Re: [LSF/VM TOPIC] Handling of invalid requests in virtual HBAs

[Vladislav Bolkhovitin <vst@vlnb.net>]

Nicholas A. Bellinger, on 04/14/2010 12:45 AM wrote:
> On Tue, 2010-04-13 at 23:23 +0400, Vladislav Bolkhovitin wrote:
>> Nicholas A. Bellinger, on 04/13/2010 10:37 PM wrote:
>>> On Tue, 2010-04-13 at 21:09 +0400, Vladislav Bolkhovitin wrote:
>>>> LIO doesn't support 1 to many pass-through devices sharing, so SCST in 
>>>> the only option.
>>> Sorry, but this statement about your perceived limitiations wrt TCM/LIO
>>> is completely incorrect.
>>>
>>> Using a single passthrough backstore device (eg: plain /dev/sdX) with
>>> TCM/pSCSI (or any TCM subsystem plugin) has been supported since the
>>> dawn of time to allow for any number of TCM_Loop Virtual SAS Ports with
>>> SG_IO going into KVM Guest.  The same is also true for LIO-Target
>>> (iSCSI) and TCM_FC (FCoE) ports as well regardless of TCM subsystem
>>> backstore.
>>>
>>> Perhaps you would be so kind to provide a TCM/LIO source code reference
>>> from where you came up with this make-believe notion..?
>> I've just rechecked with the latest 
>> git://git.kernel.org/pub/scm/linux/kernel/git/nab/lio-core-2.6.git and 
>> still wasn't able to find in LIO required pieces of functionality to 
>> support 1 to many pass-through. I see only code for non-enforced 1 to 1 
>> pass-through (single initiator only). You can use with it more 
>> initiators only as the SCSI violation, from pointing on which I started 
>> my participation in this discussion.
>>
> 
> Sorry, but listing the path to my git tree on k.o is not the same as
> citing a actual TCM/LIO source+line reference.

That's interesting. How can I point to the code which doesn't exist?

>> Particularly, I can't see the code, which in pass-through mode upon 
>> receive of RESERVE command sends it also to the backend device, if 
>> necessary (i.e. only the first time), and then sends RELEASE command to 
>> the device upon the reservation holder removal. Could you point me on 
>> the *exact* code which implements that?
>>
> 
> Again, since *you* are the one making a claim, *you* are the one
> expected to back it up with a concrete code reference and example
> scenario.   Please, do the leg-work yourself wrt TCM SPC-3 persisetent
> reservations and legacy SPC-2 reservations instead of expecting me to do
> the actual work for you to verify your own fanciful claims about target
> mode, seriously..
> 
> So, short of you being able to produce a response that is concrete and
> human-readable, your claim will once again be dismissed as generic
> hand-waving.

Well, this is an Open Source and in Open Source code auditing is the 
most important way to verify an implementation. I reviewed TCM/LIO and 
my conclusion is pretty clear: it misses too many important bits to be a 
1 to many pass-through implementation, so it implements only 
not-enforced 1 to 1 pass-through. You implemented reservations 
emulation, but that's _NOT_ enough. I gave an example of the one among 
many missed functionality.

I'm not going to go any deeper than I've already gone doing your home 
work for you and explaining you the SCSI basics (although for me it's 
really hard to believe that a person implemented something as big as 
TCM/LIO can't see such obvious things), because when I did it before, 
when you asserted that if a pass-trough device supports Persistent 
Reservations, LIO in pass-through mode with this device automatically 
supports them too, which is obviously wrong (see the end of 
http://lkml.org/lkml/2008/7/14/273), I didn't have even a bit of 
appreciation for my effort. Instead I received privately from you 
insulting and threatening e-mails.

If you are not agree with my conclusion about pass-through 
implementation in TCM/LIO, you can either point us on the code 
implementing the necessary functionality, which I missed, or argue why 
it isn't needed. But if you continue attacking me personally trying to 
discredit me, everybody will see that my conclusions are definitely 
correct as well as which is your preferred way of cooperation.

Vlad