Even if this improves ceph-disk/ceph-volume performances then it also
impact the ceph-osd process.
The ceph-osd process shouldn't use 1024:4096 value for the max open
files.
Removing the ulimit option from the container engine and doing this kind
of change on the container side [1].
[1] https://github.com/ceph/ceph-container/pull/1497
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1702285
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
On containerized deployment, the OSD entrypoint runs some ceph-volume
commands (lvm/simple scan and/or activate) which perform badly without
the ulimit option.
This option was added for all previous ceph-volume commands but not on
the ceph-osd container startup.
Also updating hard limit value to 4096 to reflect default baremetal
value.
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
This environment variable was added in cb381b4 but was removed in
4d35e9e.
This commit reintroduces the change.
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
When using podman, the systemd unit scripts don't have a dependency
on the network. So we're not sure that the network is up and running
when the containers are starting.
With docker this behaviour is already handled because the systemd
unit scripts depend on docker service which is started after the
network.
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
Otherwise content in /run/udev is mislabeled and prevent some services
like NetworkManager from starting.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
In containerized deployment the default osd cpu quota is too low
for production environment using NVMe devices.
This is causing performance degradation compared to bare-metal.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1695880
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
We don't need to set After=docker.service when the container_binary
variable isn't set to docker.
It doesn't break anything currently but it could be confusing when
using podman.
Signed-off-by: Dimitri Savineau <dsavinea@redhat.com>
This reverts commit bb2bbeb941.
Looks like when not passing `--pid=host` we are facing some issues when
deploying more than 2 OSDs in containerized environment.
At the moment, we are still troubleshooting this issue but we prefer to
revert this commit so it doesn't block any PR in the CI.
As soon as we have a fix; we will push a new PR to remove `--pid=host`
(a revert of revert...)
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
instead of using `RuntimeDirectory` parameter in systemd unit files,
let's use a systemd `tmpfiles.d` to ensure `/run/ceph`.
Explanation:
`podman` doesn't create the `/var/run/ceph` if it doesn't exist the time
where the container is run while `docker` used to create it.
In case of `switch_to_containers` scenario, `/run/ceph` gets created by
a tmpfiles.d systemd file; when switching to containers, the systemd
unit file complains because `/run/ceph` already exists
The better fix would be to ensure `/usr/lib/tmpfiles.d/ceph-common.conf`
is removed and only rely on `RuntimeDirectory` from systemd unit file parameter
but we come from a non-containerized environment which is already running,
it means `/run/ceph` is already created and when starting the unit to
start the container, systemd will still complain and we can't simply
remove the directory if daemons are collocated.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
/var/run/ceph resides in a non persistent filesystem (tmpfs)
After a reboot, all daemons won't start because this directory will be
missing.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
without this, the command `ceph-volume lvm list --format json` hangs and
takes a very long time to complete.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
Applying and passing the OSD_BLUESTORE/FILESTORE on the fly is wrong for
existing clusters as their config will be changed.
Typically, if an OSD was prepared with ceph-disk on filestore and we
change the default objectstore to bluestore, the activation will fail.
The flag osd_objectstore should only be used for the preparation, not
activation. The activate in this case detects the osd objecstore which
prevents failures like the one described above.
Signed-off-by: Sébastien Han <seb@redhat.com>
If an existing cluster runs this config, and has ceph-disk OSD, the
`expose_partitions` won't be expected by jinja since it's inside the
'old' if. We need it as part of the osd_scenario != 'lvm' condition.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1640273
Signed-off-by: Sébastien Han <seb@redhat.com>
We don't need to pass the device and discover the OSD ID. We have a
task that gathers all the OSD ID present on that machine, so we simply
re-use them and activate them. This also handles the situation when you
have multiple OSDs running on the same device.
Signed-off-by: Sébastien Han <seb@redhat.com>
We don't need to pass the hostname on the container name but we can keep
it simple and just call it ceph-osd-$id.
