Re: [Xen-devel] [RFC PATCH] blkif.h: document scsi/0x12/0x83 node

[David Vrabel <david.vrabel@xxxxxxxxxx>]

On 22/03/16 14:10, Konrad Rzeszutek Wilk wrote:
> On Tue, Mar 22, 2016 at 01:41:43PM +0000, David Vrabel wrote:
>> On 22/03/16 12:55, Bob Liu wrote:
>>>
>>> On 03/17/2016 07:12 PM, Ian Jackson wrote:
>>>> David Vrabel writes ("Re: [Xen-devel] [RFC PATCH] blkif.h: document 
>>>> scsi/0x12/0x83 node"):
>>>>> On 16/03/16 13:59, Bob Liu wrote:
>>>>>> But we'd like to get the VPD information(of underlying storage device) 
>>>>>> also in Linux blkfront, even blkfront is not a SCSI device.
>>>>>
>>>>> Why does blkback/blkfront need to involved here?  This is just some
>>>>> xenstore keys that can be written by the toolstack and directly read by
>>>>> the relevant application in the guest.
>>>>
>>>
>>> They want a more generic way because the application may run on all kinds 
>>> of environment including baremetal.
>>> So they prefers to just call ioctl(SG_IO) against a storage device.
>>>
>>>> I'm getting rather a different picture here than at first.  Previously
>>>> I thought you had some 3rd-party application, not under your control,
>>>> which expected to see this VPD data.
>>>>
>>>> But now I think that you're saying the application is under your own
>>>> control.  I don't understand why synthetic VPD data is the best way to
>>>> give your application the information it needs.
>>>>
>>>> What is the application doing with this VPD data ?  I mean,
>>>> which specific application functions, and how do they depend on the
>>>> VPD data ?
>>>>
>>>
>>> From the feedbacks I just got, they do *not* want the details to be in 
>>> public.
>>
>> It is difficult to suggest how it should be done correctly without this
>> information.
> 
> Just think of it as a black box.

This isn't sufficient.

You are presenting a solution but have not properly described the
problem so no one can evaluate whether the solution is appropriate.

>> I also find it difficult to see a use case where running the storage
>> software in the guest (instead of in the backend) is sensible or desirable.
> 
> Are you suggesting that doing backend drivers is not sensible?

I do not understand your question.

>>> Anyway, I think this is not a block of this patch.
>>> In Windows PV block driver, we already use the same way to get the raw 
>>> INQUIRY data.
>>>  * The Windows PV block driver accepts ioctl(SG_IO).
>>>  * Then it reads this /scsi/0x12/0x83 node.
>>>  * Then return the raw INQURIY data back to ioctl.
>>>
>>> Since Linux guest also wants to do the same thing, let's making this 
>>> mechanism to be a generic interface!
>>> I'll post a patch adding ioctl(SG_IO) support to xen-blkfront together with 
>>> a updated version of this patch soon.
>>
>> I do not think this feature is generally useful outside of this
>> unspecified use case.  I do not think that supplying details about
>> underlying storage device (beyond generic properties) to guests is
>> sensible (e.g., what if the guest snapshot is restored on different
>> storage?).
> 
> The restore process (xl) can update the XenStore key with the new storage.

And how is this going to be communicated to the application using the data?

>> And thus I do not not think we should either: a) make this part of the
>> blkif ABI; or b) add support to xen-blkfront or xen-blkback.
> 
> It is already coded in Windows PV drivers so I am not following why
> codyfing this in the blkif.h is harmful?

My understanding was that this was a legacy hack that should be removed
(we do not currently make use of it).

David

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel