disk cachemodes should be restricted with multiattached volumes

[Eric Harney]

If using multiattach, the "writeback" and "unsafe" disk_cachemode options presumably break the semantics that an application writing to a clustered datastore would rely on for data consistency between multiple nodes.

Volumes should not be allowed to attach to multiple instances (multiattach) with unsafe cache modes.

(This may even include writethrough? I'm not sure.)

[Matt Riedemann (mriedem)]

Have you tested this or just guessing that the libvirt driver in nova isn't doing the right thing? Because multiattach disks are always set to cache mode "none":

https://github.com/openstack/nova/blob/20bc0136d0665bafdcd379f19389a0a5ea7bf310/nova/virt/libvirt/driver.py#L423-L426

        # Shareable disks like for a multi-attach volume need to have the
        # driver cache disabled.
        if getattr(conf, 'shareable', False):
            conf.driver_cache = 'none'

[Sean McGinnis (sean-mcginnis)]

Description only refers to nova settings. I don't see anything here for Cinder, so I am going to close it as Invalid. If there is something, please reopen and provide more detail what Cinder is doing (or not doing) that it should.