When removing a mon from quorum, there is a race condition that
can result in mon quorum going being lost at least temporarily.
The mon pod was being deleted first, and then the mon removed
from quorum. If any other mon went down between the time the
pod of the bad mon was deleted and when the mon was removed from
quorum, there may not be sufficient quorum to complete the
action of removing the mon from quorum and the operator would
be stuck.
For example, there could be 4 mons temporarily due to timing
of upgrading K8s nodes where mons may be taken down for some
number of minutes. Say a new mon is started while the down
mon also comes back up. Now the operator sees it can remove
the 4th mon from quorum, so it starts to remove it. Now say
another mon goes down on another node that is being updated
or otherwise drained. Since the 4th mon pod was deleted
and another mon is down, there are only two mons remaining
in quorum, but 3 mons are required in quorum when there
are 4 mons. Therefore, the quorum is stuck until the
third mon comes back up.
The solution is to first remove the extra mon from quorum
before taking down the mon pod.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
(cherry picked from commit 8987d26a6d)
Previously, Rook would fail to create monitors for a `CephCluster` spec that
required them to be located in distinct failure domains:
```yaml
spec:
mon:
count: 3
failureDomainLabel: topology.kubernetes.io/zone
zones:
- eu-central-1a
- eu-central-1b
- eu-central-1c
```
This was due to the fact that Rook was mistakenly getting the
`failureDomainLabel` from `spec.mon.stretchCluster.failureDomainLabel`.
This commit ensures `spec.mon.failureDomainLabel` is being used instead, which
fixes the cluster deployment for the spec above.
Signed-off-by: Benoît Knecht <bknecht@protonmail.ch>
(cherry picked from commit 1c45677f5e)
adding changes to create CSI-operator config CR based on
every radosnamesapce and subvolumegroup.
Signed-off-by: subhamkrai <srai@redhat.com>
(cherry picked from commit 667e044013)
adding code changes,rbac changes required for create the new
Ceph-CSI operator CR named cephCluster in api group 'csi.ceph.io'.
Signed-off-by: subhamkrai <srai@redhat.com>
(cherry picked from commit 4b0b3a55d9)
For an advanced network configuration, the mon endpoints
expected for the ceph public network may be different from
the public IP address that is set on the K8s node. When
host networking is enabled in rook, the IP address of the
node was only allowed previously. Now, the IP address
used for a mon when scheduled on a node can be determined
by setting an annotation on each node where mons can run:
network.rook.io/mon-ip: <ip-address>
This annotation must be set before the mons are created.
Mon endpoints cannot be changed after they are already
running.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
When IPv6-only network mode is enabled and when mons are configured to
listen only on msgrV2 port, Rook has to add the port to the mon
--public-bind-addr in order to force mons to stop listening on the
msgrV1 port. This can cause problems for IPv6 clusters where adding
':<port>' to the end of an IP addr looks like a valid IPv6 address.
In IPv6 mode, Rook needs to set this to '[<ip>]:<port>' with the
brackets to be a non-ambiguous IPv6 address.
In IPv4 mode, Rook needs to keep its existing behavior. IPv4 mode should
also continue to be the defualt assumption if the network config isn't
specified.
If 'dualStack: true' is specified, Rook is informed that the environment
is dual stack, but Rook doesn't know which IP family to expect from the
pod IP. In this scenario, it is safest for Rook not to apply the port,
since assuming either IPv4 or IPv6 format could be incorrect. In this
scenario, mons may still listen on msgrV1 ports, but this is preferable
to an environment where mons perpetually crash.
Signed-off-by: Blaine Gardner <blaine.gardner@ibm.com>
Someone testing the Multus holder pod removal feature encountered an
issue where the migration process failed to lead to a system state where
PVCs could be created successfully.
The root cause was found to be a ceph csi config map wherein the primary
CephCluster entry was lacking a value for the "namespace" field.
I observed this once in my development on the holder pod removal
feature, but I was unable to reproduce and assumed it was my own error.
