Dimitri Savineau [Tue, 17 Aug 2021 15:27:57 +0000 (11:27 -0400)]
ceph-dashboard: fix oject gateway integration
Since [1] multiple ceph dashboard commands have been removed and this is
breaking the current ceph-ansible dashboard with RGW automation.
This removes the following dashboard rgw commands:
Dimitri Savineau [Thu, 12 Aug 2021 15:08:27 +0000 (11:08 -0400)]
ceph-volume: hide OSD keyring during creation
When using ceph-volume lvm create/prepare/batch then the keyring of each
OSD created is displayed in the output.
Let's replace those by some '*' chars.
With OpenSSL version prior 1.1.1 (like CentOS 7 with 1.0.2k), the -addext
doesn't exist.
As a solution, this uses the default openssl.cnf configuration file as a
template and add the subjectAltName in the v3_ca section. This temp openssl
configuration file is removed after the TLS certificate creation.
This patch also move the run_once statement at the block level.
infra: use dedicated variables for balancer status
The balancer status is registered during the cephadm-adopt, rolling_update
and swith2container playbooks. But it is also used in the ceph-handler role
which is included in those playbooks too.
Even if the ceph-handler tasks are skipped for rolling_update and
switch2container, the balancer_status variable is erased with the skip task
result.
Teoman ONAY [Tue, 3 Aug 2021 14:06:53 +0000 (16:06 +0200)]
podman pids.max default value is 2048, docker's one is 4096 which are
sufficient for the default value (512) of rgw thread pool size.
But if its value is increased near to the pids-limit value,
it does not leave place for the other processes to spawn and run within
the container and the container crashes.
pids-limit set to unlimited regardless of the container engine.
When the ceph dashboard user creation fails then the issue is hidden
as we don't check the return code and don't print the error message
in the module output.
This ends up with a failure on the ceph dashboard set roles command saying
that the user doesn't exist.
By failing on the user creation, we will have an explicit explaination of
the issue (like weak password).
infrastructure-playbooks: Get Ceph info in check mode
In the `set osd flags` block, run the Ceph commands that gather information
from the cluster (and don't make any changes to it) even when running in check
mode.
This allows the tasks that depend on the variables set by those tasks to
succeed in check mode.
When the rgw_multisite_proto variable is set to https then we shoudn't use
the IP address in the zone endpoints list but the node FQDN to match the
TLS certificate CN.
This refactor merges the two playbooks so we only have to maintain 1
playbook.
(Symlink the old purge-container-cluster.yml playbook for backward
compatibility).
1303611 introduced tasks for disabling the pg_autoscaler on pools and
the balancer but thoses tasks are already executed on the first monitor
node so we don't need to add the run_once statement.
Populating the ceph_mgr_modules list in the mgr_modules doesn't make sense
since that file is only executed if the list isn't empty or we're using the
dashboard.
We already have config override variables for existing block (like
ganesha_ceph_export_overrides, ganesha_log_overrides, etc...) or a
global one (ganesha_conf_overrides) but redefining the NFS_CORE_PARAM
block in that variable will erase all previous values (currently only
Bind_Addr).
If one a the monitor is out of the quorum then nothing prevents the upgrade
playbook to run.
We only check if we have at least three monitor nodes but we should also
check if those monitor nodes are correctly present in the quorum.
dashboard: remove "certificate is valid for" error
When deploying dashboard with ssl certificates generated by
ceph-ansible, we enforce the CN to 'ceph-dashboard' which can makes
application such alertmanager complain like following:
`err="Post https://mgr0:8443/api/prometheus_receiver: x509: certificate is valid for ceph-dashboard, not mgr0" context_err="context deadline exceeded"`
The idea here is to add alternative names matching all mgr/mon instances
in the certificate so this error won't appear in logs.
The dashboard/monitoring stack can be deployed via the dashboard_enabled
variable. But there's nothing similar if we can to remove that part only
and keep the ceph cluster up and running.
The current purge playbooks remove everything.
Add any_errors_fatal: true in cephadm-adopt playbook.
We should stop the playbook execution when a task throws an error.
Otherwise it can lead to unexpected behavior.
Dimitri Savineau [Wed, 16 Dec 2020 19:18:08 +0000 (14:18 -0500)]
ceph-facts: move device facts to its own file
Instead of reusing the condition 'inventory_hostname in groups[osds]'
on each device facts tasks then we can move all the tasks into a
dedicated file and set the condition on the import_tasks statement.
