that doesn't seem to be 100% reproducible but it shows up after a
reboot. The only workaround we came up with at the moment is to run
`podman rm --storage <container>` before starting it.
Benoît Knecht [Wed, 7 Oct 2020 07:44:29 +0000 (09:44 +0200)]
ceph-facts: Fix osd_pool_default_crush_rule fact
The `osd_pool_default_crush_rule` is set based on `crush_rule_variable`, which
is the output of a `grep` command.
However, two consecutive tasks can set that variable, and if the second task is
skipped, it still overwrites the `crush_rule_variable`, leading the
`osd_pool_default_crush_rule` to be set to `ceph_osd_pool_default_crush_rule`
instead of the output of the first task.
This commit ensures that the fact is set right after the `crush_rule_variable`
is assigned, before it can be overwritten.
Gaudenz Steinlin [Mon, 28 Oct 2019 09:41:26 +0000 (10:41 +0100)]
config: Always use osd_memory_target if set
The osd_memory_target variable was only used if it was higher than the
calculated value based on the number of OSDs. This is changed to always
use the value if it is set in the configuration. This allows this value
to be intentionally set lower so that it does not have to be changed
when more OSDs are added later.
When deploying the ceph OSD via the packages then the ceph-osd@.service
unit is configured as enabled-runtime.
This means that each ceph-osd service will inherit from that state.
The enabled-runtime systemd state doesn't survive after a reboot.
For non containerized deployment the OSD are still starting after a
reboot because there's the ceph-volume@.service and/or ceph-osd.target
units that are doing the job.
When switching to containerized deployment we are stopping/disabling
ceph-osd@XX.servive, ceph-volume and ceph.target and then removing the
systemd unit files.
But the new systemd units for containerized ceph-osd service will still
inherit from ceph-osd@.service unit file.
As a consequence, if an OSD host is rebooting after the playbook execution
then the ceph-osd service won't come back because they aren't enabled at
boot.
This patch also adds a reboot and testinfra run after running the switch
to container playbook.
Add ceph_client tag to execute or skip the playbook
There are some use cases where there's a need to skip the execution
of the ceph-ansible client role even though the client section of the
inventory isn't empty.
This can happen in contexts where the services are colocated or when
a all-in-one deployment is performed.
The purpose of this change is adding a 'ceph_client' tag to avoid
altering the ceph-ansible execution flow but at the same time be able
to include or exclude a set of tasks using this tag.
Signed-off-by: Francesco Pantano <fpantano@redhat.com>
Gaudenz Steinlin [Tue, 27 Aug 2019 13:15:35 +0000 (15:15 +0200)]
osd: Fix number of OSD calculation
If some OSDs are to be created and others already exist the calculation
only counted the to be created OSDs. This changes the calculation to
take all OSDs into account.
Dimitri Savineau [Fri, 30 Oct 2020 14:54:16 +0000 (10:54 -0400)]
rolling_update: fix mgr start with mon collocation
cec994b introduced a regression when a mgr is collocated with a mon.
During the mon upgrade, the mgr service is masked to avoid to be
restarted on packages update.
Then the start mgr task is failing because the service is still masked.
Instead we should unmask it.
Dimitri Savineau [Mon, 26 Oct 2020 23:35:06 +0000 (19:35 -0400)]
rolling_update: use ceph health instead of ceph -s
The ceph status command returns a lot of information stored in variables
and/or facts which could consume resources for nothing.
When checking the cluster health, we're using the health structure in the
ceph status output.
To optimize this, we could use the ceph health command which contains
the same needed information.
$ ceph status -f json | wc -c
2001
$ ceph health -f json | wc -c
46
Dimitri Savineau [Mon, 26 Oct 2020 21:49:47 +0000 (17:49 -0400)]
rgw/rbdmirror: use service dump instead of ceph -s
The ceph status command returns a lot of information stored in variables
and/or facts which could consume resources for nothing.
When checking the rgw/rbdmirror services status, we're only using the
servicmap structure in the ceph status output.
To optimize this, we could use the ceph service dump command which contains
the same needed information.
This command returns less information and is slightly faster than the ceph
status command.
$ ceph status -f json | wc -c
2001
$ ceph service dump -f json | wc -c
1105
$ time ceph status -f json > /dev/null
real 0m0.557s
user 0m0.516s
sys 0m0.040s
$ time ceph service dump -f json > /dev/null
Dimitri Savineau [Mon, 26 Oct 2020 21:33:45 +0000 (17:33 -0400)]
monitor: use quorum_status instead of ceph status
The ceph status command returns a lot of information stored in variables
and/or facts which could consume resources for nothing.
When checking the quorum status, we're only using the quorum_names
structure in the ceph status output.
To optimize this, we could use the ceph quorum_status command which contains
the same needed information.
This command returns less information.
$ ceph status -f json | wc -c
2001
$ ceph quorum_status -f json | wc -c
957
$ time ceph status -f json > /dev/null
real 0m0.577s
user 0m0.538s
sys 0m0.029s
$ time ceph quorum_status -f json > /dev/null
Dimitri Savineau [Mon, 26 Oct 2020 15:23:01 +0000 (11:23 -0400)]
osds: use pg stat command instead of ceph status
The ceph status command returns a lot of information stored in variables
and/or facts which could consume resources for nothing.
