On Tue, Sep 10, 2002 at 01:47:48PM -0500, Steve Lord wrote:
> Hmm, the code I am thinking of went on August 16th, so probably you
> already had it.
Yes, I'm pretty sure I would have. I will check though.
> So one open for the raid resync, and one for the mount - both of
> those are outside XFS's control.
It seems for any other FS other than / you can remount the
raid-superblock RO with "ref == 1" as the reference that is open is
the process calling the ioctl to remount RO. For / fs, you always
have a second reference and this screws you.
I'm actually not sure how / is supposed to work on RAID so perhaps I
have this wrong.
> The attempt to replay the log by making the filesystem writable in
> the middle of this operation is probably what gets thrown here. We
> remount read/write, run recovery, then remount readonly again,
> finally user space does its own remount read/write.
Possibly, but this should get released again making XFS no different
than any other fs... so I'm not sure XFS is to blame.
I will try to get call-chains from *all* calls to the open on /dev/md0
and see what falls out. I should probably also get release
information too :)