yahoo-eng-team team mailing list archive

[OpenStack Infra <1821696@xxxxxxxxxxxxxxxxxx>]

Reviewed:  https://review.opendev.org/649951
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=56ca4d32ddf944b541b8a6c46f07275e7d8472bc
Submitter: Zuul
Branch:    master

commit 56ca4d32ddf944b541b8a6c46f07275e7d8472bc
Author: Lee Yarwood <lyarwood@xxxxxxxxxx>
Date:   Thu Apr 4 09:09:04 2019 +0100

    libvirt: Avoid using os-brick encryptors when device_path isn't provided
    
    When disconnecting an encrypted volume the Libvirt driver uses the
    presence of a Libvirt secret associated with the volume to determine if
    the new style native QEMU LUKS decryption or original decryption method
    using os-brick encrytors is used.
    
    While this works well in most deployments some issues have been observed
    in Kolla based environments where the Libvirt secrets are not fully
    persisted between host reboots or container upgrades. This can lead to
    _detach_encryptor attempting to build an encryptor which will fail if
    the associated connection_info for the volume does not contain a
    device_path, such as in the case for encrypted rbd volumes.
    
    This change adds a simple conditional to _detach_encryptor to ensure we
    return when device_path is not present in connection_info and native
    QEMU LUKS decryption is available. This handles the specific use
    case where we are certain that the encrypted volume was never decrypted
    using the os-brick encryptors, as these require a local block device on
    the compute host and have thus never supported rbd.
    
    It is still safe to build an encryptor and call detach_volume when a
    device_path is present however as change I9f52f89b8466d036 made such
    calls idempotent within os-brick.
    
    Change-Id: Id670f13a7f197e71c77dc91276fc2fba2fc5f314
    Closes-bug: #1821696


** Changed in: nova
       Status: In Progress => 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/1821696

Title:
  Failed to start instances with encrypted volumes

Status in kolla-ansible:
  In Progress
Status in kolla-ansible rocky series:
  New
Status in kolla-ansible stein series:
  In Progress
Status in OpenStack Compute (nova):
  Fix Released

Bug description:
  Description
  ===========
  We hit this bug after doing a complete cluster shutdown due to server room maintenance. The bug is however more easily reproducible.

  When cold starting an instance with an encrypted volume attached, it
  fails so start with a VolumeEncryptionNotSupported error.

  https://github.com/openstack/os-
  brick/blob/stable/rocky/os_brick/encryptors/cryptsetup.py#L52

  Steps to reproduce
  ==================

  * Deploy Openstack with Barbican support using Kolla.
  * Create an encrypted volume type
  * Create an encrypted volume
  * Create an instance and attach the encrypted folder
  * Enjoy your new instance and volume, install software and store data
  * In our case, we shut down the entire cluster and restarted it again. First all instances were stopped in Horizon using Shut down instance command. We use Ceph so we then stopped that using these procedures https://ceph.com/planet/how-to-do-a-ceph-cluster-maintenance-shutdown/ and then shut down the compute / storage nodes and then the controller nodes one by one. Then we started the cluster in the reverse order, verified all services were up and running, examined logs and then started the instances. 
  * Instances without encrypted volumes started fine.
  * Instances with encrypted volumes fail to start with VolumeEncryptionNotSupported.

  Note: It is possible to recreate the problem by using a Hard Reboot
  (possibly related https://bugs.launchpad.net/nova/+bug/1597234) or by
  shutting down instances and then restarting all Openstack services on
  that compute node.

  Expected results
  ================
  Instances with encrypted volumes should start fine, even after a Hard Reboot or a complete cluster shutdown.

  Actual results
  ==============
  Instances with encrypted volumes failed to start with VolumeEncryptionNotSupported

  https://pastebin.com/mvMbJQRb

  Environment
  ===========

  1. Openstack version
  Environment is established by Kolla (Rocky release).

  2. Hypervisor
  KVM on RHEL

  3. Storage type
  Ceph using Kolla (Rocky release)

  Analysis
  ========
  There seems to be a problem related to this code not behaving as expected:

  https://github.com/openstack/nova/blob/stable/rocky/nova/virt/libvirt/driver.py#L1049

  It seems that it is expected that the exception should be ignored and
  logged, but for some reason, the `ctxt.reraise = False` does not work
  as expected:

  self.force_reraise() is called in
  https://github.com/openstack/oslo.utils/blob/stable/rocky/oslo_utils/excutils.py#L220
  which it should not have hit since `reraise` is expected to be
  `False`.

  We did some hacking and just swallowed the exception by commenting out
  the `excutils.save_and_reraise_exception()` section and replacing it
  with a simple `pass`.

  Then the instance booted - but it could not boot from the image. But,
  it was then possible to remove the encrypted volume attachment, reboot
  the server and then reattach the encrypted volume.

To manage notifications about this bug go to:
https://bugs.launchpad.net/kolla-ansible/+bug/1821696/+subscriptions