As per #1013 it appears that BS will soon use THP to lower TLB misses,
also disabling THP hasn't demonstrated any gains so far.
Closes: https://github.com/ceph/ceph-ansible/issues/1013
Signed-off-by: Sébastien Han <seb@redhat.com>
38dc20e74b introduced a bug in the purge
playbooks because using `*` in `command` module doesn't work.
`/var/lib/ceph/*` files are not purged it means there is a leftover.
When trying to redeploy a cluster, it failed because monitor daemon was
detecting existing keyring, therefore, it assumed a cluster already
existed.
Typical error (from container output):
```
Sep 26 13:18:16 mon0 docker[31316]: 2018-09-26 13:18:16 /entrypoint.sh: Existing mon, trying to rejoin cluster...
Sep 26 13:18:16 mon0 docker[31316]: 2018-09-26 13:18:16.9323937f15b0d74700 -1 auth: unable to find a keyring on /etc/ceph/test.client.admin.keyring,/etc/ceph/test.keyring,/etc/ceph/keyring,/etc/ceph/keyring.bin,:(2) No such file or directory
Sep 26 13:18:23 mon0 docker[31316]: 2018-09-26 13:18:23 /entrypoint.sh:
SUCCESS
```
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1633563
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
Previous commit c13a3c3 has removed a condition.
This commit brings back this condition which is essential to ensure we
won't hit a false positive result in the `when` condition for the check
PGs task.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
In cluster with a large number of PGs, it can be expected some of them
scrubbing, it's a normal operation.
Preventing from scrubbing operation force to set noscrub flag before a
rolling update which is a problem because it pauses an important data
integrity operation until the end of the rolling upgrade.
This commit allows an upgrade even while PGs are scrubbing.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1616066
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
List the osd_scenarios and install options that are validated by the
ceph-validate role in the documentation.
Signed-off-by: Andrew Schoen <aschoen@redhat.com>
get more coverage by adding an RGW daemon collocated on osd0.
We've missed a bug in the past which could have been caught earlier in
the CI.
Let's add this additional daemon in order to have a better coverage.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
`+` is more idiomatic for "one or more" in a regex than `{1,}`; the
latter was introduced in a previous fix for an incorrect `{1,2}`
restriction.
Signed-off-by: Matthew Vernon <mv3@sanger.ac.uk>
After restarting each OSD, restart_osd_daemon.sh checks that the
cluster is in a good state before moving on to the next one. One of
the checks it does is that the number of pgs in the state
"active+clean" is equal to the total number of pgs in the cluster.
On large clusters (e.g. we have 173,696 pgs), it is likely that at
least one pg will be scrubbing and/or deep-scrubbing at any one
time. These pgs are in state "active+clean+scrubbing" or
"active+clean+scrubbing+deep", so the script was erroneously not
including them in the "good" count. Similar concerns apply to
"active+clean+snaptrim" and "active+clean+snaptrim_wait".
Fix this by considering as good any pg whose state contains
active+clean. Do this as an integer comparison to num_pgs in pgmap.
(could this be backported to at least stable-3.0 please?)
Closes: #2008
Signed-off-by: Matthew Vernon <mv3@sanger.ac.uk>
Previously RETRIES was set (by default to 40) once at the start of the
script; this meant that it would only ever wait for up to 40 lots of
30s across *all* the OSDs on a host before bombing out. In fact, we
want to be prepared to wait for the same amount of time after each OSD
restart for the clusters' pgs to be happy again before continuing.
Closes: #3154
Signed-off-by: Matthew Vernon <mv3@sanger.ac.uk>
...with the exception of the purge operation, since
removing Calamari would still be useful for an old
cluster.
Signed-off-by: John Spray <john.spray@redhat.com>
In case of an OpenStack "box", the Vagrantfile intend to check the
existence of os_networks and os_floating_ip_pool settings in
vagrant_variables.yml and pass them to the provider if they are set.
Due to two typos in the Vagrantfile this is not working as it checks the
wrong variable names.
This commit fixes the typos so these settings can be used.
Signed-off-by: Norbert Illés <illesnorbi@gmail.com>
For now our best guess is to count the number of devices and multiply
by osds_per_device. Ideally we'd like to run ceph-volume lvm batch
--report and get the number of OSDs that way, but currently we need
a ceph.conf in place already before we can do that. There is a tracker
ticket that would allow os to get around the need for a ceph.conf:
http://tracker.ceph.com/issues/36088
Fixes: https://github.com/ceph/ceph-ansible/issues/3135
Signed-off-by: Andrew Schoen <aschoen@redhat.com>
the default value for _rgw_hostname was took from the current node being
played while it should be took from the respective node in the loop.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1622505
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
This avoids errors when the osd scenario choosen does not require
setting devices or lvm_volumes. The default values for these are not
set because they exist in the ceph-osd role, not ceph-defaults.