Signed-off-by: Sébastien Han <seb@redhat.com>
expose_partitions is only needed on ceph-disk OSDs so we don't need to
activate this code when running lvm prepared OSDs.
Signed-off-by: Sébastien Han <seb@redhat.com>
Fixes the deprecation warning:
[DEPRECATION WARNING]: Using tests as filters is deprecated. Instead of
using `result|search` use `result is search`.
Signed-off-by: Noah Watkins <nwatkins@redhat.com>
The container runs with --rm which means it will be deleted by Docker
when exiting. Also 'docker rm -f' is not idempotent and returns 1 if the
container does not exist.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1609007
Signed-off-by: Sébastien Han <seb@redhat.com>
Useful for softwares that do data collection/monitoring like collectd.
They can connect to the socket and then retrieve information.
Even though the sockets are exposed now, I'm keeping the docker exec to
check the socket, this will allow newer version of ceph-ansible to work
with older versions.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1563280
Signed-off-by: Sébastien Han <seb@redhat.com>
We know bindmount with the :z option at the end of the -v command so
this will basically run the exact same command as we used to run. So to
speak:
chcon -Rt svirt_sandbox_file_t /var/lib/ceph
Signed-off-by: Sébastien Han <seb@redhat.com>
This commit fixes a bug that occurs especially for dmcrypt scenarios.
There is an issue where the 'disk_list' container can't reach the ceph
cluster because it's not launched with `--net=host`.
If this container can't reach the cluster, it will hang on this step
(when trying to retrieve the dm-crypt key) :
```
+common_functions.sh:448: open_encrypted_part(): ceph --cluster abc12 --name \
client.osd-lockbox.9138767f-7445-49e0-baad-35e19adca8bb --keyring \
/var/lib/ceph/osd-lockbox/9138767f-7445-49e0-baad-35e19adca8bb/keyring \
config-key get dm-crypt/osd/9138767f-7445-49e0-baad-35e19adca8bb/luks
+common_functions.sh:452: open_encrypted_part(): base64 -d
+common_functions.sh:452: open_encrypted_part(): cryptsetup --key-file \
-luksOpen /dev/sdb1 9138767f-7445-49e0-baad-35e19adca8bb
```
It means the `ceph-run-osd.sh` script won't be able to start the
`osd_disk_activate` process in ceph-container because he won't have
filled the `$DOCKER_ENV` environment variable properly.
Adding `--net=host` to the 'disk_list' container fixes this issue.
Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1543284
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
We have a scenario when we switch from non-container to containers. This
means we don't know anything about the ceph partitions associated to an
OSD. Normally in a containerized context we have files containing the
preparation sequence. From these files we can get the capabilities of
each OSD. As a last resort we use a ceph-disk call inside a dummy bash
container to discover the ceph journal on the current osd.
Closes: https://bugzilla.redhat.com/show_bug.cgi?id=1525612
Signed-off-by: Sébastien Han <seb@redhat.com>
The name docker_version is very generic and is also used by other
roles. As a result, there may be name conflicts. To avoid this a
ceph_ prefix should be used for this fact. Since it is an internal
fact renaming is not a problem.
Add the variables ceph_osd_docker_cpuset_cpus and
ceph_osd_docker_cpuset_mems, so that a user may specify
the CPUs and memory nodes of NUMA systems on which OSD
containers are run.
Provides a example in osds.yaml.sample to guide user
based on sample `lscpu` output since cpuset-mems refers
to the memory by NUMA node only while cpuset-cpus can
refer to individual vCPUs within a NUMA node.
there is no need to have a condition on this task, this test should be
always run since the result will be interpreted later.
Signed-off-by: Guillaume Abrioux <gabrioux@redhat.com>
During the initial implementation of this 'old' thing we were falling
into this issue without noticing
https://github.com/moby/moby/issues/30341 and where blindly using --rm,
now this is fixed the prepare container disappears and thus activation
fail.
I'm fixing this for old jewel images.
Also this fixes the machine reboot case where the docker logs are
purgend. In the old scenario, we now store the log locally in the same
directory as the ceph-osd-run.sh script.
