On Apr 20 12:04, Klaus Jensen wrote: > On Apr 20 12:03, Dmitry Tikhov wrote: > > Current implementation have two problems: > > First in the read part of copy command. Because there is no metadata > > mangling before nvme_dif_check invocation, reftag error is thrown for > > blocks of namespace that have not been previously written to. > > Yes, this is definitely a bug and the fix is good, thanks! > > > Second in the write part. Reftag in the protection information section > > of the source metadata should not be copied as is to the destination. > > Hmm, says who? > > > Source range start lba and destination range start lba could differ so > > recalculation of reftag is always needed. > > > > If PRACT is 0, we really should not touch the protection information. My > interpretation of the Copy command is that it is simply just screwed if > used with PRACT 0 and Type 1. PRACT bit is specifically to allow the > controller to generate appropriate PI in this case. > > On the other hand, I can totally see your interpretation as valid as > well. Let me ask some spec people about this, and I will get back to > you.
Discussed this with the TP authors and, no, reftag should not be re-computed for PRACT 0, regardless of the PI type.
signature.asc
Description: PGP signature
