This commit makes this playbook gathering facts from all other nodes but
clients.
When collocating OSDs on other nodes it can fail like following:
```
fatal: [vm252-11]: FAILED! => {
"msg": "'ansible.vars.hostvars.HostVarsVars object' has no attribute 'ansible_hostname'"
}
```
In that case, a fact from a RGW node is called when rendering the
`ceph.conf.j2` but it fails because facts are gathered only from mon and
osd nodes.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1806765
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
just like site.yml and rolling_update, let's exclude clients node from
the fact gathering.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 332c39376b)
(cherry picked from commit 5c3ba0787c)
This commit excludes client nodes from facts gathering, they are not
needed and can speed up this task.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 865d2eac9b)
Since ed36a11 we move the crush rules creation code from the ceph-mon to
the ceph-osd role.
To keep the backward compatibility we kept the possibility to set the
crush variables on the mons side but we didn't move the default values.
As a result, when using crush_rule_config set to true and wanted to use
the default values for crush_rules then the crush rule ansible task
creation will fail.
"msg": "'ansible.vars.hostvars.HostVarsVars object' has no attribute
'crush_rules'"
This patch move the default crush variables from ceph-mon to ceph-osd
role but also use those default values when nothing is defined on the
mons side.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1798864
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
(cherry picked from commit 1fc6b33714)
These tasks should have been removed with backport #4756
Note:
This should have been backported from master but it's not possible
because of too many change between master and stable-3.2
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1740463
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
in containerized context, using the binary provided in atomic os won't
work because it's an old version provided by ceph-common based on
10.2.5.
Using a container could be an idea but for large cluster with hundreds
of client nodes, that would require to pull the image of each of them
just to unmap the rbd devices.
Let's use the sysfs method in order to avoid any issue related to ceph
version that is shipped on the host.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1766064
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 3cfcc7a105)
There is no need to run these tasks n times from each monitor.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit c878e99589)
1. set noout and nodeep-scrub flags,
2. upgrade each OSD node, one by one, wait for active+clean pgs
3. after all osd nodes are upgraded, unset flags
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
Co-authored-by: Rachana Patel <racpatel@redhat.com>
(cherry picked from commit 548db78b95)
When upgrading from jewel to luminous we can execute the crush rule tasks
only when the 'osd require-osd-release luminous' command.
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
The ceph lockbox partition (part number 5) used with non lvm scenarios
and in non containerized deployment don't have a valid PARTUUID.
The value is set to 00000000-0000-0000-0000-000000000000 for each OSD
devices.
$ blkid -t PARTLABEL="ceph lockbox" -o value -s PARTUUID
00000000-0000-0000-0000-000000000000
00000000-0000-0000-0000-000000000000
00000000-0000-0000-0000-000000000000
00000000-0000-0000-0000-000000000000
00000000-0000-0000-0000-000000000000
When switching to containerized deployment we manually mount the lockbox
partition by using the PARTUUID.
Unfortunately because we have most of the time multiple OSD on the same
node we can't have the right symlink in /dev/disk/by-partuuid because it
will point to only one partition.
/dev/disk/by-partuuid/00000000-0000-0000-0000-000000000000 -> ../../sdb5
After the switch_to_containers playbook then only one OSD will restart
correctly and the other will try to access to the wrong device causing
error like 'xxxx is still in use'.
When deploying with containers and dmcrypt OSDs we force a PARTUUID
value during the ceph-disk prepare task.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1616159
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
When using fqdn in inventory, that playbook fails because of some tasks
using the result of ceph osd tree (which returns shortname) to get
some datas in hostvars[].
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1779021
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 6d9ca6b05b)
When a container is already running on a non containerized node then the
umount ceph partition task is skipped.
This is due to the container ps command which always returns 0 even if
the filter matches nothing.
We should run the umount task when:
1/ the container command is failing (not installed) : rc != 0
2/ the container command reports running ceph-osd containers : rc == 0
Also we should not fail on the ceph directory listing.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1616159
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
(cherry picked from commit 39cfe0aa65)
On non containerized deployment the ceph-mon hostname/fqdn systemd
service are stopped at the beginning of the mon upgrade.
