On Wed, 2014-04-02 at 09:51 +0300, Sagi Grimberg wrote:
> On 4/1/2014 8:45 PM, Nicholas A. Bellinger wrote:
> > On Tue, 2014-04-01 at 20:27 +0300, sagi grimberg wrote:
> >> On 4/1/2014 8:09 PM, Martin K. Petersen wrote:
> >>>>>>>> "Sagi" == Sagi Grimberg <sagig@xxxxxxxxxxxxxxxxxx> writes:
> >>> Sagi> I originally wrote the code to support that. But it got left
> >>> Sagi> behind since I figured it is not an interesting use-case. If your
> >>> Sagi> beckend doesn't support T10-PI why should the target publish it
> >>> Sagi> support it and ask the device to strip/insert it? I suppose it is
> >>> Sagi> to allow the initiator to protect half-way, but I don't know how
> >>> Sagi> interesting it is if the data is not stored with protection...
> >>>
> >>> That depends what you do on the backend. There are several devices out
> >>> there that expose PI to the host but use a different protection scheme
> >>> internally. And then synthesize PI on the host-facing side. Some even do
> >>> T10 PI to an internal protection scheme and then back to T10 PI when
> >>> talking to the disk drives in the back end.
> >>>
> >> Hey Martin,
> >>
> >> I understand, but even for internal different T10-PI schemes, is
> >> stripping protection from incoming data
> >> at the fabric level (and then do whatever with it in the backend level)
> >> the right thing to do here?
> >> I mean we basically lose protection across the PCI with this scheme
> >> aren't we?
> >>
> > The WRITE_STRIP + READ_INSERT case would be still be useful for IBLOCK
> > backends that don't support real hw PI, so that at least the protection
> > can be in place for data movement between physical machines.
> >
> > Also, I think the amount of changes required to support this type of
> > configuration in target-core are quite small.
>
> So trying to understand how this will come to use.
> Target will publish the fabric T10-PI support based only on the
> transport configuration (not accounting the backing devices configuration).
Yes, passing in the transport configuration for PI at
transport_init_session() time seems to make the most sense here in order
to address all fabric types.
> Then upon each cmd the target will look on {backstore configuration,
> PROTECT bit, transport configuration} - then will decide on protection
> operation (STRIP/INSERT/PASS).
>
> Looks right?
>
Correct.
I'm thinking it makes sense for target-core to perform the WRITE_INSERT
+ READ_STRIP (in software) when the transport does not directly support
PI, but the backend has PI enabled.
--nab
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