Since this has been seen in a user environment, it must be that the
error is a race condition, and I am unable to determine the exact source
of the bug.
I do not believe this bug would be present if the code that updates the
CSI configmap were properly idempotent, but it has many conditions based
on prior states, and I was unable to determine how to resolve this
underlying impelementation pattern issue.
Instead, I opted to separate the `clusterKey` parameter into two clear
parts:
1. `clusterID` for when `clusterKey` is used as an analogue for
`clusterID`
2. `clusterNamespace` for when `clusterKey` is used as an analogue for
`clusterNamespace`
I added unit tests to ensure that SaveClusterConfig() will be able to
detect when the namespace is currently missing, and using the new
`clusterNamespace` field, it should always know what value to use as
input when correcting the bug in already-installed systems.
I also verified that this update works when the function simultaneously
removes netNamespaceFilePath entries, and that those entries are removed
properly.
Finally, manual testing also verifies the change.
Signed-off-by: Blaine Gardner <blaine.gardner@ibm.com>
The mon was listening on both v1 and v2 ports inside the daemon
even though the service endpoint would not redirect any traffic
to the v1 port. When msgr2 is required, disable the mon from
listening on the v1 port for reduced security footprint.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
in case of portable mons, mon scaledown was skipped
with below code
```
if mon.NodeName == "" {
logger.Debugf("mon %q is not scheduled to a specific host", mon.DaemonName)
continue
}
```
which skips mon removal if mon nodeName is empty
but if mons are scale down in case of portable mons
in that case also we want to remove extra mons to match
the cephCluster configuration.
Signed-off-by: subhamkrai <srai@redhat.com>
If the spec.mon.count is 5 or more, then create
mon pdbs with maxUnavailable count as 2. We can
afford to have more than 1 mons down while
draining the nodes in this case.
Signed-off-by: sp98 <sapillai@redhat.com>
Reduce CRD size by removing some descriptions for things that
(a) are long, (b) are repeated, and (c) are commong kubernetes types.
Remove the `Status` field from VolumeClaimTemplates (a.k.a. PVCs) in
types.go to remove all unnecessary status-related fields.
Use `yq` to go through generated `crds.yaml` to strip out descriptions
for well-known k8s types that are long and repeated where we can't
remove them from types.go. This is mostly placement descriptions and
volumeSource descriptions.
The overall `crds.yaml` file is reduced to less than 60% of its original
size.
Signed-off-by: Blaine Gardner <blaine.gardner@ibm.com>
Signed-off-by: Redouane Kachach <rkachach@redhat.com>
Co-authored-by: Blaine Gardner <blaine.gardner@ibm.com>
This commit uses an csi config map API exposed by Ceph-CSI.
This will reduce the efforts of keeping the configuration in
sync with Ceph-CSI.
Signed-off-by: Praveen M <m.praveen@ibm.com>
During certain maintenance tasks the admin will own running
operations on the ceph mgr, rgw, mds and rbd-mirror daemons
and the operator should not interfere with those operations.
Co-authored-by: gauravsitlani <gaurav.sitlani@live.com>
Signed-off-by: subhamkrai <srai@redhat.com>
Change how Rook detects network CIDRs for Multus networks. The IPAM
configuration is only defined as an arbitrary string JSON blob with a
"type" field and nothing more. Rook's detection of CIDRs for whereabouts
had already grown out of date since the initial implementation.
Additionally, Rook did not support DHCP IPAM, which is a reasonable
choice for users. And more, Rook did not support CNI plugin chaining,
which further complicates NADs. Based on the CNI spec, network chaning
can result in any changes to network CIDRs from the first-given plugin.
All these problems make it more and more difficult for Rook to support
Multus by inspecting the NAD itself to predict network CIDRs. Instead,
it is better for Rook to treat the CNI process as a black box. To
preserve legacy functionality of auto-detecting networks and to make
that as robust as possible, change to a canary-style architecture like
that used for Ceph mons, from which Rook will detect the network CIDRs
if possible.
