Comment 14 for bug 1739323

Reviewed: https://review.openstack.org/529635
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=52679374a404b14578c834f47d6d4d35e85df2e1
Submitter: Zuul
Branch: stable/ocata

commit 52679374a404b14578c834f47d6d4d35e85df2e1
Author: Matthew Booth <email address hidden>
Date: Sat Dec 16 20:27:08 2017 +0000

    Fix an error in _get_host_states when deleting a compute node

    _get_host_states returns a generator which closes over seen_nodes, which
    is local, and self.host_state_map, which is global. It also modifies
    self.host_state_map, and will remove entries whose compute nodes are no
    longer present.

    If a compute node is deleted while a filter is still evaluating the
    generator returned by _get_host_states, the entry in self.host_state_map
    will be deleted if _get_host_states is called again. This will cause a
    KeyError when the first generator comes to evaluate the entry for the
    deleted compute node.

    We fix this by modifying the returned generator expression to check
    that a host_state_map entry still exists before returning it. An
    existing unit test is modified to exhibit the bug.

    Conflicts:
          nova/scheduler/filter_scheduler.py
          nova/scheduler/host_manager.py

    NOTE(mriedem): The conflict in filter_scheduler.py is due to
    48268c73e3f43fa763d071422816942942987f4a in Pike which is not
    needed here. The conflict in host_manager.py is due to
    d1de5385233ce4379b17a7404557c6724dc37cd4 in Pike which is not
    needed here.

    Change-Id: Ibb7c43a0abc433f93fc3de71146263e6d5923666
    Closes-Bug: #1739323
    (cherry picked from commit d72b33b986525a9b2c7aa08b609ae386de1d0e89)
    (cherry picked from commit b6a1a192955da3d18f50dfa3130c95fc15d7a66a)