[pnfs] Linux pNFS status meeting 09/03

[Rick McNeal]

On Sep 8, 2009, at 10:34 AM, J. Bruce Fields wrote:

> On Tue, Sep 08, 2009 at 07:28:19PM +0300, Boaz Harrosh wrote:
>> On 09/08/2009 06:59 PM, Rick McNeal wrote:
>>>
>>> The block layout specification requires a unique signature to be
>>> location within the data blocks of the device. I didn't want to  
>>> create
>>> some new process and looked for existing solutions. EFI was the
>>> logical choice since there is a unique signature found in the label.
>>> When I first looked at the specification I actually wanted to see it
>>> changed. I felt there was no reason to require a signature to be
>>> located within the data blocks, which could be accidently  
>>> overwritten,
>>> when a perfectly good signature is available via the SCSI INQUIRY  
>>> page
>>> 0x83 command.
>>>
>>
>> This is a serious short comings in the specification. I would even  
>> go on
>> to say, a show stopper. We should talk about this in Beckathon and  
>> push
>> for a standard update.
>>
>> I would suggest that the "SCSI INQUIRY page 0x83" should be an  
>> alternative
>> option to the on data-block signature and the server should be able  
>> to specify
>> the first (for these that have it so today) or the new "SCSI  
>> INQUIRY page 0x83".
>>
>> Also in objects-layout "SCSI INQUIRY page 0x83" is used as a  
>> system_id. I have
>> hacks in store for scsi-tgt that lets you specify that id in  
>> configuration or
>> command-line.
>>
>> Some of the more important pnfs-blocks I can envision is over  
>> iscsi_tcp with
>> of-the-shelf target systems. It looks like an ugly administrator  
>> nightmare to
>> have to hack up an iscsi-target-backend files just for these  
>> system_ids
>
> I thought one of the requirements for the block specification was that
> it be completely block-protocol-neutral?  (Not that I really  
> understand
> why.)
>

I never understood the block-protocol-neutral aspect myself.  
Transportation neutral I understand. Today there's ATA or SCSI. Even  
if a new block protocol was created certain SCSI features are a must  
if the storage is to be shared. Persistent locking, globally unique  
identifiers. The specification could state that a data block signature  
could be used or SCSI INQUIRY. Then if, and only if, a new block  
protocol emerges the specification could be amended.

> I also don't understand why the client administrator has to  
> individually
> connect to all the block devices, when I'd expect there to be common
> cases where the the MDS already knows addresses for them that the  
> client
> could use.
>

The administrators need to use iSNS. The MDS has no means to provide  
the target addresses to a client.

> --b.

----
Rick McNeal
rick.mcneal at lsi.com

“Never blame a legislative body for not doing something. When they do  
nothing, they don’t hurt anybody. When they do something is when they  
become dangerous.” —Will Rogers