Open Stack

Hi Chris, you are right – and I think there are several synchronization issues in the code there.

The migration process should be synchronized as it is in build_run_instance etc. At the moment if pre_live_migrate fails due to an RPC timeout (i.e. it takes too long for some reason) the rollback can be initiated while it is still running. So you can get volume attachments and vif plugging being built and torn down at the same time, not good. [I’m not sure if that has a bug report – I’ll file one if not].

We need to pay attention to these cases as we refactor the live migration process. This has begun with:
https://specs.openstack.org/openstack/nova-specs/specs/newton/approved/async-live-migration-rest-check.html
https://specs.openstack.org/openstack/nova-specs/specs/newton/approved/live_migration_compute_communcation.html

Paul

From: lương hữu tuấn [mailto:tuantuluong at gmail.com]
Sent: 10 June 2016 10:20
To: OpenStack Development Mailing List (not for usage questions) <openstack-dev at lists.openstack.org>
Subject: Re: [openstack-dev] [nova] theoretical race between live migration and resource audit?

Hi,

Yes, it is actually a race and we have already faced a negative effect when using evacuation. Some information of cpu pinning is lost. Imagine that, in some cases, we do some re-scheduling actions (evacuate, live-migration, etc.) then immediately do the next actions (delete, resize, etc.) before the resource_tracker updates in the next period. In this case, it fails. Actually, it has some negative side in writing tests based on the scenario described above.

Br,

Tutj

On Fri, Jun 10, 2016 at 10:36 AM, Matthew Booth <mbooth at redhat.com<mailto:mbooth at redhat.com>> wrote:
Yes, this is a race.

However, it's my understanding that this is 'ok'. The resource tracker doesn't claim to be 100% accurate at all times, right? Otherwise why would it update itself in a period task in the first place. It's my understanding that the resource tracker is basically a best effort cache, and that scheduling decisions can still fail at the host. The resource tracker will fix itself next time it runs via its periodic task.

Matt (not a scheduler person)

On Thu, Jun 9, 2016 at 10:41 PM, Chris Friesen <chris.friesen at windriver.com<mailto:chris.friesen at windriver.com>> wrote:
Hi,

I'm wondering if we might have a race between live migration and the resource audit.  I've included a few people on the receiver list that have worked directly with this code in the past.

In _update_available_resource() we have code that looks like this:

instances = objects.InstanceList.get_by_host_and_node()
self._update_usage_from_instances()
migrations = objects.MigrationList.get_in_progress_by_host_and_node()
self._update_usage_from_migrations()


In post_live_migration_at_destination() we do this (updating the host and node as well as the task state):
            instance.host = self.host
            instance.task_state = None
            instance.node = node_name
            instance.save(expected_task_state=task_states.MIGRATING)


And in _post_live_migration() we update the migration status to "completed":
        if migrate_data and migrate_data.get('migration'):
            migrate_data['migration'].status = 'completed'
            migrate_data['migration'].save()


Both of the latter routines are not serialized by the COMPUTE_RESOURCE_SEMAPHORE, so they can race relative to the code in _update_available_resource().


I'm wondering if we can have a situation like this:

1) migration in progress
2) We start running _update_available_resource() on destination, and we call instances = objects.InstanceList.get_by_host_and_node().  This will not return the migration, because it is not yet on the destination host.
3) The migration completes and we call post_live_migration_at_destination(), which sets the host/node/task_state on the instance.
4) In _update_available_resource() on destination, we call migrations = objects.MigrationList.get_in_progress_by_host_and_node().  This will return the migration for the instance in question, but when we run self._update_usage_from_migrations() the uuid will not be in "instances" and so we will use the instance from the newly-queried migration.  We will then ignore the instance because it is not in a "migrating" state.

Am I imagining things, or is there a race here?  If so, the negative effects would be that the resources of the migrating instance would be "lost", allowing a newly-scheduled instance to claim the same resources (PCI devices, pinned CPUs, etc.)

Chris

__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe<http://OpenStack-dev-request@lists.openstack.org?subject:unsubscribe>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev



--
Matthew Booth
Red Hat Engineering, Virtualisation Team

Phone: +442070094448<tel:%2B442070094448> (UK)


__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe<http://OpenStack-dev-request@lists.openstack.org?subject:unsubscribe>
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20160614/3d4efa0c/attachment.html>