OSSA-2023-003: Unauthorized volume access through deleted volume attachments¶
May 10, 2023
Cinder: <20.2.1, >=21.0.0 <21.2.1, ==22.0.0
Glance_store: <3.0.1, >=4.0.0 <4.1.1, >=4.2.0 <4.3.1
Nova: <25.1.2, >=26.0.0 <26.1.2, ==27.0.0
Os-brick: <5.2.3, >=6.0.0 <6.1.1, >=6.2.0 <6.2.2
An unauthorized access to a volume could occur when an iSCSI or FC connection from a host is severed due to a volume being unmapped on the storage system and the device is later reused for another volume on the same host.
Scope: Only deployments with iSCSI or FC volumes are affected. However, the fix for this issue includes a configuration change in Nova and Cinder that may impact you on your next upgrade regardless of what backend storage technology you use. See the Configuration change section below, and item 4(B) in the Patches and Associated Deployment Changes for details.
This data leak can be triggered by two different situations.
Accidental case: If there is a problem with network connectivity during a normal detach operation, OpenStack may fail to clean the situation up properly. Instead of force-detaching the compute node device, Nova ignores the error, assuming the instance has already been deleted. Due to this incomplete operation OpenStack may end up selecting the wrong multipath device when connecting another volume to an instance.
Intentional case: A regular user can create an instance with a volume, and then delete the volume attachment directly in Cinder, which neglects to notify Nova. The compute node SCSI plumbing (over iSCSI/FC) will continue trying to connect to the original host/port/LUN, not knowing the attachment has been deleted. If a subsequent volume attachment re-uses the host/port/LUN for a different instance and volume, the original instance will gain access to it once the SCSI plumbing reconnects.
To prevent the intentional case, the Block Storage API provided by Cinder must only accept attachment delete requests from Nova for instance-attached volumes. A complicating factor is that Nova deletes an attachment by making a call to the Block Storage API on behalf of the user (that is, by passing the user’s token), which makes the request indistinguishable from the user making this request directly. The solution is to have Nova include a service token along with the user’s token so that Cinder can determine that the detach request is coming from Nova. The ability for Nova to pass a service token has been supported since Ocata, but has not been required until now. Thus, deployments that are not currently sending service user credentials from Nova will need to apply the relevant code changes and also make configuration changes to solve the problem.
Patches and Associated Deployment Changes¶
Given the above analysis, a thorough fix must include the following elements:
The os-brick library must implement the
forceoption for fibre channel, which which has only been available for iSCSI until now (covered by the linked patches).
Nova must call os-brick with the
forceoption when disconnecting volumes from deleted instances (covered by the linked patches).
In deployments where Glance uses the cinder glance_store driver, glance must call os-brick with the
forceoption when disconnecting volumes (covered by the linked patches).
Cinder must distinguish between safe and unsafe attachment delete requests and reject the unsafe ones. This part of the fix has two components:
The Block Storage API will return a 409 (Conflict) for a request to delete an attachment if there is an instance currently using the attachment, unless the request is being made by a service (for example, Nova) on behalf of a user (covered by the linked patches).
In order to recognize that a request is being made by a service on behalf of a user, Nova must be configured to send a service token along with the user token. If this configuration change is not made, the cinder change will reject any request to delete an attachment associated with a volume that is attached to an instance. Nova must be configured to send a service token to Cinder, and Cinder must be configured to accept service tokens. This is described in the following document and IS NOT AUTOMATICALLY APPLIED BY THE LINKED PATCHES: (Using service tokens to prevent long-running job failures) https://docs.openstack.org/cinder/latest/configuration/block-storage/service-token.html The Nova patch mentioned in step 2 includes a similar document more focused on Nova: doc/source/admin/configuration/service-user-token.rst
The cinder glance_store driver does not attach volumes to instances; instead, it attaches volumes directly to the Glance node. Thus, the Cinder change in step 4 will recognize an attachment-delete request coming from Glance as safe and allow it. (Of course, we expect that you will have applied the patches in steps 1 and 3 to your Glance nodes.)
An additional nova patch is required to fix a minor regression in periodic tasks and some nova-manage actions (errata 1). Also a patch to tempest is needed to account for behavior changes with fixes in place (errata 2). The stable/wallaby branch fix for nova introduced a regression which was addressed through subsequent adjustment of its patches after the advisory was published (errata 3).
https://review.opendev.org/882836 (2023.1/antelope cinder)
https://review.opendev.org/882851 (2023.1/antelope glance_store)
https://review.opendev.org/882858 (2023.1/antelope nova)
https://review.opendev.org/882859 (2023.1/antelope nova errata 1)
https://review.opendev.org/882843 (2023.1/antelope os-brick)
https://review.opendev.org/882835 (2023.2/bobcat cinder)
https://review.opendev.org/882834 (2023.2/bobcat glance_store)
https://review.opendev.org/882847 (2023.2/bobcat nova)
https://review.opendev.org/882852 (2023.2/bobcat nova errata 1)
https://review.opendev.org/882840 (2023.2/bobcat os-brick)
https://review.opendev.org/882876 (2023.2/bobcat tempest errata 2)
https://review.opendev.org/882869 (Wallaby nova)
https://review.opendev.org/882870 (Wallaby nova errata 1)
https://review.opendev.org/882839 (Xena cinder)
https://review.opendev.org/882855 (Xena glance_store)
https://review.opendev.org/882867 (Xena nova)
https://review.opendev.org/882868 (Xena nova errata 1)
https://review.opendev.org/882848 (Xena os-brick)
https://review.opendev.org/882838 (Yoga cinder)
https://review.opendev.org/882854 (Yoga glance_store)
https://review.opendev.org/882863 (Yoga nova)
https://review.opendev.org/882864 (Yoga nova errata 1)
https://review.opendev.org/882846 (Yoga os-brick)
https://review.opendev.org/882837 (Zed cinder)
https://review.opendev.org/882853 (Zed glance_store)
https://review.opendev.org/882860 (Zed nova)
https://review.opendev.org/882861 (Zed nova errata 1)
https://review.opendev.org/882844 (Zed os-brick)
Jan Wasilewski from Atman (CVE-2023-2088)
Gorka Eguileor from Red Hat (CVE-2023-2088)
Limited Protection Against Accidents… If you are only concerned with protecting against the accidental case described earlier in this document, steps 1-3 above should be sufficient. Note, however, that only applying steps 1-3 leaves your cloud wide open to the intentional exploitation of this vulnerability. Therefore, we recommend that the full fix be applied to all deployments.
Using Configuration as a Short-Term Mitigation… An alternative approach to mitigation can be found in OSSN-0092 https://wiki.openstack.org/wiki/OSSN/OSSN-0092
The stable/xena and stable/wallaby branches are under extended maintenance and will receive no new point releases, but patches for them are provided as a courtesy where available.
2023-05-15 - Errata 3
2023-05-10 - Errata 2
2023-05-10 - Errata 1
2023-05-10 - Original Version