[PATCH v2 2/3] block: support PI at non-zero offset within metadata

[Martin K. Petersen]

Keith,

> Besides the passthrough interface, though, the setup uses kernel block
> layer to write the partition tables. Upgrading from 6.8 -> 6.9 won't
> be able to read the partition table on these devices. I'm still not
> sure the best way to handle this, though.

That's why, despite Type 2 supporting offset ref tags, the kernel always
uses the LBA. This is to ensure that any application which doesn't have
prior knowledge of the contents of each block's PI can perform a read
without failing. Including the kernel itself when reading the partition
table. But it's not just a kernel problem, the BIOS also needs to be
able to read arbitrary blocks, at least at the beginning and end of the
device.

Many of the features of the original DIF spec were squarely aimed at use
inside storage arrays where there was a single entity, the array
firmware, performing all reads and writes.

With DIX, since the interface needed to be used by array firmware and
general purpose operating systems alike, we needed to add the all
required features to the interface to accommodate every use case. And
NVMe inherited this. But NVMe did not really capture which things are
suitable only in the "single application" scenario.

I'm not sure there's a good fix. Other than as a passthrough user, the
burden is on you to ensure that things can be read by somebody else.
Could be backup applications, virus scanners, system firmware, etc.

-- 
Martin K. Petersen	Oracle Linux Engineering