Comment 28 for bug 1669054

Revision history for this message
OpenStack Infra (hudson-openstack) wrote : Fix merged to nova (stable/pike)

Reviewed: https://review.openstack.org/649387
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=efab235f88ac040ef9654c782236b83ffcb473e3
Submitter: Zuul
Branch: stable/pike

commit efab235f88ac040ef9654c782236b83ffcb473e3
Author: Matt Riedemann <email address hidden>
Date: Mon Mar 25 11:27:57 2019 -0400

    Do not persist RequestSpec.ignore_hosts

    Change Ic3968721d257a167f3f946e5387cd227a7eeec6c in Newton
    started setting the RequestSpec.ignore_hosts field to the
    source instance.host during resize/cold migrate if
    allow_resize_to_same_host=False in config, which it is by
    default.

    Change I8abdf58a6537dd5e15a012ea37a7b48abd726579 also in
    Newton persists changes to the RequestSpec in conductor
    in order to save the RequestSpec.flavor for the new flavor.
    This inadvertently persists the ignore_hosts field as well.

    Later if you try to evacuate or unshelve the server it will ignore
    the original source host because of the persisted ignore_hosts
    value. This is obviously a problem in a small deployment with only
    a few compute nodes (like an edge deployment). As a result, an
    evacuation can fail if the only available host is the one being
    ignored.

    This change does two things:

    1. In order to deal with existing corrupted RequestSpecs in the DB,
       this change simply makes conductor overwrite RequestSpec.ignore_hosts
       rather than append during evacuate before calling the scheduler so
       the current instance host (which is down) is filtered out.

       This evacuate code dealing with ignore_hosts goes back to Mitaka:

         I7fe694175bb47f53d281bd62ac200f1c8416682b

       The test_rebuild_instance_with_request_spec unit test is updated
       and renamed to actually be doing an evacuate which is what it was
       intended for, i.e. the host would not change during rebuild.

    2. This change makes the RequestSpec no longer persist the ignore_hosts
       field like several other per-operation fields in the RequestSpec.
       The only operations that use ignore_hosts are resize (if
       allow_resize_to_same_host=False), evacuate and live migration, and
       the field gets reset in each case to ignore the source instance.host.

    The related functional recreate test is also updated to show the
    bug is fixed. Note that as part of that, the confirm_migration method
    in the fake virt driver needed to be implemented otherwise trying to
    evacuate back to the source host fails with an InstanceExists error since
    the confirmResize operation did not remove the guest from the source host.

    Conflicts:
          nova/objects/request_spec.py
          nova/tests/unit/conductor/test_conductor.py
          nova/tests/unit/objects/test_request_spec.py

    NOTE(mriedem): The (test_)request_spec.py conflicts are due to not
    having change I2a78f0754c63381c57e7e1c610d0938b6df0f537 in Pike.
    The test_conductor.py conflict is due to not having change
    I70b11dd489d222be3d70733355bfe7966df556aa in Pike.

    NOTE(mriedem): The functional test is adjusted slightly to pass
    a string rather than list to _wait_for_migration_status since
    change I752617066bb2167b49239ab9d17b0c89754a3e12 is not in Pike.

    Change-Id: I3f488be6f3c399f23ccf2b9ee0d76cd000da0e3e
    Closes-Bug: #1669054
    (cherry picked from commit e4c998e57390c15891131205f7443fee98dde4ee)
    (cherry picked from commit 76dfb9d0b6cee1bbb553d3ec11139ca99a6f9474)
    (cherry picked from commit 31c08d0c7dc73bb2eefc69fa0f014b6478f3d149)
    (cherry picked from commit 8f1773a7af273e2739843a804fb4fe718c0a242d)