Dimitri Savineau [Tue, 15 Dec 2020 22:34:34 +0000 (17:34 -0500)]
ceph-validate: check logical volumes
We currently don't check if the logical volume used in lvm_volumes list
for either bluestore data/db/wal or filestore data/journal exist.
We're only doing this on raw devices for batch scenario.
Dimitri Savineau [Tue, 15 Dec 2020 20:08:00 +0000 (15:08 -0500)]
ceph-validate: check db/journal/wal devices too
When using dedicated devices for db/journal/wal objecstore with
ceph-volume lvm batch then we should also validate that those devices
exist and don't use a gpt partition table in addition of the devices
and lvm_volume.data variables.
Dimitri Savineau [Tue, 15 Dec 2020 19:49:57 +0000 (14:49 -0500)]
ceph-validate: check devices from lvm_volumes
2888c08 introduced a regression as the check_devices tasks file was
only included based on the devices variable.
But that file also validate some devices from the lvm_volumes variable.
If multi-realms were deployed with several instances belonging to the same
realm and zone using the same port on different nodes, the service id
expected by cephadm will be the same and therefore only one service will
be deployed. We need to create a service called
`<node>.<realm>.<zone>.<port>` to be sure the service name will be unique
and well deployed on the expected node in order to preserve backward
compatibility with the rgws instances that were deployed with
ceph-ansible.
When running the switch-to-containers playbook with multisite enabled,
the fact "rgw_instances" is only set for the node being processed
(serial: 1), the consequence of that is that the set_fact of
'rgw_instances_all' can't iterate over all rgw node in order to look up
each 'rgw_instances_host'.
Adding a condition checking whether hostvars[item]["rgw_instances_host"]
is defined fixes this issue.
When calling the `ceph_key` module with `state: info`, if the ceph
command called fails, the actual error is hidden by the module which
makes it pretty difficult to troubleshoot.
The current code always states that if rc is not equal to 0 the keyring
doesn't exist.
`state: info` should always return the actual rc, stdout and stderr.
When no `[mgrs]` group is defined in the inventory, mgr daemon are
implicitly collocated with monitors.
This task currently relies on the length of the mgr group in order to
tell cephadm to deploy mgr daemons.
If there's no `[mgrs]` group defined in the inventory, it will ask
cephadm to deploy 0 mgr daemon which doesn't make sense and will throw
an error.
Since we fire up much less VMs than other job, we can affoard allocating
more memory here for this job.
Each VM hosts more daemon so 1024Mb can be too few.
When monitors and rgw are collocated with multisite enabled, the
rolling_update playbook fails because during the workflow, we run some
radosgw-admin commands very early on the first mon even though this is
the monitor being upgraded, it means the container doesn't exist since
it was stopped.
This block is relevant only for scaling out rgw daemons or initial
deployment. In rolling_update workflow, it is not needed so let's skip
it.
0990ae41099d8b9d678f546fc8790033f421711f changed the filter in
selectattr() from 'match' to 'equalto' but due to an incompatibility with
the Jinja2 version for python 2.7 on el7 we must stick to using 'match'
filter.
When osd nodes are collocated in the clients group (HCI context for
instance), the current logic will exclude osd nodes since they are
present in the client group.
The best fix would be to exclude clients node only when they are not
member of another group but for now, as a workaround, we can enforce
the addition of osd nodes to fix this specific case.
temporary work around vagrant cloud issue which seems broken at the time
of pushing this commit.
Let's pull images from cloud.centos.org for now since vagrant cloud
hosted images return a 403 error.
Since we need to revert 33bfb10, this is an alternative to initial approach.
We can avoid maintaining this file since it is present in container
image. The idea is to simply get it from the image container and write
it to the host.
This is a workaround for an issue in ansible.
When trying to stop/mask/disable this service in one task, the stop
didn't actually happen, the task doesn't fail but for some reason the
container is still present and running.
Then the task starting the service in the role ceph-crash fails because
it can't start the container since it's already running with the same
name.
ceph-ansible leaves a ceph-crash container in containerized deployment.
It means we end up with 2 ceph-crash containers running after the
migration playbook is complete.
When migrating from a cluster with no MDS nodes deployed,
`{{ cephfs_data_pool.name }}` doesn't exist so we need to create a pool
for storing nfs export objects.
projects / ceph-ansible.git / log