Volume stuck in creating, or in error right after create¶
Symptom¶
A user creates a volume (openstack volume create --size <n> …, possibly --image, --snapshot or --type). The
request is accepted, but the volume never becomes available:
$ openstack volume show <volume-id> -c status -c os-vol-host-attr:host -c type -c created_at
| status | error | # within seconds of the request
...
| status | creating | # or downloading, minutes or hours after the request
A request refused at once (quota exceeded, invalid volume type, size too small for the image) returns an error and
creates no volume: that is not this page (openstack quota show <project-id> shows the volume and gigabytes limits).
Impact¶
The volumes being created, for one project or for everyone, depending on the cause. Existing volumes and their
attachments are not affected. A volume in error or stuck in creating holds no user data yet, but it counts
against the project's quota until it is deleted. The risk comes from the fixes: forcing a status by hand can hide a
volume that the backend is in fact still creating.
Quick checks¶
Read-only. With admin credentials, from any host with the OpenStack client:
# 1. Status, the backend chosen (None if no backend was chosen) and the type
# (the column is "type"; python-openstackclient 9.0.0 silently ignores -c volume_type)
openstack volume show <volume-id> -c status -c os-vol-host-attr:host -c type -c size -c created_at
# 2. Cinder's own explanation (volume API 3.3 or later), for this volume
openstack --os-volume-api-version 3.3 volume message list | grep <volume-id>
# 3. Are the Cinder services up? A backend's cinder-volume is listed as <host>@<backend>.
openstack volume service list
# 4. What each pool has left, as the scheduler sees it
openstack volume backend pool list --long
# 5. The volume type's properties: volume_backend_name must match a configured backend
openstack volume type show <type> -c properties
How to read them:
| You see | Go to |
|---|---|
error, empty host, check 2 says Could not find any available weighted backend. |
Cause 1 |
error, a host is set, check 2 says Driver failed to create the volume. or Image used for creating volume exceeds available space., or check 2 shows nothing and the cinder-volume log has the error |
Cause 2 |
creating or downloading for longer than the image size explains, and the cinder-volume of its host is down in check 3 |
Cause 3 |
creating, empty host, every service up |
Cause 4 |
creating (during the Glance download) or downloading, a host set, the service up, no error in the log |
Cause 5 |
Diagnosis¶
Cinder's path for a new volume¶
cinder-api creates the volume record in creating and casts the request to cinder-scheduler over RabbitMQ. The
scheduler filters the backends (by default AvailabilityZoneFilter, CapacityFilter and CapabilitiesFilter),
weighs the rest, records the chosen backend as os-vol-host-attr:host and casts to that backend's cinder-volume.
cinder-volume asks the driver to create the volume. From an image, it sets the status to downloading, then
writes the image to the volume. On success, the status becomes available. If the scheduler finds no backend, it
sets error itself, which is why error within seconds and an empty host point to the scheduler.
A cinder-volume or cinder-scheduler service that starts again sets the volumes it had left in creating or
downloading to error: it does not resume them (Cleaning Volume with id <volume-id> and status creating in its
log). A request it had not yet received is different: it waits in RabbitMQ and is carried out when the service is
back, and the volume becomes available.
Logs¶
Search for the volume ID:
Kolla-Ansible: /var/log/kolla/cinder/cinder-scheduler.log on the hosts running cinder_scheduler, and
/var/log/kolla/cinder/cinder-volume.log on the hosts running cinder_volume (the cinder group, the control
hosts by default in 2026.1).
Any: the cinder-scheduler and cinder-volume service logs.
What to look for:
- in the scheduler log:
Filtering removed all hosts for the request with volume ID '<volume-id>'. Filter results: …, which shows the filter that removed the backends (for exampleCapacityFilter: (start: 1, end: 0)), next toNo valid backend was found(that line does not carry the volume ID); - in the volume log: the driver's exception (LVM, RBD, NFS, a vendor driver), or the image download and conversion;
- no trace of the volume ID in either log: the request was lost before reaching a service (cause 4).
