Make the Ceph operator more cautious about when it decides to remove
nodes from the Rook-Ceph cluster which are acting as osd hosts.
When `useAllNodes` is set to `true` we assume that the user wants to
have the most hands-off experience. Node removals are allowed when a
node is delted from Kubernetes and when a node has its taints/affinities
modified by the user (but not by automatic k8s modification as much as
possible).
When `useAllnodes` is set to `false` the only time a node is removed is
if it is removed from the Ceph cluster definition.
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
The basic idea here is to replace the release name in the ClusterInfo
object with the fine grained version information queried at runtime.
This patch also removes the Name field from the cluster spec, which was
also runtime determined and was redundant with ClusterInfo relase name.
Signed-off-by: Noah Watkins <noahwatkins@gmail.com>
If someone sets limits to pod, we want to ensure the possible
experience, so we want to make sure that people do not configure
inapropriate values for certain daemons.
We decide to fail if the memory.limit is too low.
Signed-off-by: Sébastien Han <seb@redhat.com>
If the spec of the OSD daemon contains resource limits, especially
memory.limit we will use that value for setting the osd memory target
value so the bluestore OSD knows the amount of memory available for its
cache.
Signed-off-by: Sébastien Han <seb@redhat.com>
The mgr orchestrator modules need the container image to run
the ceph image for blinking the lights when a disk is down
Signed-off-by: travisn <tnielsen@redhat.com>
this is needed as cryptsetup in the pod and udev on the host coordinate
via a semaphore that must be visible in both contexts.
note that additionally the host needs to have dmcrypt compiled with udev
synchronization, which at least one minikube version does not have.
fixes: #923
Signed-off-by: Noah Watkins <noahwatkins@gmail.com>
Configure the Ceph mds daemon completely from the operator a la the
recent changes to the Ceph mon and mgr operators.
Create the mds deployments first and then
create the keyring secrets for them with their owner reference as the
corresponding deployment. This will mean that the secrets do not need to
be micromanaged. When the deployment is deleted, the secret is also
deleted. This has not been necessary for the mons or the manager since
the mons share a keyring with a lifespan of the cluster, as does the
mgr, which currently has single-mgr support only.
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
All devices detected by the discovery pod were being passed to the OSD provisioning pod
thus not always honoring the desired device list that should be provisioned.
Now the provisioning pod will be given the desired state from the crd,
then apply that state depending on the actual devices detected.
Also added a helper to ensure OSDsPerDevice is always valid.
Signed-off-by: travisn <tnielsen@redhat.com>
This commit introduces the necessary changes to support the new
messenger feature coming with Ceph Nautilus (currently in development).
What changes? Now the monitor listens on two port:
* old 6789 for messenger v1, which will help us support older client
(e,g: krbd)
* new 3300 for messengers v2, which brings new improvement in the
messaging layer. This new transport layer brings numerous advantages
such as encryption improvement, speed improvement, pluggable nature to
support different network stack than TCP and many more.
We still have one Service IP, however it has 2 ports, see:
```
NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE
rook-ceph-mon-a ClusterIP 10.106.217.160 <none> 3300/TCP,6789/TCP 4h
rook-ceph-mon-b ClusterIP 10.99.36.175 <none> 3300/TCP,6789/TCP 4h
rook-ceph-mon-c ClusterIP 10.108.220.74 <none> 3300/TCP,6789/TCP 4h
````
The `ceph.conf` has changed and we don't force the port when using an IP
address (public addr etc). Ceph, depending on its version will naturally
start the monitors on their right port, 6789.
A new --ceph-version-name CLI argument has been added to the Rook binary
so that when the pod starts it passes the ceph version name and the
configuration of the ceph.conf, as well as daemon startup flags, happen
properly.
Given that the Rook Operator remembers the port of all the monitors it
deployed (through Pod definition), this change is not an issue and will
maintain backward compatibility.
Note that to test this you must build rook with dev container image,
which contains the dev Nautilus version. So you should do something
like:
`make -j4 BASEIMAGE='ceph/daemon-base:latest-master' IMAGES='ceph' build`
Resolves: #2525
Signed-off-by: Sébastien Han <seb@redhat.com>
Rook's behavior should not be to create a default OSD in dataDirHostPath
when no devices are present on a node. Any preexisting fallback osd
should be kept as long as no dirs or disks have been specified to keep
the legacy behavior for those clusters running with these osds in place.
The legacy deletion behavior is also kept, removing the osd as soon as
any dir or device is specified other than the dataDirHostPath dir.
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
Ceph expects to be able to reason about physical hardware / nodes.
Before this patch the NODE_NAME that was available referred to the pod
name. This patch exposes the underlying node name, pod, and pod
namespace for the ceph daemons to consume.
fixes: #2078
Signed-off-by: Noah Watkins <nwatkins@redhat.com>
Rook no longer supports Kubernetes < 1.7
Remove a workaround for k8s < 1.6.4 where `runAsNonRoot` could not be
explicitly set to false.
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
We don't need to handle other osd(s) related tasks if there's
no valid nodes available in the cluster to run osd(s), log a
warning message to help user identify the cause of the issue.
Signed-off-by: Dennis Chen <dennis.chen@arm.com>
Fix some basic spellcheck errors. Also remove trailing spaces and make
sure files have a newline (my editor does automatically).
Signed-off-by: Blaine Gardner <blaine.gardner@suse.com>
This fixes issues with a new OSD Pod in the same Deployment coming up and
directly entering CrashLoopBackOff because the old OSD Pod hasn't been fully
terminated yet (released it's locks).
Signed-off-by: Alexander Trost <galexrt@googlemail.com>
Previously, if this failed mid-way it would then throw
an exception at the first operation when retrying (because it's
trying to reweight an OSD that might not be in the map any more)
Signed-off-by: John Spray <john.spray@redhat.com>