On Mon, Jun 04, 2012 at 05:02:26PM +0200, Stefan Behrens wrote: > > According to the kern.1.log file that you have sent (which is not > visible on the mailing list because it exceeded the 100,000 chars limit > of vger.kernel.org), a rebalance operation was active when the disks or > the RAID controller started to cause IO errors. > > There seems to be a bug! Like that a write failure is ignored in btrfs. > For instance, the result of barrier_all_devices() is ignored. Afterwards > the superblocks are written referencing trees which have not been > completely written to disk.
This may be also what happened when my hardware RAID blew up. I was left with two completely inconsistent/unusable btrfs which I am still attempting to recover. Assuming that the general mount options to remount read-only on errors are correctly handled by btrfs, that would seem to be the wise thing to do. IMO it seems a volume which experiences a metadata write error to the underlying medium should be made immediately read-only anyway. -- Ryan C. Underwood, <neme...@icequake.net>
signature.asc
Description: Digital signature