quota_reserve headroom might be wrong if project_quotas != user_quotas

[Matt Riedemann]

Looking at this code:

https://github.com/openstack/nova/blob/2014.2.b3/nova/db/sqlalchemy/api.py#L3343

Notice the headroom variable is created based on usages and user_quotas:

        headroom = dict((res, user_quotas[res] -
                             (usages[res]['in_use'] + usages[res]['reserved']))
                        for res in user_quotas.keys())

but the usages variable is based on whether or not project_quotas == user_quotas:

        if project_quotas == user_quotas:
            usages = project_usages
        else:
            usages = user_usages

So it appears that headroom could be incorrect if project_quotas != user_quotas.

Looking at what uses headroom, the compute API uses this in the instance create and resize flows. For resize it's just using headroom to plug into an error message for the TooManyInstances exception.

In the create flow (_check_num_instances_quota) it's used for a bit more advanced logic with recursion.

We should probably just remove the headroom calculation from quota_reserve and let the caller figure it out and what needs to be done with it. It's also odd that this is happening in the DB API because it's dealing with instance quotas but maybe I'm not doing anything with instance quotas, maybe I'm doing things with security group or fixed IP quotas - so this code seems to be in the wrong place. Maybe it's just conveniently placed here given the other data already in scope from the database.

[Chris Behrens (cbehrens)]

Agree with fixing this by removing the headroom calculation from the db layer. I don't like the hardcoding of things like 'cores' and so forth in there related to this. I think returning current usage+reservations along with quotas is enough to let the caller to compute headroom itself on the resources it cares about, etc...

[Matt Riedemann (mriedem)]

Maybe user_quotas is used since that's the only thing checked when calculating headroom for cores and ram (flavor info), but still seems like this is a hack and should be done on the calling side rather than in the DB API.

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.openstack.org/121786

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/121786
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=0045f7d50cad345c26ae02394ad57716b270d268
Submitter: Jenkins
Branch: master

commit 0045f7d50cad345c26ae02394ad57716b270d268
Author: liyingjun <email address hidden>
Date: Sat Sep 6 09:10:31 2014 +0800

    Removing the headroom calculation from db layer

    It's odd that this is happening in the DB API because it's dealing
    with instance quotas but maybe I'm not doing anything with instance
    quotas, maybe I'm doing things with security group or fixed IP quotas.

    This patch moves the headroom calculation from db layer to the
    calling side.

    Change-Id: I8379471eea4ea7b114c2a235224f6fe61f8f981f
    Closes-bug: #1369696