Bug#422255: [etch] Installing with root on JFS in LVM fails

[David Härdeman <david@hardeman.nu>]

On Fri, May 11, 2007 1:38, Frans Pop said:
> The second was using an Etch netinst image, and there I can reliably
> reproduce the hangs you are seeing. They seem to occur during the
> unpacking of tarballs. At least, in both cases the last command that is
> visible in the output of 'ps' is "tar -xf -" (obviously a pipe from
> another command).
>
> I've seen two hangs (in different installs):
> - one at 31%; ps shows:
>      zcat /usr/lib/debootstrap/devices.tar.gz | tar -xf -
> - one at 28% "extracting tzdata"; no 'zcat' visible in ps, just:
>      tar -xf - (in state D)
>
> State "D" is "Uninterruptible sleep (usually IO)"; there is no processor
> activity; killing processes does not help, the system remains frozen.
>
> I've also tried a normal install (no crypto, no LVM) with / on jfs, and
> that succeeded without problems.
> However, with jfs on / using LVM without crypto, I can also reproduce the
> problem, so the fact that encryption was used is irrelevant. The issue
> seems to be with using jfs in LVM (or maybe jfs with device-mapper).
>
> It does look like we've got a kernel issue here in the 2.6.18 kernel that
> has apparently been fixed in 2.6.20.

Perhaps it's a kernel stack overflow issue? I remember there being a lot
of fixes going into the kernel to reduce stack usage when the stack was
reduced to 4kB and that many of the patches were against "stacked"
filesystems (e.g. anyfs-on-lvm-on-md-on-something).

Does dmesg show any interesting kernel messages after the installation
barfs? (like stack overflow warnings)

-- 
David Härdeman