the properties target_size_ratio, target_size_bytes and pg_num_min are
used to fine-tune the pg_autoscaler and are set on a pool. The updated
pool list shows now autoscale settings & status. Including the new
(optimal) target PGs. To make it easier for new users to get/set the
correct amount of PGs.
Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
We want to check explicitly for type host, so filter for that first
and create a hash map for easier usage afterwards.
Drop the error when there's no tree, as either RADOS error'd on bad
command already, or there really is no tree (but RADOS worked OK), in
which case we simply return that the OSD did not belong to this node.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Allow destroying only OSDs that belong to the node that has been specified in
the API path.
So if
- OSD 1 belongs to node A and
- OSD 2 belongs to node B
then
- pvesh delete nodes/A/ceph/osd/1 is allowed but
- pvesh delete nodes/A/ceph/osd/2 is not
Destroying an OSD via GUI automatically inserts the correct node
into the API path.
pveceph automatically insert the local node into the API call, too.
Consequently, it can now only destroy local OSDs (fix#2053).
- pveceph osd destroy 1 is allowed on node A but
- pveceph osd destroy 2 is not
Signed-off-by: Dominic Jäger <d.jaeger@proxmox.com>
We get the device list from ceph-volume lvm list, and decode the json
output, which at that point is tainted (perlsec (1)).
Untaint it here before calling, because it is currently the only
call-site using the information in a problematic way (run_command).
(the only other call-site being in pve5to6)
Alternatively we could untaint while reading the information, but then
should only return a small subset of the ceph-volume output.
The issue is most likely due to
cb9db10c1a9855cf40ff13e81f9dd97d6a9b2698 in pve-common ('run_command:
improve performance for logging and long lines'),
Tested on a virtual testsetup by creating OSDs with second DB disk,
and destroying it via GUI (did not manage to get the error without the
DB disk)
Reported via our community forum:
https://forum.proxmox.com/threads/insecure-dependency-in-exec-during-osd-destroy.79574/
Signed-off-by: Stoiko Ivanov <s.ivanov@proxmox.com>
In some situations Ceph's auto-detection doesn't recognize the device
class correctly. The option allows to set it directly on osd create,
instead of altering it afterwards. This way the cluster doesn't need to
shift data back and forth unnecessarily.
Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
Luminous, Nautilus and Octopus. In Octopus the mon_status was dropped.
Also the ceph status was cleaned up and doesn't provide the mgrmap and
monmap.
The rados queries used in the ceph status API endpoints (cluster / node)
were factored out and merged to one place.
Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
The public_addr option for creating a new MON is only valid for manual
startup (since Ceph Jewel) and is just ignored by ceph-mon during setup.
As the MON is started after the creation through systemd without an IP
specified. It is trying to auto-select an IP.
Before this patch the public_addr was only explicitly written to the
ceph.conf if no public_network was set. The mon_address is only needed
in the config on the first start of the MON.
The ceph-mon itself tries to select an IP on the following conditions.
- no public_network or public_addr is in the ceph.conf
* startup fails
- public_network is in the ceph.conf
* with a single network, take the first available IP
* on multiple networks, walk through the list orderly and start on
the first network where an IP is found
Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
instead of having multiple regexes in various places for the name,
define a 'SERVICE_REGEX' in PVE::Ceph::Services, and use that
everywhere in the api where we need it
additionally limit new sevices to 200 characters, since
systemd units have a limit of 256 characters[0] (including suffix), and
200 seems to be enough.
users can now create ceph services on machines with hostnames
longer than 32 characters
0: https://www.freedesktop.org/software/systemd/man/systemd.unit.html
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
the network and the cluster network are optional in the ceph config
and with 'pveceph init', so only check if we have an ip address
from those networks if it is actually configured
otherwise, the createosd call dies with an 'ip' error message
even if it would work
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
It's possible to have a situation where the cluster network (used for
inter-OSD traffic) is not configured on a node. The OSD can still be
created but can't communicate.
This check will abort the creation if there is no IP within the subnet
of the cluster network present on the node. If there is no dedicated
cluster network the public network is used as a failsafe even though
this situation should not occur.
Signed-off-by: Aaron Lauterer <a.lauterer@proxmox.com>
Ceph MGR fails to start when installed on a node without existing
symlink to /etc/pve/ceph.conf.
Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
It was intended that for partitioned disks, we create one and use it.
Instead the code died always when the disk was used and not of type 'LVM'
We now check correctly the 2 cases:
* used for partitions and has gpt
* used and lvm
The remaining api call handles those two cases correctly
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
we want to improve the version hints in the osd tree gui and need
the version at the host nodes
we could (and want to) workaround it in the gui to have that
info for both versions of the api call
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
commit 970f96fdbb did not account for
getting the correct size parameter from the api call, so we ignored
it always resulting in uses not be able to set an explicit db/wal
size
Originally-fixed-by: Dominik Csapak <d.csapak@proxmox.com>
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
instead of silently ignoring them. since we are in a task worker here
this is especially important - otherwise the task status/result is also
wrong!
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
nautilus puts non running monitors also in the monmap, so only show
as running when it has quorum
this is also not 100% correct, but the only 'correct' alternative is
to try and get/parse the systemd status of the units and broadcast it
to the pmxcfs
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
this fixes an issue where only one monitor is in mon_host, which is
offline, prevents a client connection
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
ms_bind_ipv4 is default true and osds look for both
ipv6 and ipv4 addresses in cluster network/public network
since we only allow for one network each (which must be either
ipv4 or ipv6) we disallow ipv4 if ipv6 is detected
this fixes not starting osds on an ipv6 enabled, newly-setup cluster
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
when there is no 'public_network' in the config, the monitor
can only find an ip if it is given explicitly, either via commandline
(not possible with systemd) or via the ceph.conf
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
a 'mon remove' does this already for us, so do not stop it
this lead to a race where we could stop the next to the last monitor
before it was removed from the cluster, leading to a state
where two monitor were needed for quorum, but only one did exist
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
we need to remove an ip, ip:port or a ipvector from monhost
so use multiple regex search and replaces for this
this looks not really nice, but due to the strange format
of the line (e.g. ',' is a seperator inside and outside of a vector,
also ipv6 adresses may be surrounded with [] but so are vectors),
i found no better way
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
otherwise it is possible that multiple users create monitors at the same
time, resulting in a wrong ceph.conf and probably worse
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
in nautilus, the default msgr protocol is v2, but it has to be
explicitely given to monmaptool, also we don't want to use the
monitor sections anymore so only update mon_host
ceph can cope with mixed mon_host and monitor sections, so this is
not a problem
also the ceph-create-keys part is not necessary anymore since
this is done by the monitor itself now
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
by using our new 'get_services_info'
this already checks for nautilus+ style 'mon_host' key in the ceph.conf
for the ip address
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
it makes no sense to have the mon key inside the client.admin.keyring
also the order and operations did not make much sense
also create the client admin keyring when creating the config
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
if we already have a monitor, we can try to get the public_network via
the ceph configuration database
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
in a case where we cannot connect to any monitor, we did not get
any info even if we have them via the pmxcfs
so get the RADOS object in an eval, and get the info we have from the
config/pmxcfs, and set the state to unknown if we cannot query via RADOS
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
The aim of this patch is to reorder/rework the code of the api call
so that it gets more readable
it adds comments of what/why something is done, removes
code duplication between db/wal checks/creation
There are two changes in behaviour:
* when a device is given more than once via the api,
the user gets a parameter exception for the db or wal
with the information that the explicit defined devices must be
different
* we check the usage for db/wal before the worker, so that the user
gets instant feedback if a device is already in use
(this is more for api users than for gui users, since we do those
checks there also)
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
since the size of an LV can only be a multiple of 512b, we round
down to the next kib
we then have to mulitply it by 1024 for the partition, since
append_partition expects bytes and not kib
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
similar to the MDS api, so that DELETE and POST calls can operate on
the same path. This does not changes the CLI pveceph interface
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
As in a situation where we /had/ a manager but destroyed it this
key's value is a empty string, and if we pass that to the WebUI we
get strange results form of a ghost MGR entry with ExtJS auto-ID
generation as name -> pretty confusing.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
no point in first building a list if we can just remove it directly
afterwards, it's eval-ed anyway and $osd_list did not get touched
in-between.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
with this, osd destruction is left to ceph-volume if the osd was created
with ceph-volume, else our old code remains mostly the same since
we want to be able to destroy upgraded osds
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
this completely rewrites the ceph os creation api call using ceph-volume
since ceph-disk is not available anymore
breaking changes:
no filestore anymore, journal_dev -> db_dev
it is now possible to give a specific size for db/wal, default
is to read from ceph db/config and fallback is
10% of osd for block.db and 1% of osd for block.wal
the reason is that ceph-volume does not autocreate those itself
(like ceph-disk) but you have to create it yourself
if the db/wal device has an lvm on it with naming scheme 'ceph-UUID'
it uses that and creates a new lv
if we detect partitions, we create a new partition at the end
if the disk is not used at all, we create a pv/vg/lv for it
it is not possible to create osds on luminous with this api call anymore,
anyone needing this has to use ceph-disk directly
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
since we will have a seperate gui for the manager, we do not need this
anymore
this is a breaking api change
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
and use the broadcast when a service is added/removed
we will use 'get_cluster_service' in the future when we generate a list
of services of a specific type
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
From Nautilus release changelog[0]:
> The auid property for cephx users and RADOS pools has been removed.
