[sheepdog] [PATCH v1] sheep/md: leave cluster if disk number equals zero in handling EIO

[Valerio Pachera]

2014-09-18 9:04 GMT+02:00 Hitoshi Mitake <mitake.hitoshi at lab.ntt.co.jp>:
> Thanks for your reporting. BTW, I don't like the current way of
> handling -EIO (not limited to md, and not works well
> currently). Because if sheep has no disks, it means it cannot store
> log. It will make trouble shooting almost impossible. Rather than
> calling leave_cluster(), simply calling panic() or exit() would be
> enough. How do you think, Valerio and Bingpeng?

I was considering this scenario:
  sheep /var/lib/sheepdog,/mnt/sheep/0
being /mnt/sheep/0 the mount point of a different device.[1]
In such case, it's always safe to unplug a disk (logs and metadata are
elsewhere).
Metadata dir should never be on the same device where objects are stored.
IMHO, instead of a 'good practise', this should be forced requiring at
least 2 folders in sheep syntax (and even better checking by 'stat'
that they are not on the same device).
[1] it may be the same physical disk, but another partition.

In the case of
  sheep /var/lib/sheepdog
it's right to exit the cluster, because we have a single device and
it's giving EIO.
Metadata and log may have problem being written, so better exiting the cluster.