Possible causes¶
Cause 1: no backend passes the scheduler¶
- Capacity: no pool has enough free space for the size, once
reserved_percentageand, for thin provisioning,max_over_subscription_ratioare applied. Check 4 showsfree_capacity_gbper pool. - Capabilities: the volume type's
volume_backend_name(or another extra spec) matches no configured backend, for example after a backend was renamed incinder.conf. - Availability zone: the requested zone has no
cinder-volumeservice. - Backend down: the only matching backend's
cinder-volumeisdown, so the scheduler does not see it. Right after acinder-volumerestart, the scheduler may not have its capabilities yet: on a lab a volume created 8 s after the service wasupagain went toerrorwithAvailabilityZoneFilter: (start: 0, end: 0).
Cause 2: the backend failed to create it¶
The backend was chosen and the driver failed: an LVM volume group full (vgs on the cinder-volume host shows its
free space; with thin provisioning, Kolla-Ansible's default, the space is in the thin pool: lvs cinder-volumes,
column Data%), a Ceph pool full or unreachable, an NFS share not mounted, a vendor array refusing the request. For a
volume from an image: no space in image_conversion_dir on the cinder-volume host to download or convert the
image, or a Glance error. An image whose content does not match its disk_format (for example qcow2 uploaded as
raw) is refused by Cinder 2026.1: ImageUnacceptable: … The image format was claimed to be 'raw' but the image data
appears to be in a different format., with no user message in check 2.
Cause 3: the service that had the request stopped¶
cinder-volume stopped or lost RabbitMQ while it was creating the volume. The volume stays creating or
downloading until that service starts again, which then sets it to error. A clean stop (systemctl stop or
restart of the Kolla-Ansible unit) waits for the creation in progress to finish; a crash or a kill does not. On
Kolla-Ansible the unit restarts a crashed container after RestartSec (Restart=always; 11 s with the Kolla defaults,
59 s on the lab, which sets docker_client_timeout: 600).
Cause 4: the request never reached a service¶
The record exists, but the cast to the scheduler or to cinder-volume was lost (RabbitMQ unavailable or
partitioned at that moment). Nothing will move the volume: no service knows about it.
Cause 5: the image is still being written¶
Not a failure. A large image, or one that must be converted to raw, takes time to download and write. The
cinder-volume log shows nothing for the volume while the image downloads, then Image download <n> MB at <n> MB/s
and, when converted, Converted <n> MB image at <n> MB/s. On a lab the status stayed creating during the download
and showed downloading only after it. Cinder's image-volume cache makes later volumes from the same image
much faster.
Resolution¶
Work by risk. The quick checks are read-only. Then fix the cause, from the least to the most intrusive step. Most
of the time a volume in error cannot be retried: fix the cause and create a new volume.
For cause 1¶
Read-only first: compare the requested size with free_capacity_gb (check 4), and the type's volume_backend_name
with the volume_backend_name of each backend in cinder.conf. Then:
- capacity: free space on the backend, or add one; raising
max_over_subscription_ratioon a thin backend is a capacity decision, not a fix; - a type that names no backend: correct the type's property (a change for every future volume of that type);
⚠️ State-changing. Changes the volume type for every future volume of that type; existing volumes keep their backend. Undo: set the previous value back. Cinder refuses the change with
Volume Type is currently in use. (HTTP 400)while any volume of that type exists, including the ones inerror: delete those first (see the clean-up below), or create a new type with the right property.
openstack volume type set --property volume_backend_name=<backend-name> <type>
- a backend
down: see cause 3.
For cause 2¶
Fix the backend (space in the volume group or pool, the NFS mount, the array) or the conversion directory, as the
cinder-volume log says. Then create a new volume.
For cause 3¶
Find why the service stopped (its log, RabbitMQ) and start it again. On start, it sets the volumes it had left in
creating or downloading to error.
⚠️ State-changing. Restarts the
cinder-volumeservice of this host. Volumes being created or attached through it at that moment can fail and need retrying; volumes already attached to instances stay attached. Fix what stopped it first: a restart does not fix a full backend or a broken RabbitMQ.
Kolla-Ansible:
sudo systemctl restart kolla-cinder_volume-container.service
Any:
sudo systemctl restart <cinder-volume-unit>
For cause 4¶
Check RabbitMQ first (docker exec rabbitmq rabbitmqctl cluster_status on Kolla-Ansible). If the volume has no host
and no service knows it, nothing will finish it: its status has to be set by hand, then the volume deleted.
