Name/label/id metadata: how do I make it go away

Paul Kraus paul at kraus-haus.org
Wed Sep 23 17:07:55 UTC 2015


On Sep 22, 2015, at 17:43, Chris Stankevitz <chris at stankevitz.com> wrote:

> And if I want to dig deeper into root cause I can ask ZFS "why do you sometimes select from the consumer collection and sometimes from the provider collection when putting a pool together".  Or if I don't want to dig deeper I can "deal with it" or I can disable diskid using kern.geom.label.disk_ident.enable

Assuming 10.x, on boot the ZFS code scans all attached devices and attempts to reassemble any zpools that it finds that were owned by this system. I suspect which device ZFS uses is based on which it scans (and finds) first.

I actually prefer the /dev/diskid/nnn names as those are tied to the physical drives. By using them I guarantee that even if a drive physically (or logically if drives or controllers are added) moves the system can still find and import the zpool. In the early days of ZFS one of the best ways to damage a zpool was to rearrange drives so that the ZFS label (and cache) no longer agreed with reality. I was in the habit of manually exporting critical zpools before making any hardware changes and after the changes were complete I would import the pool (sometime with new device names). ZFS _should_ be robust enough to handle device movement today, but I am slightly paranoid when it comes to critical data.

--
Paul Kraus
paul at kraus-haus.org



More information about the freebsd-questions mailing list