When checking the pgs state, we're using the pgmap structure in the ceph
status output.
To optimize this, we could use the ceph pg stat command which contains
the same needed information.
This command returns less information (only about pgs) and is slightly
faster than the ceph status command.
$ ceph status -f json | wc -c
2000
$ ceph pg stat -f json | wc -c
240
$ time ceph status -f json > /dev/null
real 0m0.529s
user 0m0.503s
sys 0m0.024s
$ time ceph pg stat -f json > /dev/null
real 0m0.426s
user 0m0.409s
sys 0m0.016s
The data returned by the ceph status is even bigger when using the
nautilus release.
$ ceph status -f json | wc -c
35005
$ ceph pg stat -f json | wc -c
240
wangxiaotong [Sat, 24 Oct 2020 13:59:17 +0000 (21:59 +0800)]
osds: use ceph osd stat instead of ceph status
Improve the checked way of the OSD created checking process.
This replaces the ceph status command by the ceph osd stat command.
The osdmap structure isn't needed anymore.
$ ceph status -f json | wc -c
2001
$ ceph osd stat -f json | wc -c
132
$ time ceph status -f json > /dev/null
real 0m0.563s
user 0m0.526s
sys 0m0.036s
$ time ceph osd stat -f json > /dev/null
Benoît Knecht [Wed, 28 Oct 2020 15:09:58 +0000 (16:09 +0100)]
ceph-mon: Don't set monitor directory mode recursively
After rolling updates performed with
`infrastructure-playbooks/rolling_updates.yml`, files located in
`/var/lib/ceph/mon/{{ cluster }}-{{ monitor_name }}` had mode 0755 (including
the keyring), making them world-readable.
This commit separates the task that configured permissions recursively on
`/var/lib/ceph/mon/{{ cluster }}-{{ monitor_name }}` into two separate tasks:
1. Set the ownership and mode of the directory itself;
2. Recursively set ownership in the directory, but don't modify the mode.
Dimitri Savineau [Fri, 23 Oct 2020 18:50:52 +0000 (14:50 -0400)]
library/ceph_key: add output format parameter
The ceph_key module currently only supports the json output for the
info state.
When using this state on an entity then we something want the output
as:
- plain for copying it to another node.
- json in order to get only a subset information of the entity (like
the key or caps).
This patch adds the output_format parameter which uses json as a
default value for backward compatibility. It removes the internal and
hardcoded variable also called output_format.
In addition of json and plain outputs, there's also xml and yaml
values available.
Gaudenz Steinlin [Mon, 10 Aug 2020 09:52:56 +0000 (11:52 +0200)]
openstack: use ceph_keyring_permissions by default
Otherwise this task fails if no permission is set on the item.
Previously the code omited the mode parameter if it was not set, but
this was lost with commit ab370b6ad823e551cfc324fd9c264633a34b72b5.
Dimitri Savineau [Thu, 22 Oct 2020 14:59:15 +0000 (10:59 -0400)]
podman: force log driver to journald
Since we've changed to podman configuration using the detach mode and
systemd type to forking then the container logs aren't present in the
journald anymore.
The default conmon log driver is using k8s-file.
This file is currently deployed with '0644' ownership making this file
readable by any user on the system.
Since it contains sensitive information it should be readable by the
owner only.
When running rhel8 containers on a rhel7 host, after zapping an OSD
there's a discrepancy with the lvmetad cache that needs to be refreshed.
Otherwise, the host still sees the lv and can makes the user confused.
If user tries to redeploy an OSD, it will fail because the LV isn't
present and need to be recreated.
ie:
```
stderr: lsblk: ceph-block-8/block-8: not a block device
stderr: blkid: error: ceph-block-8/block-8: No such file or directory
stderr: Unknown device, --name=, --path=, or absolute path in /dev/ or /sys expected.
usage: ceph-volume lvm prepare [-h] --data DATA [--data-size DATA_SIZE]
[--data-slots DATA_SLOTS] [--filestore]
[--journal JOURNAL]
[--journal-size JOURNAL_SIZE] [--bluestore]
[--block.db BLOCK_DB]
[--block.db-size BLOCK_DB_SIZE]
[--block.db-slots BLOCK_DB_SLOTS]
[--block.wal BLOCK_WAL]
[--block.wal-size BLOCK_WAL_SIZE]
[--block.wal-slots BLOCK_WAL_SLOTS]
[--osd-id OSD_ID] [--osd-fsid OSD_FSID]
[--cluster-fsid CLUSTER_FSID]
[--crush-device-class CRUSH_DEVICE_CLASS]
[--dmcrypt] [--no-systemd]
ceph-volume lvm prepare: error: Unable to proceed with non-existing device: ceph-block-8/block-8
```
Benoît Knecht [Mon, 19 Oct 2020 09:23:59 +0000 (11:23 +0200)]
ceph-mon: Fix check mode for deploy monitor tasks
Skip the `get initial keyring when it already exists` task when both commands
whose `stdout` output it requires have been skipped (e.g. when running in check
mode).
