lilo + raid + kernel-2.4.x failure to boot

[linas@backlot.linas.org (Linas Vepstas)]

[-- Attachment #1: Type: text/plain, Size: 2336 bytes --]


Hi,

another zinger that I am sending to LKML because I don't know where else to send it ...

I've discovered a deadly combination of kernel & lilo (and raid).  This may be a pure
lilo bug, but I assume that the kernel+raid aids & abets the problem...:


I am running kernel-2.4.x.  Two ide hard drives, with partitions 1,5,6,7,8 in use.
The partitions on the two drives are mirrored using RAID-1 to create /dev/md1, /dev/md5,
/dev/md6, etc.  The root fs is on /dev/md1.  Thus, lilo.conf looks like:


boot=/dev/md1
map=/boot/map
install=/boot/boot.b
prompt
timeout=50
linear
default=linux

image=/boot/vmlinuz-2.4.2
        label=linux
        read-only
        root=/dev/md1

For nearly a year, this combo has worked just fine (running 2.3.99 back then).   

Just fine, that is, using the redhat-6.2 rpm for LILO, i.e. version 
lilo-0.21-15.i386.rpm  which reports itself to be: 

% /sbin/lilo -V
LILO version 21

Recently, this machine went over to debian-unstable from redhat: 

% dpkg -s lilo
Package: lilo
Status: install ok installed
Priority: important
Section: base
Installed-Size: 271
Maintainer: Russell Coker <russell@coker.com.au>
Version: 1:21.7-3
Depends: libc6 (>= 2.2.1-2), debconf (>= 0.2.26), logrotate

The debian version of lilo writes a boot sector that hangs hard for the above
kernel+raid+lilo.conf configuration: specifically: 

LIL-     after a reboot.   Needless to say, recovery was painful.  But I was able to verify
that the redhat lilo rpm always functioned correctly, and the debian-unstable dpkg always
hung in this way.  Although at one point, during my twisting & turning, I got the debian lilo
to get to only LI  before hanging.  I have no idea of what I did different to get to that as
opposed to LIL- 


BTW, I noticed that oddly, every time I ran lilo, and then ran lilo -q -v -v, it reported
different sector numbers for the kernel images.  This freaked me out at first, but I came to
accept it as normal: doesn't affect bootability.  But is this really w.a.d?  (I was
assuming, appearently erroneously, that lilo -q -v -v was reporting the physical location
of the kernel image on the disk; but since the numbers bounce around, that can't be right.
Or is this just weird bios head/cyl/sect math flakiness?)


--linas

 




[-- Attachment #2: Type: application/pgp-signature, Size: 232 bytes --]