yahoo-eng-team team mailing list archive

Thread
Date
[Bug 2019190] Re: [SRU][RBD] Retyping of in-use boot volumes renders instances unusable (possible data corruption)

To: yahoo-eng-team@xxxxxxxxxxxxxxxxxxx
From: James Page <2019190@xxxxxxxxxxxxxxxxxx>
Date: Thu, 25 Jan 2024 16:30:09 -0000
Reply-to: Bug 2019190 <2019190@xxxxxxxxxxxxxxxxxx>
Sender: noreply@xxxxxxxxxxxxx
Included in most recent snapshots for Caracal

** Changed in: cinder (Ubuntu Noble)
       Status: New => Fix Released

-- 
You received this bug notification because you are a member of Yahoo!
Engineering Team, which is subscribed to OpenStack Compute (nova).
https://bugs.launchpad.net/bugs/2019190

Title:
  [SRU][RBD] Retyping of in-use boot volumes renders instances unusable
  (possible data corruption)

Status in Cinder:
  New
Status in Cinder wallaby series:
  New
Status in Ubuntu Cloud Archive:
  New
Status in Ubuntu Cloud Archive antelope series:
  New
Status in Ubuntu Cloud Archive bobcat series:
  New
Status in Ubuntu Cloud Archive caracal series:
  New
Status in Ubuntu Cloud Archive yoga series:
  New
Status in Ubuntu Cloud Archive zed series:
  New
Status in OpenStack Compute (nova):
  Invalid
Status in cinder package in Ubuntu:
  Fix Released
Status in cinder source package in Jammy:
  New
Status in cinder source package in Lunar:
  New
Status in cinder source package in Mantic:
  New
Status in cinder source package in Noble:
  Fix Released

Bug description:
  [Impact]

  See bug description for full details but short summary is that a patch
  landed in Wallaby release that introduced a regression whereby
  retyping an in-use volume leaves the attached volume in an
  inconsistent state with potential for data corruption. Result is that
  a vm does not receive updated connection_info from Cinder and will
  keep pointing to the old volume, even after reboot.

  [Test Plan]

  * Deploy Openstack with two Cinder RBD storage backends (different pools)
  * Create two volume types
  * Boot a vm from volume: openstack server create --wait --image jammy --flavor m1.small --key-name testkey --nic net-id=8c74f1ef-9231-46f4-a492-eccdb7943ecd testvm --boot-from-volume 10
  * Retype the volume to type B: openstack volume set --type typeB --retype-policy on-demand <volume>
  * Go to compute host running vm and check that the vm is now copying data to the new location e.g.

      <disk type='network' device='disk'>
        <driver name='qemu' type='raw' cache='none' discard='unmap'/>
        <auth username='cinder-ceph'>
          <secret type='ceph' uuid='01b65a79-22a3-4672-80e7-5a47b0e5581a'/>
        </auth>
        <source protocol='rbd' name='cinder-ceph/volume-b68be47d-f526-4f98-a77b-a903bf8b6c65' index='1'>
          <host name='10.5.2.236' port='6789'/>
        </source>
        <mirror type='network' job='copy'>
          <format type='raw'/>
          <source protocol='rbd' name='cinder-ceph-alt/volume-c6b55b4c-a540-4c39-ad1f-626c964ae3e1' index='2'>
            <host name='10.5.2.236' port='6789'/>
            <auth username='cinder-ceph-alt'>
              <secret type='ceph' uuid='e089e27e-3a2f-49d6-b6d9-770f52177eb1'/>
            </auth>
          </source>
          <backingStore/>
        </mirror>
        <target dev='vda' bus='virtio'/>
        <serial>b68be47d-f526-4f98-a77b-a903bf8b6c65</serial>
        <alias name='virtio-disk0'/>
        <address type='pci' domain='0x0000' bus='0x00' slot='0x03' function='0x0'/>
      </disk>

  which will eventually settle and change to:

      <disk type='network' device='disk'>
        <driver name='qemu' type='raw' cache='none' discard='unmap'/>
        <auth username='cinder-ceph-alt'>
          <secret type='ceph' uuid='e089e27e-3a2f-49d6-b6d9-770f52177eb1'/>
        </auth>
        <source protocol='rbd' name='cinder-ceph-alt/volume-c6b55b4c-a540-4c39-ad1f-626c964ae3e1' index='2'>
          <host name='10.5.2.236' port='6789'/>
        </source>
        <backingStore/>
        <target dev='vda' bus='virtio'/>
        <serial>b68be47d-f526-4f98-a77b-a903bf8b6c65</serial>
        <alias name='virtio-disk0'/>
        <address type='pci' domain='0x0000' bus='0x00' slot='0x03' function='0x0'/>
      </disk>

  * And lastly a reboot of the vm should be successfull.

