On Fri, 2014-12-12 at 17:53 -0800, Nicholas A. Bellinger wrote: > On Sat, 2014-12-13 at 00:52 +0100, James Bottomley wrote: > > On Fri, 2014-12-12 at 14:59 -0800, Nicholas A. Bellinger wrote: > > > Hello T10/SCSI folks, > > > > > > A question came up on the LIO target-devel mailing list recently wrt the > > > proper handling of Persistent Reservations + All Registrants logic in a > > > couple of different scenarios. > > > > I can tell you how it was originally designed over a decade ago to > > support clustering products (being in the business then of negotiating > > with the manufacturers). > > > > > The crux of the confusion stems from how registrations are treated > > > before an explicit PR-OUT RESERVE service action with All Registrants > > > type occurs, and after that explicit PR-OUT RESERVE of All Registrants > > > type has been released. > > > > > > My original interpretation was that an explicit PR-OUT RESERVE with All > > > Registrants type had to first occur before other non reservation holding > > > registrations (existing or not) would gain the implicit reservation. > > > > You seem to be asking in the presence of registrations would an > > unregistered initiator receive a reservation conflict before the PR_OUT > > RESERVE is issued? The answer, categorially is no because there's no > > reservation until a registered initiator establishes one. > > > > Not exactly. The question is if a RESERVE with type=All Registrants > still needs to happen in order for the other registrations to have their > 'implicit' reservation activated. > > As mentioned below, the reasoning was because I'd only ever seen RESERVE > populate the reservation TYPE field (eg: not REGISTER), which would > imply that a 'explicit' RESERVE needs to happen from one initiator in > order for the target to know if All Registrants logic needs to kick in > for other existing + future registrations. I'm not sure I can be more clear: You can't have a reservation without and initiator issuing a reserve command. The TYPE parameter is ignored for REGISTER (specified in table 116). > > > This was primarily because the reservation type was not known until the > > > explicit PR-OUT RESERVE occurred, and AFAICT at the time there where no > > > host implementations that actually specified the reservation type at > > > registration time. > > > > > > So Is it correct to assume that one explicit PR-OUT RESERVE with All > > > Registrants type needs to occur before other non reservation holding > > > registrations gain an implicit reservation..? > > > > I don't understand what you mean implicit reservation. Once an all > > registrants reservation is established, *all* registered initiators > > become the reservation holder ... they're all equal. This is mandated > > in SPC-3 5.6.9 > > Implicit reservation meaning a registration for an I_T nexus to a device > with a different I_T nexus holding an All Registrants reservation. > > Eg, an I_T Nexus issued a RESERVE with type=All Registrants, and now the > other I_T nexus with an active registration on the same device have an > implicit reservation. There's no such distinction in the standards; all registered initators become the reservation holder once an all registrants reservation is issued. > > > How about when the explicit reservation is removed..? Do the implicit > > > reservations need to remain in effect for the existing registrations, > > > even though the explicit reservation of All Registrants type no longer > > > exists..? > > > > Again, I don't understand the distinction between implicit and explicit. > > For an all registrants reservation, all registered initiators are the > > reservation holder so any one of them may release the reservation. Once > > the reservation is released, there's no reservation at all. > > So some initiator still needs to perform the actual RESERVE, right..? Yes, there's no reservation until a reserve is issued. > That is, a number of registrations don't become 'implicit' reservation > holders until one initiator performs an 'explicit' RESERVE with type=All > Registrants. Once one initiator issues an all registrants reservation, all registants become the reservation holder. > > > The second point that Ilias raised is what should be the correct > > > R_HOLDER bit status for PRIN READ_FULL_STATUS descriptors when All > > > Registrants type is specified. My original interpretation was that > > > R_HOLDER was to signal the explicit reservation holder obtained with > > > PR-OUT RESERVE, and not for the implicit reservations from the other > > > registrations. Is this the correct approach..? > > > > R_HOLDER is set for any initiator that would be defined to be the > > persistent reservation holder in section 5.6.9 > > > > Well, section 5.6.9 only mentions the use of RESERVE, so that would seem > to indicate the original interpretation + current code for > READ_FULL_STATUS is correct. It should be set for every registered initiator in the all registrants case (because they're all the reservation holder). > > > Also, does anyone have any experience with a host side implementation of > > > All Registrants to use as a reference for what's already out in the > > > wild..? > > > > I know what was negotiated to support clustering (which is a pretty tiny > > subset). However, I think your point of confusion is trying to > > distinguish between what you call implicit and explicit reservations ... > > there's no such thing. There's only a reservation holder which may be > > many initiators, So it's binary: a given initiator either is or is not > > the reservation holder. Section 5.6 needs to be read in that context. > > > > The confusion is around when registrations who don't issue an explicit > RESERVE become implicit reservation holders. Is it when one initiator > does an explicit RESERVE + type=All Registrants, or is it when REGISTER > + type=All Registrants occurs..? There is no reservation until a reserve command is issued. Once an all registrants reservation is issued, by any registered initator, all registered initators then become the reservation holder. James -- 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