🛑 Destructive. Forces Cinder's record of the volume to
error. If a backend is in fact still creating it (a host is set and that service is working on it), the record then lies about the backend, and the delete below can fail or leave an orphan on the storage. Precondition:os-vol-host-attr:hostis empty, or thecinder-volumelog of that host shows no activity for the volume ID; and the volume was neveravailable.
openstack volume set --state error <volume-id>
Clean up volumes left in error¶
A volume in error holds no user data when it was never available. Delete it to give the quota back:
🛑 Destructive. Deletes the volume and, if the backend created anything, its storage. Precondition: the volume was never
available(check itscreated_atand the logs) and is not attached (attachmentsempty). Cannot be undone.
openstack volume delete <volume-id>
For cause 5¶
Wait, and follow the copy in the cinder-volume log. For the next volumes from that image, enable the
image-volume cache on the backend, and,
with Ceph RBD for both Glance and Cinder, keep images in raw so that Cinder clones them instead of copying.
Verification¶
openstack volume service list
# expected: every cinder-scheduler and cinder-volume service up
openstack volume create --size 1 runbook-check
openstack volume show runbook-check -c status -c os-vol-host-attr:host
# expected: available, on the expected backend
openstack volume delete runbook-check
The verification volume is created and deleted on purpose:
⚠️ State-changing. Creates a 1 GB test volume, deleted right after. Needs volume and gigabytes quota.
Lab observations¶
On a lab: OpenStack 2026.1 (Cinder 28.0.1), Kolla-Ansible 22.2.0, all-in-one, Cinder LVM (thin, 19 GB pool,
max_over_subscription_ratio 20) over iSCSI, RabbitMQ 4.2 with quorum queues, python-openstackclient 9.0.0.
- Cause 1, capacity.
--size 500:errorwithin seconds, hostNone, check 2schedule allocate volume:Could not find any available weighted backend., scheduler logInsufficient free virtual space (380.0GB) to accommodate thin provisioned 500GB volume. - Cause 1, capabilities. A type with
volume_backend_name=lvm-renamed: same message,CapabilitiesFilter: (start: 1, end: 0).volume type set --propertywas refused while the error volume of that type existed, and worked once it was deleted; the next volume of the type wasavailable. - Cause 2. A qcow2 image uploaded as
raw:errorin 4 s, host set, no user message,ImageUnacceptableincinder-volume.log. The backend had created the logical volume;volume deleteremoved it. - Cause 3.
docker kill cinder_volume37 s into a 2 GB image download: the volume stayedcreatingwith its host set, the service wasdownafter about 55 s, systemd restarted the container after 60 s, and the start-up clean-up set the volume toerror. Asystemctl stopduring a creation instead waited 21 s for it to finish (available). - Cause 4.
cinder-schedulerstopped, a volume created, thecinder-schedulerqueue purged, the scheduler started: the volume stayedcreating, hostNone, every serviceup, no trace of its ID in any Cinder log.volume set --state errorthenvolume deleteworked as written. - Verification ran as written.
Not run: cause 5 as a separate case (seen during cause 3), Ceph, NFS, a real RabbitMQ outage.
Validity and limits¶
- Written from Cinder 2026.1 (the scheduler and volume managers, their clean-up on start, the default scheduler
filters, user messages,
service_down_time) and Kolla-Ansible 2026.1. Causes 1 to 4 reproduced on an all-in-one lab (see Lab observations). - The user messages quoted come from
cinder/message/message_field.py; how they appear depends on the client. - Not covered: volumes created by Nova for a boot from volume (see the related runbook), backups and restores,
snapshots stuck in
creating, replication, multi-attach, encrypted volumes, and migrations or retypes stuck inmigratingorretyping.
Related¶
- Instance fails to boot from volume: the same failures seen from Nova, when it creates the volume.
- Volume stuck in attaching, reserved or detaching
- Stale volume attachment after the instance was deleted or evacuated
- Cinder 2026.1: scheduler filters, image-volume cache, messages API
Provenance: written by agent:ozarland-author, from the Cinder and Kolla-Ansible 2026.1 documentation and source. Reviewed in its pull request (see the page history). Checked by a human operator: not yet.
Worked for you, or not? Tell us in one line: your release, your deployment method, and what happened.