Signed-off-by: Andrew Schoen <aschoen@redhat.com>
BlueStore's cache is sized conservatively by default, so that it does
not overwhelm under-provisioned servers. The default is 1G for HDD, and
3G for SSD.
To replace the page cache, as much memory as possible should be given to
BlueStore. This is required for good performance. Since ceph-ansible
knows how much memory a host has, it can set
`bluestore cache size = max(total host memory / num OSDs on this host * safety
factor, 1G)`
Due to fragmentation and other memory use not included in bluestore's
cache, a safety factor of 0.5 for dedicated nodes and 0.2 for
hyperconverged nodes is recommended.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1595003
Signed-off-by: Neha Ojha <nojha@redhat.com>
Co-Authored-by: Guillaume Abrioux <gabrioux@redhat.com>
- Adds loop in bash to satisfy the 1:n relation between `osd_hosts` and the
different device lists.
- Fixes some container name which were using the host hostname instead
of the actual container one.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
`mon_group_name` isn't defined here, we must hardcode it.
Typical error:
```
The task includes an option with an undefined variable. The error was: 'mon_group_name' is undefined
```
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
Once the OSD is destroyed we also have to purge the associated devices,
this means purging journal, db , wal partitions too.
This now works for container and non-container.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1572933
Signed-off-by: Sébastien Han <seb@redhat.com>
The commit:
commit 1164cdc002
Author: Guillaume Abrioux <gabrioux@redhat.com>
Date: Thu Aug 2 11:58:47 2018 +0200
iscsigw: install ceph-iscsi-cli package
installs the cli package but does not start and enable the
rbd-target-api daemon needed for gwcli to communicate with the igw
nodes. This patch just enables and starts it for the non-container
setup. The container setup is already doing this.
This fixes bz https://bugzilla.redhat.com/show_bug.cgi?id=1613963
Signed-off-by: Mike Christie <mchristi@redhat.com>
ce1dd8d introduced the purge osd on containers but it was incorrect.
`resolve parent device` and `zap ceph osd disks` tasks must be delegated to
their respective OSD nodes.
Indeed, they were run on the ansible node, it means it was trying to
resolve parent devices from this node where it should be done on OSD
nodes.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1612095
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
As of rhel 7.6, it has been decided it doesn't make sense to confine
`ganesha_t` anymore. It means this domain won't exist anymore.
Let's add a `failed_when: false` in order to make the deployment not
failing when trying to run this command.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1626070
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
using sphinx 1.8.0 breaks our doc test CI job.
Typical error:
```
Exception occurred:
File
"/home/jenkins-build/build/workspace/ceph-ansible-docs-pull-requests/docs/.tox/docs/lib/python2.7/site-packages/sphinx/highlighting.py", line 26, in <module>
from sphinx.ext import doctest
SyntaxError: unqualified exec is not allowed in function 'run' it contains a nested function with free variables (doctest.py, line 97)
```
See: https://github.com/sphinx-doc/sphinx/issues/5417
Pinning to 1.7.9 to fix our CI.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
If this is set to anything other than the default value of 1 then the
--osds-per-device flag will be used by the batch command to define how
many osds will be created per device.
Signed-off-by: Andrew Schoen <aschoen@redhat.com>
This command line is not supported.
According to official documentation:
```
Note that shell command lines are not directly supported.
If shell command lines are to be used,
they need to be passed explicitly to a shell implementation of some kind.
```
We must run this using /bin/sh instead.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
let's add ansible_hostname as a default value for rgw_hostname if no
hostname in servicemap matches ansible_fqdn.
Fixes: #3063
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1622505
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
Using `UPDATE_*` environment variables here will make an upgrade of the
ceph release when running switch_to_containers scenario which is not
correct.
Eg:
If ceph luminous was first deployed, then we should switch to ceph
luminous containers, not to mimic.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
This commit is adding quotes that make keyring unusuable
eg:
```
client.john
key: AQAN0RdbAAAAABAAH5D3WgMN9Rxw3M8jkpMIfg==
caps: [mds] ''
caps: [mgr] 'allow *'
caps: [mon] 'allow rw'
caps: [osd] 'allow rw'
```
Trying to import such a keyring and use it will result:
```
Error EACCES: access denied
```
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1623417
This reverts commit 424815501a.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
When ceph-nfs is deployed containerized and ceph-common is not
installed on the host the start_nfs task fails because the rados
command is missing on the host.