But the parameter enabled is set to true for both task so even if we're
not using the fqdn then it will enabled the systemd unit based on it.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1649617
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
This commit adds a default value in the with_dict because when using
python 2.7, if a task using a with_dict has a condition, it is
evaluated anyway whereas in python 3 it isn't.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1766499
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
There's no need to use the default filter on active/standby groups
because if the group doesn't exist then the play is just skipped.
Currently this generates warnings like:
[WARNING]: Could not match supplied host pattern, ignoring: |
[WARNING]: Could not match supplied host pattern, ignoring: default([])
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
(cherry picked from commit 2ca79fcc99)
The active mds host should be based on the inventory hostname and not on
the ansible hostname.
The value returns under the mdsmap structure is based on the OS hostname
so we need to find the right node in the inventory with this value when
doing operation on inventory nodes.
Othewise we could see error like:
The task includes an option with an undefined variable. The error was:
"hostvars[foobar]" is undefined
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
(cherry picked from commit f1f2352c79)
Let's skip this part of the code if there's no mds node in the
inventory.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 5ec906c3af)
When switching from a baremetal deployment to a containerized deployment
we only umount the OSD data partition.
If the OSD is encrypted (dmcrypt: true) then there's an additional
partition (part number 5) used for the lockbox and mount in the
/var/lib/ceph/osd-lockbox/ directory.
Because this partition isn't umount then the containerized OSD aren't
able to start. The partition is still mount by the system and can't be
remount from the container.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1616159
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
(cherry picked from commit 19edf707a5)
9e7972a introduced a regression via the container_binary variable
which is undefined.
The CEPH_CONTAINER_BINARY environment variable isn't used at all.
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
By changing the set ownership command from using the file module in combination with a with_items loop to a raw chown command, we can achieve a 98% performance increase here.
On a ceph cluster with a significant amount of directories and files in /var/lib/ceph, the file module has to run checks on ownership of all those directories and files to determine whether a change is needed.
In this case, we just want to explicitly set the ownership of all these directories and files to the ceph_uid
Added context note to all set proper ownership tasks
Signed-off-by: Kevin Jones <kevinjones@redhat.com>
(cherry picked from commit 47bf47c9d8)
we must use the ids instead of device names in the tasks executed in
`post_tasks` for the osd rolling update otherwise it ends up with old
systemd units enabled.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1739209
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
Since bedc0ab we now manage ceph-osd systemd unit scripts based on ID
instead of device name but it was not present in the shrink-osd
playbook (ceph-disk version).
To keep backward compatibility on deployment that didn't do yet the
transition on OSD id then we should stop unit scripts for both device
and ID.
This commit adds the ulimit nofile container option to get better
performance on ceph-disk commands.
It also fixes an issue when the OSD id matches multiple OSD ids with
the same first digit.
$ ceph-disk list | grep osd.1
/dev/sdb1 ceph data, prepared, cluster ceph, osd.1, block /dev/sdb2
/dev/sdg1 ceph data, prepared, cluster ceph, osd.12, block /dev/sdg2
Finally removing the shrinked OSD directory.
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
When shrinking an OSD, its corresponding 'prepare container' should be
removed otherwise it prevent from redeploying a new osd because of this
leftover.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
Removing the gpt header on devices will ease ceph-disk to ceph-volume
migration when using shrink-osd + add-osd playbooks.
ceph-disk requires GPT header where ceph-volume will complain if GPT
header is present.
That won't break ceph-disk (re)deployment since we check and add the GPT
header if needed when deploying ceph-disk ODs.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1613735
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
This tries to first unmount any cephfs/nfs-ganesha mount point on client
nodes, then unmap any mapped rbd devices and finally it tries to remove
ceph kernel modules.
If it fails it means some resources are still busy and should be cleaned
manually before continuing to purge the cluster.
This is done early in the playbook so the cluster stays untouched until
everything is ready for that operation, otherwise if you try to redeploy
a cluster it could end up by getting confused by leftover from previous
deployment.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1337915
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 20e4852888)
ceph-facts should be run before we play ceph-validate since it has
reference to facts that are set in ceph-facts role.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
The ceph-agent role was used only for RHCS 2 (jewel) so it's not
usefull anymore.
The current code will fail on CentOS distribution because the rhscon
package is only avaible on Red Hat with the RHCS 2 repository and
this ceph release is supported on stable-3.0 branch.