Gaudenz Steinlin [Mon, 10 Aug 2020 09:38:47 +0000 (11:38 +0200)]
ceph-crash: Only deploy key to targeted hosts
The current task installs the ceph-crash key to "most" hosts via
"delegate_to". This key is only used by the ceph-crash daemon and should
just be installed on all hosts targeted by this role. There is no need
for using a delegated task.
The container_binding_name package was only mandatory when we were
using the docker modules (docker_image and docker_container) but since
we manage both docker and podman containers without using the dedicated
module then we can remove it.
Dimitri Savineau [Wed, 14 Oct 2020 00:43:53 +0000 (20:43 -0400)]
ceph-osd: don't start the OSD services twice
Using the + operation on two lists doesn't filter out the duplicate
keys.
Currently each OSDs is started (via systemd) twice.
Instead we could use the union filter.
It's time to remove this backward compatibility. Users had enough time
to convert their openstack_keys and key values.
We now fail in ceph-validate if the caps key isn't set.
Currently the `ceph_key` module doesn't support using a different
keyring than `client.admin`.
This commit adds the possibility to use a different keyring.
When rgw and osd are collocated, the current workflow prevents from
scaling out the radosgw_num_instances parameter when rerunning the
playbook in baremetal deployments.
When ceph-osd notifies handlers, it means rgw handlers are triggered
too. The issue with this is that they are triggered before the role
ceph-rgw is run.
In the case a scaleout operation is expected on `radosgw_num_instances`
it causes an issue because keyrings haven't been created yet so the new
instances won't start.
ceph-volume recently introduced a breaking change because of a `lvm
batch` refactor.
when rerunning `lvm batch --report --format json` on existing OSDs, it
doesn't output a valid json on stdout.
ceph-facts: add get default crush rule from running monitor
In case of deploying new monitor node to an existing cluster,
osd_pool_default_crush_rule should be taken from running monitor because
ceph-osd role won't be run and the new monitor will have different
osd_pool_default_crush_role from other monitors.
In non containerized deployment we check if the service is running
via the socket file presence.
This is done via the xxx_socket_stat variable that check the file
socket in the /var/run/ceph/ directory.
In some scenarios, we could have the socket file still present in
that directory but not used by any process.
That's why we have the xxx_stat variable which clean those leftovers.
The problem here is that we're set the variable for the handlers status
(like handler_mon_status) based on xxx_socket_stat instead of xxx_stat.
That means we will trigger the handlers if there's an old socket file
present on the system without any process associated.
af9f6684 introduced a regression on the ceph iscsi pool creation
because it was delegated to the first monitor node before that change.
This patch restores the initial worflow.
When the iscsi node doesn't have the admin keyring then the pool
creation fails.
This commit also ensures that the pool creation is only executed once
when having multiple iscsi nodes.
Otherwise this will generate an ansible warning about the missing
filter.
[DEPRECATION WARNING]: evaluating xxx as a bare variable, this behaviour
will go away and you might need to add |bool to the expression in the
future.
Also see CONDITIONAL_BARE_VARS configuration toggle.. This feature will
be removed in version 2.12.
When using the "absent" state on a non existing pool then the ceph_pool
module will fail and return a python traceback.
Instead we should check if the pool exit or not and execute the pool
deletion according to the result.
The state changed is now set when the pool is actually deleted.
This also disable add_file_common_args because we don't manipulate
files with this module.
ansible.cfg: remove cfg file in infrastructure-playbooks
There's no need ot have a copy of this file in infrastructure-playbooks
directory.
playbooks in that directory can be run from the root dir of
ceph-ansible.
As of 2.10, group names containing a dash are invalid.
However, setting this option makes it still possible to use a dash in
group names and prevent this warning to show up.
It might need to be definitely addressed in a future ansible release.
facts: fix 'set_fact rgw_instances with rgw multisite'
the current condition doesn't work, as soon as the first iteration is
done the condition makes next iterations skip since `rgw_instances` got
set with the first iteration.
When using a http(s) proxy with either docker or podman we can rely on
the HTTP_PROXY, HTTPS_PROXY and NO_PROXY environment variables.
But with ansible, even if those variables are defined in a source file
then they aren't loaded during the container pull/login tasks.
This implements the http(s) proxy support with docker/podman.
Both implementations are different:
1/ docker doesn't rely en the environment variables with the CLI.
Thos are needed by the docker daemon via systemd.
2/ podman uses the environment variables so we need to add them to
the login/pull tasks.
If the OSD directory is using symlinks for referencing devices (like
block, db, wal for bluestore and journal for filestore) then the chown
command could fail to change the owner:group on some system.
When running the switch2container playbook on a Debian based system
then the systemd unit path isn't the same than Red Hat based system.
Because the systemd unit files aren't removed then the new container
systemd unit isn't take in count.
projects / ceph-ansible.git / log