Run rados commands from a ceph container instead so that
they will succeed.
Signed-off-by: Tom Barron <tpb@dyncloud.net>
If the ceph-radosgw target is not enabled, then enabling the
ceph-radosgw@ service has no effect since nothing will pull
it on the next reboot. As such, we need to ensure that the
target is enabled.
Signed-off-by: Markos Chandras <mchandras@suse.de>
Sometime /var/lib/ceph is mounted on a device so we won't be able to
remove it (device busy) so let's remove its content only.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1615872
Signed-off-by: Sébastien Han <seb@redhat.com>
The dummy client container currently wont work on non-x86_64 hosts.
This PR creates a filtered client group that contains only hosts
that are x86_64 - which can then be the group to run the
dummy container against.
This is for the specific case of a containerized_deployment where
there is a mixture of non-x86_64 hosts and x86_64 hosts. As such
the filtered group will contain all hosts when running with
containerized_deployment: false.
Currently ppc64le is not supported for Ceph server components.
Signed-off-by: Andy McCrae <andy.mccrae@gmail.com>
We have been supporting multiple devices for journalin containerized
deployments for a while now and forgot about this.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1622393
Signed-off-by: Sébastien Han <seb@redhat.com>
Add a message for when PV creation fails.
This message alerts users that FS/GPT/RAID
signatures could still on the device and the
reason for the failures.
`wipefs -a $device` needs to be run to fix this issue.
Signed-off-by: Ali Maredia <amaredia@redhat.com>
Same problem again... ceph_release_num[ceph_release] is only set in
ceph-docker-common/common roles so putting the condition on that role
will never work. Removing the condition.
The downside of this is we will be installing packages and then skip the
role on the node.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1622210
Signed-off-by: Sébastien Han <seb@redhat.com>
If we play site-docker.yml, we are already in a
containerized_deployment. So the condition is not needed.
Signed-off-by: Sébastien Han <seb@redhat.com>
There is no point of using hosts running on atomic AND centos hosts. So
let's run containerized scenarios on Atomic only.
This solves this error here:
```
fatal: [client2]: FAILED! => {
"failed": true
}
MSG:
The conditional check 'ceph_current_status.rc == 0' failed. The error was: error while evaluating conditional (ceph_current_status.rc == 0): 'dict object' has no attribute 'rc'
The error appears to have been in '/home/jenkins-build/build/workspace/ceph-ansible-nightly-luminous-stable-3.1-ooo_collocation/roles/ceph-defaults/tasks/facts.yml': line 74, column 3, but may
be elsewhere in the file depending on the exact syntax problem.
The offending line appears to be:
- name: set_fact ceph_current_status (convert to json)
^ here
```
From https://2.jenkins.ceph.com/view/ceph-ansible-stable3.1/job/ceph-ansible-nightly-luminous-stable-3.1-ooo_collocation/37/consoleFull#1765217701b5dd38fa-a56e-4233-a5ca-584604e56e3a
What's happening here is all the hosts excepts the clients are running atomic, so here: https://github.com/ceph/ceph-ansible/blob/master/site-docker.yml.sample#L62
The condition will skipped all the nodes excepts the clients, thus when running ceph-default, the task "is ceph running already?" is skipped but the task above needs the rc of the skipped task.
This is not an error from the playbook, it's a CI setup issue.
Signed-off-by: Sébastien Han <seb@redhat.com>
A couple if things were wrong in the initial commit:
* ceph_release_num[ceph_release] >= ceph_release_num['luminous'] will
never work since the ceph_release fact is set in the roles after. So
either ceph-common or ceph-docker-common set it
* we can easily re-use the initial command to check if a cluster is
running, it's more elegant than running it twice.
* set the fact rgw_hostname on rgw nodes only
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1618678
Signed-off-by: Sébastien Han <seb@redhat.com>
This file is not needed, if you want to generate a key you can run:
python -c "import os ; import struct ; import time; import base64 ; key
= os.urandom(16) ; header =
struct.pack('<hiih',1,int(time.time()),0,len(key)) ;
print(base64.b64encode(header + key).decode())"
Signed-off-by: Sébastien Han <seb@redhat.com>
Sébastien Han
Rishabh Dave
Guillaume Abrioux
Andrew Schoen
Giulio Fidente
Matthew Vernon
John Spray
Norbert Illés
Neha Ojha
Mike Christie
Tom Barron
Markos Chandras
Alfredo Deza
Andy McCrae
Ali Maredia