[libvirt-users] Storage Pools & nodedev-create specific scsi_host#?

[Osier Yang]

On 2012年06月29日 08:16, mdecheser at comcast.net wrote:
> Hello everyone,
>
> Current host build is RHEL 6.2, soon to be upgrading.
>
> I'm in the process of mapping out a KVM/RHEV topology.  I have questions about the landscape of storage pools and the instantiation of specific scsi_host IDs using virsh nodedev-create and some magic XML definitions.  I'm grouping these questions together because the answer to one may impact the other.
>
> High-level goals:
>
> - fence off I/O by providing NPIV paths to VMs individually
> - provide each VM with a pair of paths to the same storage for redundancy
>
> Problem 1 - Storage Pools:
>
> I've successfully created a storage pool by starting with an initial SCSI host based off one of my virtual HBAs.  I see my LUNs.  I performed a virsh pool-dumpxml and captured the configuration of the storage pool using just one SCSI host.  I then modified the XML file to include a 2nd SCSI host, stopped&  deleted the original storage pool, then created&  started a new storage pool.  In the GUI, I am able to see the same LUNs, listed twice (once for each path).

It sounds to me a bug. But I may be wrong as not farmiliar with
the scsi storage backend yet.

> Here's my XML file:
>
> <pool type='scsi'>
>    <name>VM1POOL</name>
>    <uuid>f0465f3f-3c9c-766c-3c96-678a99ec81cb</uuid>
>    <capacity>144813588480</capacity>
>    <allocation>144813588480</allocation>
>    <available>0</available>
>    <source>
>      <adapter name='host5'/>
>    </source>
>    <source>
>      <adapter name='host6'/>
>    </source>
>    <target>
>      <path>/dev/disk/by-path</path>
>      <permissions>
>        <mode>0700</mode>
>        <owner>-1</owner>
>        <group>-1</group>
>      </permissions>
>    </target>
> </pool>
>
> Question 1 - Is this the correct approach for presenting a VM with SAN storage in a redundant fashion?  Or should I instead use the mpath approach and find some way with the storage pool XML file to restrict which mpath devices are utilized (thereby restricting which LUNs the storage pool sees?).
>
> It seems my current approach provides each LUN twice, however a LUN has two different volume identifiers based on its path:
>
> UNIT:0:0:13
> UNIT:0:1:13
>
> Problem 2 - Specifying scsi_host# in a device XML file
>
> As I mentioned, I've successfully created a virtual HBA using virsh nodedev-create and an XML config.  Here's a sample config:
>
> [root at host ~]# cat vm1vhba1.xml
> <device>
>    <parent>scsi_host3</parent>
>    <capability type='scsi_host'>
>      <capability type='fc_host'>
>        <wwnn>a97a3b9220010002</wwnn>
>        <wwpn>a97a3b9220010001</wwpn>
>      </capability>
>      <capability type='vport_ops' />
>    </capability>
> </device>
>
> Pretty straightforward.  The parent is physical SCSI host 3.  The resulting SCSI host is 5.  This is one of the SCSI hosts used in the storage pool above.
>
> So, I wanted to test what would happen to my VM if I ran virsh nodedev-destroy scsi_host5.   Essentially, testing the redundancy of the paths to the VM.  The device was successfully deleted, however a few other things happened, both expected and unexpected.
>
> - Since I only associated one of the two LUN units from my storage
pool with the VM,I expected this to cause the VM to fail.  It did fail,
however in a headless fashion. Unexpected.  The VM is still up and running
in memory.  Any effort to write out to the disks results in failures like
"read-only filesystem".  Expected.

The default write error policy is "report", so that's expected you see
the error, while the VM is still running, even the device is destroyed.
And generally filesystem will be readonly if the underlying physical
device is detected by kernel as broken or something else.

You can set the error_policy of the vm to "stop" or "ignore". With
error_policy="stop", you will see the VM will be stopped in this case.

See http://libvirt.org/formatdomain.html#elementsDisks for more details.

> - The Storage Pool entered into an inactive state.  Unexpected.

Have you restarted or reload libvirtd? I have no other idea of what
trigers the pool refreshing otherwise. We don't integrate with things
like inotify with storage driver yet.

Efforts to restart the storage pool using virsh pool-start failed 
because one of the scsi_host devices was not present.  Expected.

Yes, that's expected.

>
> I tried to recreate the device using virsh nodedev-create vm1vhba1.xml, and the device was created successfully but as scsi_host9 (the next available scsi_host beyond the original scsi_host5).
>
> So, unless I can recreate this scsi_host5, I won't be able to restart my storage pool and restore I/O to my VM.

That's auto-numbered by scsi driver I think, not sure if there is way
to give a fix name for the new created vhba.

>
> Question 2:  Is there a definition I can use in the XML config to specify the resulting scsi_host #?
>
> I've been told that RHEL 6.3 is supposed to provide some methodology of auto-instantiating my virtual HBAs upon the reboot of a host so that the VMs on the will see their storage and can start automatically on boot.  I have a feeling that the answer to my question may be brushing up against this functionality.
>
> I am also assuming that if it's possible to specify the scsi_host# in my XML config, that I'll be able to restart my storage pool successfully and that will enable my VM to reestablish the connection to their storage.
>

As far as I known, no support for this yet.

Regards,
Osier