Deadlock creating >5 servers at once using xenapi driver

[Johannes Erdfelt]

A deadlock can occur when a compute node is working on more than 5 servers at once. This only affects systems using the xenapi driver.

The problem is caused by a nested use of _get_session(). If multiple greenthreads are attempting to communicate with dom0 at once, then the outer call to _get_session() can end up with all of the sessions acquired leaving the inner calls to _get_session() waiting for one of the outer calls to release.

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.openstack.org/3631

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/3631
Committed: http://github.com/openstack/nova/commit/093e4d38d511c7bb8d48fceebbfd8e350b533075
Submitter: Jenkins
Branch: master

commit 093e4d38d511c7bb8d48fceebbfd8e350b533075
Author: Johannes Erdfelt <email address hidden>
Date: Wed Feb 1 17:07:16 2012 +0000

    Make sure multiple calls to _get_session() aren't nested

    Fixes bug 924918

    async_call_plugin() acquires a xenapi session as does the nested call to
    get_xenapi_host(). This can cause a deadlock if multiple greenthreads
    all block waiting for the outer sessions to be freed to allocate the
    inner session. This change moves the call to get_xenapi_host() to outside
    the with statement to ensure calls to _get_session() aren't nested.

    Change-Id: I8f5490f40a9ccaf74a276187f66519a5d5f52b2e