GEOM_Journal - "Taste"-order

[Miroslav Lachman]

Andrea Brancatelli wrote:
> Hello everybody.
>
> I'm facing a funny issue with GEOM_Journal and an ASUS MB with integrated
> Raid.
>
> For reason I cannot clearly understand, after creating an hardware mirror,
> the "standard" devices (ad0 and ad4 if I can remember correctly) didn't
> "disappear" and only a new device, ar0 appeared, representing the HW Mirror.
>
> I then geom_journaled a partition inside ar0, inserted
> geom_journal_load="YES" in the usual places but when the machine booted
> geom_journal found the journaling on ad0 instead of ar0 -- and actually
> started it!
>
> Obviously fstab was referring to ar0 and thus the boot stopped.
>
> I tried some different approaches but none worked. The only convenient way
> to make everything work was to disable geom_journal_load and have a script
> run in the end of the boot doing a gjournal load and a manual mount -- but
> this obviously sucks.
>
> I tried searching for a way to "disable" tasting for ad0 or ad4 or change
> the order of the devices so that ar0 would be tasted first, but could not
> find anything related.
>
> Does anyone have any suggestion?

It is the same issue as discussed in thread: it's a race between gmirror 
and UFS labels
http://lists.freebsd.org/pipermail/freebsd-geom/2010-September/004381.html

AFAIK there is no solution. GEOM classes are "buggy" in several things.
1] gmirror is dropping disks instead of holding them as "broken"
2] taste ordering priority
3] getting GPT even if it should be considered as "broken" (if gpt is on 
dropped disk from gmirror - and it can be easily detected, because there 
is known gmirror metadata in place where gpt/gpart is looking for 
secondary table)
4] impossibility of stopping gjournal tasting for some device (on 
systems with more then one journal, one cannot prevent loading of 
journal just for one device, gjournal must be stopped on all and kernel 
modul unloaded) This is problem for iSCSI devices (I posted about this 
issue 9 month ago)

So some of the "nice" features (as mounting by labels instead of device 
names) can't be used seriously in production, because it is too risky in 
the case of disk failure in gmirror or in your case with ar0 RAID.

Miroslav Lachman