gmirror regression from 7.1 to current? (was: Re: GEOM and moving to CURRENT from 7.1)

From: Ben Kaduk <minimarmot_at_gmail.com>
Date: Thu, 15 Jan 2009 21:54:59 -0500
On Tue, Jan 13, 2009 at 4:35 PM, Ben Kaduk <minimarmot_at_gmail.com> wrote:
> On Tue, Jan 13, 2009 at 4:11 PM, Randy Bush <randy_at_psg.com> wrote:
>> installed amd64 7.1 from cdrom
>> partitioned two sata drives to single partitions
>> labeled and gmirrored
>>
>> upgraded to 8-current
>> ad0 started falling off mirror
>> ad2 started reporting smart errors
>
> Interesting.  I have been seeing similar behavior when trying to
> update my 7.1-prerelase box to current, except that it seems to be
> random _which_ disk falls off the mirror.  I'm also seeing panics if I
> try to stress the (then-degraded) mirror with the current kernel.  The
> mirror is rock-solid with the 7.1-pre kernel.
>
> The panic is "initiate_write_inodeblock_ufs2: already started", and
> I've got some details of the other messages I've seen logged up here:
> http://stuff.mit.edu/afs/sipb.mit.edu/user/kaduk/freebsd/periphrasis/20090107/panic.txt


Actually, it looks like the panic is a red herring.  Having done a bit
more poking around, I am getting the machine in  a state where one of
the disks is
trying to rebuild, and the other one is marked as broken, so there is
basically nothing available from the disks, but it keeps trying to
find something.  I can boot into 7.1 and rebuild the array, and it
takes everything I throw at it when running the 7.1 kernel.

That would seem to make this a gmirror/ata_dma regression from 7.1,
unless you think that a weird label would cause a disk to be dropped
from the array on a later-than-first access (I'm mostly sure it's
not happening on the first access).

Any thoughts?

The ident.kernel.old (RELENG_7) and ident.kernel{.updated} files in
http://stuff.mit.edu/afs/sipb.mit.edu/user/kaduk/freebsd/periphrasis/20090107/
might be of interest ...


[it turns out I didn't actually send this when I meant to, which
allows me to include the statement that memtest86+ has completed an
entire pass without errors.  I'll leave it running overnight, just for
completeness, though.]

-Ben Kaduk
Received on Fri Jan 16 2009 - 01:55:00 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:39:40 UTC