[pve-devel] NACK: [PATCH storage 2/3] krbd: remove 'exclusive-lock' from blacklist, kernel-4.13

[Alwin Antreich]

On Thu, Mar 01, 2018 at 10:59:15AM +0100, Fabian Grünbichler wrote:
> this is problematic because it potentially hides bugs in our application
> logic, with currently no benefit.
>
> with exclusive-locks disabled, mapping on multiple hosts is possible,
> but mounting the same image is not (e.g., when attempting to mount, all
> but the first successful node will fail).
>
> with exclusive-locks enabled, mapping and accessing/mounting is
> sometimes possible (it seems a bit racy?), but since ext4 is not a
> cluster FS, this will cause undesired behaviour / inconsistencies /
> corruption.
>
> OTOH, with exclusive-locks enabled we would have the option of exclusive
> mapping - if we find a way to make this work with Qemu live-migration it
> might solve all of our problems
>
> TL;DR: I think we should postpone this pending further investigations
> into potential pros and cons
>
I did some testing on this topic. Qemu dies when live migrating, as rbd
can not get the lock on the new destination. A offline migration is
possible.

I guess, if we want to use it, then either we get qemu to work
differntly on migration or only allow offline migration when
exclusive mapping is requested.

I am up for discussion.