the upgrade checks
are skipped. The messages were confusing if running a Rook
upgrade.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
We create CSI resources when we create the first Ceph cluster.
This change deletes all CSI resources when we delete the last Ceph
cluster.
Partially resolves: #4234
Signed-off-by: Elise Gafford <egafford@redhat.com>
In order to remove Ceph-CSI resources after deletion of the last
Ceph cluster, the operator must be able to perform post-processing
actions on cluster deletion. This change creates a hook for the
addition of post-deletion operator callback functions.
Partially resolves: #4234
Signed-off-by: Elise Gafford <egafford@redhat.com>
In order to facilitate the deletion of CSI resources, this patch
adds ownerReferences to all CSI deployments, daemonsets, and
services. This ownerRef refers to the CSI ConfigMap in all cases.
Partially resolves#4234
Signed-off-by: Elise Gafford <egafford@redhat.com>
cephcsi v1.2.2 release is out and the image is available
for use, updating rook to use latest available release
image
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
Earlier, simultaneous drains could allow the drain detection to switch
disabled PDBs before the Ceph health had fully reflected the effects.
Signed-off-by: Rohan CJ <rohantmp@gmail.com>
With DM_DISABLE_UDEV, LVM will avoid interaction with udev.
LVM will manage the relevant nodes in /dev directly.
Signed-off-by: Sébastien Han <seb@redhat.com>
A recent PR #4172 added the ability for rook to create
the CSIDriver object on a kubernetes cluster. The change also
made it so that attachRequired value in the ceph CSI drivers
were set to false.
This leads to the entire attach skipping the AD controller
that deals with RWO attachment enforcement. Thus, if a
CephFS volume is marked RWO, 2 pods on different hosts can
end up mounting the same and consuming it at the same time.
Further for RBD, currently the image watchers are the only ones
that prevent a double mount of an image, other than the AD controller
enforcement of single attach for RWO RBD PVs. This has proven
not to be reliable in the face of MON failures, as the watcher
information is lost, and a second mount is allowed.
Thus, we need to bring back the attach requirement to leverage
kubernetes checks and balances for RWO volume type, and
prevent unwanted multi-attach use of the same.
This PR hence undoes some of the changes introduced by the
PR #4172, to bring the attacher side car back, and also declare
the attachRequired for both RBD and CephFS CSI plugins as true.
NOTE: Initially the attachRequired was set to false as this
improved attach times considerably on kubernetes, as the
entire phase was being avoided.
Signed-off-by: ShyamsundarR <srangana@redhat.com>
The app name was not being set in k8sutil's cmdreporter code, resulting
in jobs being labeled `app=` with no name. Change cmdreporter's `New`
method to validate the contents of the config struct and not the inputs
to the `New` method to help prevent an issue like this from creeping up
in the future.
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
dulltzandSatoru Takeuchi
Sébastien Han
Travis Nielsen
Kazuhito MATSUDA
Madhu Rajanna
Rohan CJ
Elise Gafford
Blaine Gardner
d-luu
ShyamsundarR
Dmitry Yusupov