Resolves: #4020
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
(cherry picked from commit 7503098ca0)
We must stop tcmu-runner after the other rbd-target-* services
because they may need to interact with tcmu-runner during shutdown.
There is also a bug in some kernels where IO can get stuck in the
kernel and by stopping rbd-target-* first we can make sure all IO is
flushed.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1659611
Signed-off-by: Mike Christie <mchristi@redhat.com>
(cherry picked from commit d7ef12910e)
Currently only rbd-target-gw service is restarted during an update.
We also need to restart tcmu-runner and rbd-target-api services
during the ceph iscsi upgrade.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1659611
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
(cherry picked from commit f1048627ea)
`lvm_volumes` and/or `devices` variable(s) can be undefined depending on
the scenario chosen.
These tasks should be run only if these variable are defined, otherwise
it ends up with undefined variable errors.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1653307
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 0180738313)
The systemd ceph-osd@.service file used for starting the ceph osd
containers is used in all osd_scenarios.
Currently purging a containerized deployment using the lvm scenario
didn't remove the ceph-osd systemd service.
If the next deployment is a non-containerized deployment, the OSDs
won't be online because the file is still present and override the
one from the package.
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
(cherry picked from commit 7cc626b72d)
otherwise, it will end up with error like following:
```
FAILED! => {"msg": "'ansible.vars.hostvars.HostVarsVars object' has no attribute 'ansible_hostname'"}
```
because facts won't have been gathered.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1670663
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit a440878533)
as of b70d54ac80 the service launched isn't
ceph-rbd-mirror@admin.service.
it's now `ceph-rbd-mirror@rbd-mirror.{{ ansible_hostname }}`
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit a915308477)
removing the content of this directory seems a bit agressive and cause a
redeployment to fail after a purge on debian based distrubition.
Typical error:
```
fatal: [mon0]: FAILED! => changed=false
attempts: 3
msg: No package matching 'ceph' is available
```
The following task will consider the cache is still valid, so apt
doesn't refresh it:
```
- name: update apt cache if cache_valid_time has expired
apt:
update_cache: yes
cache_valid_time: 3600
register: result
until: result is succeeded
```
since the task installing ceph packages has a `update_cache: no` it
fails:
```
- name: install ceph for debian
apt:
name: "{{ debian_ceph_pkgs | unique }}"
update_cache: no
state: "{{ (upgrade_ceph_packages|bool) | ternary('latest','present') }}"
default_release: "{{ ceph_stable_release_uca | default('') }}{{ ansible_distribution_release ~ '-backports' if ceph_origin == 'distro' and ceph_use_distro_backports else '' }}"
register: result
until: result is succeeded
```
/tmp/* isn't specific to ceph as well, so we shouldn't remove everything
in this directory.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 3849f30f58)
using `shell` module seems to be the only way to make this task working
on rhel based distribution AND debian based distributions.
on ubuntu, using `command` ansible module fails like following
(not due to `sudo` usage or not):
```
ok: [osd1] => changed=false
cmd: command -v ceph-volume
failed_when_result: false
msg: '[Errno 2] No such file or directory: ''command'': ''command'''
rc: 2
```
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1653307
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 89f77589fa)
Let's enforce the default value for `client_update_batch` to 20 since
`ansible_forks` isn't always available.
Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1650184
Signed-off-by: Giulio Fidente <gfidente@redhat.com>
(cherry picked from commit ff8dbe114c)
Some people use the switch playbook to perform upgrade so they end up in
the same situation than https://bugzilla.redhat.com/show_bug.cgi?id=1650572
This is applying the same fix as
729744c6a8.
We don't want to fail on key that are not present since they will get
created after the mons are updated. They will be created by the task
"create potentially missing keys (rbd and rbd-mirror)".
Signed-off-by: Sébastien Han <seb@redhat.com>
sometimes we play the whole role `ceph-defaults` just to access the
default value of some variables. It means we play the `facts.yml` part
in this role while it's not desired. Splitting this role will speedup
the playbook.
Closes: #3282
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit 0eb56e36f8)
This task has to be called after the role `ceph-defaults` has been
played, otherwise, `mon_group_name` will never be known.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
(cherry picked from commit a12de3e048)
Guillaume Abrioux
Dimitri Savineau
Kevin Jones
Mike Christie
Noah Watkins
Noah Watkins
Giulio Fidente
Sébastien Han