The problem is that the partition naming option for the kpartx invocation was changed between RHEL4/RHEL5 from the default heuristic mode (adds a 'p' only if the partition name ends in a letter) to a mode which always adds the p, regardless of the base device name:

$ grep 'kpartx -' RHEL-4/initscripts-*/rc.d/rc.sysinit 
           /sbin/dmsetup ls --target multipath --exec "/sbin/kpartx -a"
           /sbin/dmsetup ls --target multipath --exec "/sbin/kpartx -a"
           /sbin/dmsetup ls --target multipath --exec "/sbin/kpartx -a"

$ grep 'kpartx -' RHEL-5/initscripts-*/rc.d/rc.sysinit 
           /sbin/dmsetup ls --target multipath --exec "/sbin/kpartx -a -p p"
           /sbin/kpartx -a -p p "/dev/mapper/$dmname"

This is unfortunate as quite a few tools (fdisk included) use this same heuristic when constructing partition names:

        /* Heuristic: we list partition 3 of /dev/foo as /dev/foo3,
           but if the device name ends in a digit, say /dev/foo1,
           then the partition is called /dev/foo1p3. */
        w = strlen(disk_device);
        if (w && isdigit(disk_device[w-1]))
                w++;

I'm don't think util-linux is the correct component for this BZ; the problem was introduced by this change to initscripts:

$ git show 1db1dba1
commit 1db1dba1853f6a3d45b10552ab167f60e156f403
Author: Peter Jones <pjones>
Date:   Thu Aug 31 22:21:28 2006 +0000

   - make multipath partitions be named with the form "${device}p${partnum}",
     like in nash and parted and anaconda and md partitions and...
   - make dmraid partitions be named with "p" as well (and use kpartx)

   diff --git a/rc.d/rc.sysinit b/rc.d/rc.sysinit
   index a1c3e6b..00c105f 100755
   --- a/rc.d/rc.sysinit
   +++ b/rc.d/rc.sysinit
   @@ -411,17 +411,19 @@ if [ -c /dev/mapper/control ]; then
                   modprobe dm-multipath > /dev/null 2>&1
                   /sbin/multipath.static -v 0
                   if [ -x /sbin/kpartx ]; then
   -                       /sbin/dmsetup ls --target multipath --exec "/sbin/kpartx -a"
   +                       /sbin/dmsetup ls --target multipath --exec "/sbin/kpartx -a -p p"
                   fi
           fi
           
           if [ -x /sbin/dmraid ]; then
                   modprobe dm-mirror >/dev/null 2>&1
   -               for x in $(/sbin/dmraid -ay -t --ignorelocking 2>/dev/null | egrep -iv "^[nN]o " | awk -F ':' '{ print $1 }') ; do
   +               for x in $(/sbin/dmraid -ay -i -p -t 2>/dev/null |
   +                               egrep -iv "^no " |
   +                               awk -F ':' '{ print $1 }') ; do
                           dmname=$(resolve_dm_name $x)
   -                       if [ -z "$dmname" ]; then
   -                               /sbin/dmraid -ay --ignorelocking "$x"
   -                       fi
   +                       [ -n "$dmname" ] && continue
   +                       /sbin/dmraid -ay -i -p "$dmname" >/dev/null 2>&1
   +                       /sbin/kpartx -a -p p "/dev/mapper/$dmname"
                   done
           fi
       
I am not entirely convinced by this logic. There's a simple reason why md
devices always have the p - they always end with a numeric character. Same
thing for cciss and the other device types that use this convention. Ordinary
devices like IDE, SCSI etc. follow the same convention - no 'p' since their
base names don't end in numeric characters.

Changing fdisk to use this logic doesn't seem like a good idea as it will cause the wrong names to be shown for devices that were created with the default kpartx options (e.g. user manually running "kpartx -a" to map partitions on a VM's disk image).

This naming paradigm existed in Fedora for over a year and a half before RHEL 5 was released, and has existed in RHEL 5 GA, 5.1, 5.2, and 5.3.

It's too late to change it; I think the tools need to adapt.

The "if last character is numeric add p" convention is much older than that but it's a moot point.

Any suggestions on how the tools should adapt? I couldn't think of a heuristic for fdisk that will give the right behaviour under all situations short of ripping out the offending bit of code completely and actually going and looking in the file system to determine the right name for a given device.

Maybe that's the right way to go?

Actually that doesn't work for a partition table that has been modified on-disk but that the kernel/kpartx has not yet re-read - "fdisk -l" needs to make up a name for the partitions that matches what would be done if we re-read the partition table.

Might want to check to see what parted does, and if it suffers from the same problem.

It shouldn't be a problem to use "p<N>" for all /dev/mpath/* devices. We (upstream) already have such heuristic for udev devices (/dev/disk/by-*) where the naming paradigm is "-part<N>" (IMHO this paradigm is the best one).

(In reply to comment #6)
> Might want to check to see what parted does, and if it suffers from the same
> problem.  

I have been working to add support for dmraid devices in GParted.
See GParted Bug: http://bugzilla.gnome.org/show_bug.cgi?id=317262

I can confirm from my testing that the current git version of parted uses the "if last character of device name is numeric then insert 'p' before the partition number" convention.

In my humble opinion, it would be best if dmraid also followed this convention.

An advisory has been issued which should help the problem
described in this bug report. This report is therefore being
closed with a resolution of ERRATA. For more information
on therefore solution and/or where to find the updated files,
please follow the link below. You may reopen this bug report
if the solution does not work for you.

http://rhn.redhat.com/errata/RHBA-2011-0085.html