Open Stack

On Tue, Jul 17, 2018 at 04:06:29PM -0300, Erlon Cruz wrote:
> Hi Cinder and Nova folks,
> 
> Working on some tests for our drivers, I stumbled upon this tempest test
> 'force_detach_volume'
> that is calling Cinder API passing a 'None' connector. At the time this was
> added several CIs
> went down, and people started discussing whether this (accepting/sending a
> None connector)
> would be the proper behavior for what is expected to a driver to do[1]. So,
> some of CIs started
> just skipping that test[2][3][4] and others implemented fixes that made the
> driver to disconnected
> the volume from all hosts if a None connector was received[5][6][7].

Right, it was determined the correct behavior for this was to disconnect the
volume from all hosts. The CIs that are skipping this test should stop doing so
(once their drivers are fixed of course).

> 
> While implementing this fix seems to be straightforward, I feel that just
> removing the volume
> from all hosts is not the correct thing to do mainly considering that we
> can have multi-attach.
> 

I don't think multiattach makes a difference here. Someone is forcibly
detaching the volume and not specifying an individual connection. So based on
that, Cinder should be removing any connections, whether that is to one or
several hosts.

> So, my questions are: What is the best way to fix this problem? Should
> Cinder API continue to
> accept detachments with None connectors? If, so, what would be the effects
> on other Nova
> attachments for the same volume? Is there any side effect if the volume is
> not multi-attached?
> 
> Additionally to this thread here, I should bring this topic to tomorrow's
> Cinder's meeting,
> so please join if you have something to share.
> 

+1 - good plan.