  [Regression Potential]
  Given that the current state is potential data corruption and the patch will fix this by successfully refreshing connection info I do not see a regression potential. It is in fact fixing a regression.

  -------------------------------------------------------------------------

  While trying out the volume retype feature in cinder, we noticed that after an instance is
  rebooted it will not come back online and be stuck in an error state or if it comes back
  online, its filesystem is corrupted.

  ## Observations

  Say there are the two volume types `fast` (stored in ceph pool `volumes`) and `slow`
  (stored in ceph pool `volumes.hdd`). Before the retyping we can see that the volume
  for example is present in the `volumes.hdd` pool and has a watcher accessing the
  volume.

  ```sh
  [ceph: root@mon0 /]# rbd ls volumes.hdd
  volume-81cfbafc-4fbb-41b0-abcb-8ec7359d0bf9

  [ceph: root@mon0 /]# rbd status volumes.hdd/volume-81cfbafc-4fbb-41b0-abcb-8ec7359d0bf9
  Watchers:
          watcher=[2001:XX:XX:XX::10ad]:0/3914407456 client.365192 cookie=140370268803456
  ```

  Starting the retyping process using the migration policy `on-demand` for that volume either
  via the horizon dashboard or the CLI causes the volume to be correctly transferred to the
  `volumes` pool within the ceph cluster. However, the watcher does not get transferred, so
  nobody is accessing the volume after it has been transferred.

  ```sh
  [ceph: root@mon0 /]# rbd ls volumes
  volume-81cfbafc-4fbb-41b0-abcb-8ec7359d0bf9

  [ceph: root@mon0 /]# rbd status volumes/volume-81cfbafc-4fbb-41b0-abcb-8ec7359d0bf9
  Watchers: none
  ```

  Taking a look at the libvirt XML of the instance in question, one can see that the `rbd`
  volume path does not change after the retyping is completed. Therefore, if the instance is
  restarted nova will not be able to find its volume preventing an instance start.

  #### Pre retype

  ```xml
  [...]
  <source protocol='rbd' name='volumes.hdd/volume-81cfbafc-4fbb-41b0-abcb-8ec7359d0bf9' index='1'>
      <host name='2001:XX:XX:XXX::a088' port='6789'/>
      <host name='2001:XX:XX:XXX::3af1' port='6789'/>
      <host name='2001:XX:XX:XXX::ce6f' port='6789'/>
  </source>
  [...]
  ```

  #### Post retype (no change)

  ```xml
  [...]
  <source protocol='rbd' name='volumes.hdd/volume-81cfbafc-4fbb-41b0-abcb-8ec7359d0bf9' index='1'>
      <host name='2001:XX:XX:XXX::a088' port='6789'/>
      <host name='2001:XX:XX:XXX::3af1' port='6789'/>
      <host name='2001:XX:XX:XXX::ce6f' port='6789'/>
  </source>
  [...]
  ```

  ### Possible cause

  While looking through the code that is responsible for the volume retype we found a function
  `swap_volume` volume which by our understanding should be responsible for fixing the association
  above. As we understand cinder should use an internal API path to let nova perform this action.
  This doesn't seem to happen.

  (`_swap_volume`:
  https://github.com/openstack/nova/blob/stable/wallaby/nova/compute/manager.py#L7218)

  ## Further observations

  If one tries to regenerate the libvirt XML by e.g. live migrating the instance and rebooting the
  instance after, the filesystem gets corrupted.

  ## Environmental Information and possibly related reports

  We are running the latest version of TripleO Wallaby using the hardened (whole disk)
  overcloud image for the nodes.

  Cinder Volume Version: `openstack-
  cinder-18.2.2-0.20230219112414.f9941d2.el8.noarch`

  ### Possibly related

  - https://bugzilla.redhat.com/show_bug.cgi?id=1293440

  (might want to paste the above to a markdown file for better
  readability)

To manage notifications about this bug go to:
https://bugs.launchpad.net/cinder/+bug/2019190/+subscriptions