On Thu, Feb 24, 2022 at 6:10 PM Muli Ben-Yehuda <muli@lightbitslabs.com> wrote:
>
> On Thu, Feb 24, 2022 at 3:58 PM Nir Soffer <nsoffer@redhat.com> wrote:
>>
>> On Wed, Feb 23, 2022 at 6:24 PM Muli Ben-Yehuda <muli@lightbitslabs.com> wrote:
>> >
>> > Thanks for the detailed instructions, Nir. I'm going to scrounge up some hardware.
>> > By the way, if anyone else would like to work on NVMe/TCP support, for NVMe/TCP target you can either use Lightbits (talk to me offline for details) or use the upstream Linux NVMe/TCP target. Lightbits is a clustered storage system while upstream is a single target, but the client side should be close enough for vdsm/ovirt purposes.
>>
>> I played with NVMe/TCP a little bit, using qemu to create a virtual
>> NVMe disk, and export
>> it using the kernel on one VM, and consume it on another VM.
>> https://futurewei-cloud.github.io/ARM-Datacenter/qemu/nvme-of-tcp-vms/
>>
>> One question about device naming - do we always get the same name of the
>> device in all hosts?
>
>
> No, we do not, see below how we handle migration in os_brick.
>
>> To support VM migration, every device must have unique name in the cluster.
>> With multipath we always have unique name, since we disable "friendly names",
>> so we always have:
>>
>>     /dev/mapper/{wwid}
>>
>> With rbd we also do not use /dev/rbdN but a unique path:
>>
>>     /dev/rbd/poolname/volume-vol-id
>>
>> How do we ensure cluster-unique device path? If os_brick does not handle it, we
>> can to do in ovirt, for example:
>>
>>     /run/vdsm/mangedvolumes/{uuid} -> /dev/nvme7n42
>>
>> but I think this should be handled in cinderlib, since openstack have
>> the same problem with migration.
>
>
> Indeed. Both the Lightbits LightOS connector and the nvmeof connector do this through the target provided namespace (LUN) UUID. After connecting to the target, the connectors wait for the local friendly-named device file that has the right UUID to show up, and then return the friendly name. So different hosts will have different friendly names, but the VMs will be attached to the right namespace since we return the friendly name on the current host that has the right UUID. Does this also work for you?

It will not work for oVirt.

Migration in oVirt works like this:

1. Attach disks to destination host
2. Send VM XML from source host to destination host, and start the
   VM is paused mode
3. Start the migration on the source host
4. When migration is done, start the CPU on the destination host
5. Detach the disks from the source

This will break in step 2, since the source xml refer to nvme device
that does not exist or already used by another VM.

To make this work, the VM XML must use the same path, existing on
both hosts.

The issue can be solved by libvirt hook updating the paths before qemu
is started on the destination, but I think the right way to handle this is to
have the same path.