Re: Bug#602533: parted sees useless /dev/cciss/ device

[Colin Watson]

On Fri, Nov 05, 2010 at 05:37:23PM +0000, Ian Jackson wrote:
> Package: parted
> Version: 1.8.8.git.2008.03.24-11.1
> 
> I'm running this version of parted running on an HP DL165 with
> linux-image-2.6.26-2-686-bigmem 2.6.26-25lenny1.  I have nothing
> connected to the CCISS controller; the disks are connected only to the
> ordinary SATA controller.
> 
> However, there is still a device node /dev/cciss/c0d0, without a
> corresponding entry in /proc/partitions.  parted sees this and
> complains about it.
> 
> Notably: parted -l prints:
> 
>   Error: /dev/cciss/c0d0: unrecognised disk label                           
> 
> Another consequence of this behaviour is that d-i prints this warning,
> interrupting automatic installations:
> 
>   Unable to determine geometry of file/device /dev/cciss/c0d0.  You should 
> not use Parted unless you REALLY know what you're doing!
>   Warning!
>     1. Ignore [*]
>     2. Cancel    
>   Prompt: '?' for help, default=1>
> 
> The device can be opened but looks empty:
> 
>   woodlouse:~# dd if=/dev/cciss/c0d0 of=t
>   0+0 records in
>   0+0 records out
>   0 bytes (0 B) copied, 0.000570254 s, 0.0 kB/s
>   woodlouse:~#
> 
> It would be nice if parted could be taught to ignore this spurious
> device.  Perhaps it should ignore all zero-length devices ?

I agree; there seems no point in complaining about this.  The problem
with throwing an exception is that it's difficult for partitioners such
as partman to programmatically distinguish between this kind of thing
and some more serious error that actually merits letting the user know.

Ian asked me on IRC why libparted didn't use /proc/partitions, which
doesn't list this device.  (It does, but only if /sys/block/ is not
present.)  I looked into the differences between /proc/partitions and
/sys/block/, and they are as follows:

  * /proc/partitions omits zero-length devices.
  * /proc/partitions omits non-partitionable removable devices.
  * /proc/partitions omits RAM block devices.
  * /proc/partitions is in a human-readable format, while /sys/block/ is
    machine-readable.

To a first approximation nobody cares about RAM block devices, and
libparted definitely should expose non-partitionable removable devices
because it's sometimes useful to be able to manipulate those from a
partitioning program which uses libparted for device discovery (e.g. you
might want to create a filesystem on the whole disk).  However, the
policy of /proc/partitions seems closer to what we want in the case of
zero-length devices, and therefore I suggest that we silently ignore
these.

I'll follow up to this mail with a patch.

Thanks,

-- 
Colin Watson                                       address@hidden