Also allow users to specify overrides for CIDR ranges. This allows Rook
to still support esoteric and unexpected NAD or network configurations
where a CIDR range is not detectable or where the range detected would
be incomplete. Because it may be impossible for Rook to understand the
network CIDRs wholistically while residing only on a portion of the
network, this feature should have been present from Multus's inception.
Improving CIDR auto-detection and allowing users to specify overrides
for auto-detected CIDRs rounds out Rook's Multus support for CephCluster
(core/RADOS) installations. No further architectural changes should be
needed for CephClusters as regards application of public/cluster network
CIDRs for Multus networks.
Signed-off-by: Blaine Gardner <blaine.gardner@ibm.com>
We need to restart all the ceph daemons whenever
cephCluster network settings are modified like
requiremsgr2, encryption and compression. This
required for Ceph to consider the new settings
it require new ceph daemons all over.
Signed-off-by: subhamkrai <srai@redhat.com>
* fixes mon zonal deployment of 2 per zone to 1 if zones field is
specified.
* adds zones field doc to CRDs Documentation.
* add code for picking up failureDomainLabel if specified for zones
field in mon deployment
Signed-off-by: Deepika Upadhyay <deepika@koor.tech>
The function ParseMonEndpoints existed twice, once in the mon package
and once in the controller package. It has now been deleted from the
mon package. Usages in the mon package now reference the function
in the controller package.
Signed-off-by: Henry Davies <henrydavies@hotmail.co.uk>
When msgr2 is required, ensure the mon endpoints passed to
the csi configmap are on port 3300. The mons will be
listening on both 6789 and 3300. Existing volumes can
continue using port 6789 while new volumes will use
port 3300.
Signed-off-by: travisn <tnielsen@redhat.com>
Mons that are configured on msgr1 will be using the v1 port
in the csi driver since only one port is allowed in the configmap.
To allow mons to be started on v2 while keeping quorum with the
v1 mons, the v1 mons must continue running on the v1 port.
So even if msgr2 is required, allow existing v1 mons to continue
running on that port.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
Environment variables are not recommended for secrets in pods since
they can be easily leaked if the environemnt variables are logged.
By mounting the mon secret as a file, the mgr and osd prepare pods
can read the mon secret from a file for better security.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The fsid, username, and user secrets were from legacy and no
longer needed on the osd daemon or mon pods. This removes the
unnecessary env vars from the pod specs.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
The mons that are out of quorum may cause ceph commands to timeout
or fail unnecessarily trying to connect to a mon that is no longer
online. Now the mon health check will update the mon endpoints configmap
when a mon is detected out of quorum. This also means that if the
operator is restarted during a mon failover, the failed mon will no
longer remain in the ceph.conf, thus allowing the quorum to be more
likely to respond to the mons that are still in quorum.
The update for mons out of quorum only applies if other mons are in
quorum. If quorum is down, the configmap will not keep track of the offline
mons since too many are offline.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
If the host network setting is changed in the cluster CR,
the existing mons must continue using the same network
settings or else the operator would update their pod
specs with the incorrect settings and cause mon quorum
to go down. Now the network setting is preserved so
the admin could switch between host and non-host network
configuration without reinstalling.
Signed-off-by: Travis Nielsen <tnielsen@redhat.com>
Introduce a new env variable ROOK_CSI_IMAGE_PULL_POLICY in rook operator configmap which should be used to
customize the imagePullPolicy for the csi driver and imagePullPolicy property in cephVersionSpec for ceph pods.
Signed-off-by: Avan Thakkar <athakkar@redhat.com>
if Multus is enabled the clusterinfo should be updated with
network as multus as to run the ceph cmds in remote
executor
Signed-off-by: parth-gr <paarora@redhat.com>
Travis Nielsen
Benoît Knecht
subhamkrai
Blaine Gardner
Satoru Takeuchi
piao
parth-gr
sp98
Redouane KachachandBlaine Gardner
Praveen M
dependabot[bot]
gauravsitlani
guoguangwu
Deepika Upadhyay
Henry Davies
parth-grandBlaine Gardner
Avan Thakkar