On Thu, Jun 28, 2001 at 04:46:02PM -0500, Roger Moore wrote:
>We've just updated our cluster to use XFS and have noticed a problem with
>using 'LABEL=xxx' in the fstab file.
1. fsck has a bug in the LABEL=xxx handling code. It behaves badly
(and non-deterministically) iff sizeof(/proc/partitions) > 1K.
1a. If you have a lot of partitions, and a big /proc/partitions, you could
be getting stomped by this. But it typically only affects N partitions,
where N = sizeof(/proc/partitions)/1024. That is, it pukes at each 1K
1b. This generally only happens on the initial mount. It's due to the
contents of /proc/partitions changing (getting bigger) in between the time
that fsck reads the first 1K and the next 1K from the file. The next 1K
block does not start, in terms of the file's contents, where the first 1K
block ended. The lines become misaligned.
2. Look at your /proc/partitions. Are there XFS volume labels in there?
If not, then that's your problem. fsck reads /proc/partitions to get the
label information. If they are there, then you might want to run fsck under
gdb to watch it go through the fstab file...
BTW The label code from mount is NOT the same code used in fsck (reuse?
whazzat?) but it is vulnerable to the same error condition as mentioned in 1
"Smart Tags? We don't need no steenking Smart Tags!"