Signed-off-by: Sébastien Han <seb@redhat.com>
Use "ceph_tcmalloc_max_total_thread_cache" to set the
TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES value inside /etc/default/ceph for
Debian installs, or /etc/sysconfig/ceph for Red Hat/CentOS installs.
By default this is set to 0, so the default package value will be used,
if specified this value will be changed to match the variable, and ceph
osd services will be restarted.
There was a huge resync from luminous to jewel in ceph-docker:
https://github.com/ceph/ceph-docker/pull/797
This change brought a new handy function to discover partitions tight to
an OSD. This function doesn't exist in the old image so the
ceph-osd-run.sh script breaks when trying to deploy Jewel OSD with that
old Jewel image version.
Signed-off-by: Sébastien Han <seb@redhat.com>
This will solve the following issue when starting docker containers on ubuntu:
invalid argument "1\u00a0" for --cpus=1 : failed to parse 1 as a rational number
Closes-bug: #2056
This is causing unknown issues when trying to start a dmcrypt container.
Basically the container is stuck at mount opening the LUKS device. This
is still unknown why this is causing trouble but we need to move
forward. Also, this doesn't seem to help in any ways to fix the race
condition we've seen.
Here is the log for dmcrypt:
cryptsetup 1.7.4 processing "cryptsetup --debug --verbose --key-file
key luksClose fbf8887d-8694-46ca-b9ff-be79a668e2a9"
Running command close.
Locking memory.
Installing SIGINT/SIGTERM handler.
Unblocking interruption on signal.
Allocating crypt device context by device
fbf8887d-8694-46ca-b9ff-be79a668e2a9.
Initialising device-mapper backend library.
dm version [ opencount flush ] [16384] (*1)
dm versions [ opencount flush ] [16384] (*1)
Detected dm-crypt version 1.14.1, dm-ioctl version 4.35.0.
Device-mapper backend running with UDEV support enabled.
dm status fbf8887d-8694-46ca-b9ff-be79a668e2a9 [ opencount flush ]
[16384] (*1)
Releasing device-mapper backend.
Trying to open and read device /dev/sdc1 with direct-io.
Allocating crypt device /dev/sdc1 context.
Trying to open and read device /dev/sdc1 with direct-io.
Initialising device-mapper backend library.
dm table fbf8887d-8694-46ca-b9ff-be79a668e2a9 [ opencount flush
securedata ] [16384] (*1)
Trying to open and read device /dev/sdc1 with direct-io.
Crypto backend (gcrypt 1.5.3) initialized in cryptsetup library
version 1.7.4.
Detected kernel Linux 3.10.0-693.el7.x86_64 x86_64.
Reading LUKS header of size 1024 from device /dev/sdc1
Key length 32, device size 1943016847 sectors, header size 2050
sectors.
Deactivating volume fbf8887d-8694-46ca-b9ff-be79a668e2a9.
dm status fbf8887d-8694-46ca-b9ff-be79a668e2a9 [ opencount flush ]
[16384] (*1)
Udev cookie 0xd4d14e4 (semid 32769) created
Udev cookie 0xd4d14e4 (semid 32769) incremented to 1
Udev cookie 0xd4d14e4 (semid 32769) incremented to 2
Udev cookie 0xd4d14e4 (semid 32769) assigned to REMOVE task(2) with
flags (0x0)
dm remove fbf8887d-8694-46ca-b9ff-be79a668e2a9 [ opencount flush
retryremove ] [16384] (*1)
fbf8887d-8694-46ca-b9ff-be79a668e2a9: Stacking NODE_DEL [verify_udev]
Udev cookie 0xd4d14e4 (semid 32769) decremented to 1
Udev cookie 0xd4d14e4 (semid 32769) waiting for zero
Signed-off-by: Sébastien Han <seb@redhat.com>
Dimitri Savineau
Guillaume Abrioux
Sébastien Han
Noah Watkins
Ning Yao
Christian Berendt
John Fulton
Andy McCrae