> This was an undocumented and partially implemented capability that
> allowed cephx users to map capabilities to RADOS pools that they
> “owned”. Because there are no users we have removed this support.
[0]: https://ceph.com/releases/v14-2-0-nautilus-released/
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
ceph nautilus changed the structure of 'pg dump osds'
they moved the data one level below
parse both new and old format, and bail if it returns anything else
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
Reword the error message in find_mon_ip to make it more clear, that
there is no active IP configuration for the ceph public network.
Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
If calls aren't proxied to the selected node, which seems legit in
some cases, this will cause some misleading errors while ceph is
not installed on that node. Therefor the calls should now always get
proxied.
Signed-off-by: Tim Marx <t.marx@proxmox.com>
Instead of opening proc/mounts through IO::File directly for parsing,
the patch uses ProcFSTools. This way it also takes care of eventual
decoding.
Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
When destroying an OSD over API or CLI, e.g. by executing:
'pveceph osd destroy <num> --cleanup'
all disks associated with the OSD got wiped with dd, which included
any shared and by others still in use ones, e.g., separate disks with
DB/WAL.
The patch changes 'wipe_disks' to wipe the partition instead of the
whole disk.
Signed-off-by: Alwin Antreich <a.antreich@proxmox.com>
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
It makes more sense to have it there, especially since we want to
split out the service parts into a seperate file.
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
A MDS gets only active once a FS is there, and we need an MDS active
to be able to add a storage, as the CephFS plugin does an immediate
mount check. As an MDS needs some time to get active we had a
problematic time window where this mounting could fail.
Wait for a MDS to get in active state.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Allow to create a new CephFS instance and allow to list them.
As deletion requires coordination between the active MDS and all
standby MDS next in line this needs a bit more work. One could mark
the MDS cluster down and stop the active, that should work but as
destroying is quite a sensible operation, in production not often
needed I deemed it better to document this only, and leaving API
endpoints for this to the future.
For index/list I slightly transform the result of an RADOS `fs ls`
monitor command, this would allow relative easy display of a CephFS
and it's backing metadata and data pools in a GUI.
While for now it's not enabled by default and marked as experimental,
this API is designed to host multiple CephFS instances - we may not
need this at all, but I did not want to limit us early. And anybody
liking to experiment can use it after the respective ceph.conf
settings.
When encountering errors try to rollback. As we verified at the
beginning that we did not reused pools, destroy the ones which we
created.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Co-authored-by: Alwin Antreich <a.antreich@proxmox.com>
Allow to create, list and destroy and Ceph Metadata Server (MDS) over
the API and the CLI `pveceph` tool.
Besides setting up the local systemd service template and the MDS
data directory we also add a reference to the MDS in the ceph.conf
We note the backing host (node) from the respective MDS and set up a
'mds standby for name' = 'pve' so that the PVE created ones are a
single group. If we decide to add integration for rank/path specific
MDS (possible useful for CephFS with quite a bit of load) then this
may help as a starting point.
On create, check early if a reference already exists in ceph.conf and
abort in that case. If we only see existing data directories later
on we abort but do not remove them, they could well be from an older
manual create - where it's possible dangerous to just remove it. Let
the user handle it themself in that case.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Co-authored-by: Alwin Antreich <a.antreich@proxmox.com>
