A device's physical path describes its physical location, as distinct from its logical or electrical location. Currently only ses(4) provides physical paths to da(4) disks, and only zfsd(8) consumes them. All current GEOM transforms either present blank physical paths, or pass through the underlying provider's physical paths. This is a problem for zfsd(8), because da0, da0p1, and da0p1.eli will all have the same physical path, so zfsd(8) doesn't know which to use in an autoreplace operation. I propose the following two rules for all disk-like devices: 1) The physical path should only be nonempty for providers that have a distinct physical location. For example, memory disks, mirrored pairs, and iSCSI LUNs should not have physical paths. 2) No two providers should have identical physical paths unless they both provide exactly the same data. Many GEOM transforms currently violate this rule, for example geli and gpart. gsched is an unusual exception, because both X and X.sched do provide exactly the same data.
This is done in 12. Now it's just waiting on the child bugs to MFC.