Hi there, I'm trying to solve a long-standing problem with installation
of CentOS & Fedora on our nForce-430-based motherboards.
The forcedeth kernel module has been pretty fast-moving over the last
few kernels and to support our current boards, we need a really fresh
version which is generally only found in 2.6.19 or 2.6.20. nVidia
backport these drivers to earlier kernels here:
http://www.nvidia.com/object/linux_nforce_1.21.html
Since we can only perform network network installs, I am trying to work
out a general way of retro-fitting the lastest nForce drivers into a
given Fedora Core or CentOS install, and I understand the anaconda
loader program is the key to this.
Unfortunately "use a newer distro" isn't an option in a few cases where
customers want a specific Fedora version for support of some random
commercial software - right now I am struggling to get FC5 working, but
it would be nice if this would fix Centos3/4 which are popularly
requested as well.
Also despite FC5 being listed as being included in nVidia's binaries,
they don't provide a driver disc image. They claim that the FC5 driver
should support installation at least, and the RPMs are just to bring it
bang up to date after installation. This isn't true on our newest
motherboards, it just doesn't seem to spot that the network chip exists.
What I've tried so far is to extract the stock pxeboot/initrd.img and
the contained modules/modules.cgz files. Then I add in the .ko files
from the RPMs that nVidia supply for FC5, then rebuild it all with cpio
/ gzip. Unfortunately anaconda still says it doesn't have a driver
"for the specified installation type", even though I'm 99% sure that the
forcedeth.ko file bring up an eth0 (not sure how to launch a shell from
the loader to check ;) ). I even tried copying the FC5 x86_64/os
directory onto a local drive, then re-launched the network installation
and asked it to install off a local hard drive. Again the "no driver
found for specified installation type" complaint, even though it offered
to load a driver disk off the filesystem on /dev/sda1 which it could
clearly access ;-)
So from that last problem, I guessed that even if I've put some working
driver modules in there, I needed to add some PCI IDs to a list or two.
At this point I'm guessing, but I found the modules/pci.ids file and
modules/modules.aliases files in the initrd. I've tried altering these
to contain information about the nForce drivers (10de:03ef for the
network chip, and 10de:03f6 for the SATA chip) but I still see the "no
driver found for specified installtion type" for both hard disc and
network installs.
I've attached my initrd-rebuilding script which I've used for my
experiments - can anyone with more expertise on anaconda advise on a
next step? Is "slipstreaming" kernel modules possible in this way, or
is there a smarter way of going about what I want to do?
Thanks,
--
Matthew
#!/usr/bin/ruby
#
# This script rebuilds the given redhat PXE initrd image with modules found
# in RPMs. This is useful to update installers with newer kernel modules,
# whether the manufacturer has supplied a driver disk or not - as long as
# there are some RPMs, you can use them to rebuild and run the installer.
# Save some typing
#
require 'find'
require 'fileutils'
include FileUtils
def xsys(cmd)
return if system(cmd)
STDERR.print "*** Error running: #{cmd}\n"
exit $?.exitstatus
end
class Dir
def self.mkdir_and_work_in(d)
mkdir(d) unless File.directory?(d); prev = Dir.pwd; chdir(d); yield; chdir(prev)
end
end
MODULE_PATH = /\.ko$/
# Sanity check on arguments
#
initrd, initrd_out, *rpms = *ARGV
if ARGV.length < 3 || [initrd, *rpms].select { |f| !File.exists?(f) }.length > 0
STDERR.print "Syntax: #{$0} <original initrd.img> <new initrd.img> <rpm 1> [rpm 2 ... ]\n"
exit 1
end
if File.exists?(initrd_out)
STDERR.print "Won't overwrite #{initrd_out}, please move it\n"
exit 2
end
# Do our work in a temporary directory
#
Dir.mkdir("/tmp/rebuild.#{$$}")
#at_exit { rmtree "/tmp/rebuild.#{$$}" }
Dir.chdir("/tmp/rebuild.#{$$}")
# Extract initrd
#
Dir.mkdir_and_work_in("x_initrd") do
xsys("gunzip <#{initrd} | cpio -id")
end
# Extract modules inside initrd, find correct path to put our extra modules
#
Dir.mkdir_and_work_in("x_modules") do
xsys("gunzip <../x_initrd/modules/modules.cgz | cpio -id")
end
module_path = nil
Find.find("x_modules") do |file|
next unless MODULE_PATH.match(file)
module_path = file.split("/")[0..-2].join("/")
break
end
raise ArgumentError.new("Couldn't find any .ko files in extracted modules.cgz") if module_path.nil?
# Extract all the RPMs
#
rpms.each_with_index do |rpm,idx|
Dir.mkdir_and_work_in("x_rpm_#{idx}") do
xsys("rpm2cpio #{rpm} | cpio -id")
end
end
# Scout for modules, copy them into our x_modules directory
#
rpms.length.times do |idx|
Find.find("x_rpm_#{idx}") do |file|
if MODULE_PATH.match(file)
cp(file, module_path)
STDERR.print "Copied #{file.split('/')[-1]}\n"
end
end
end
# Freshen pci.ids file - probably not necessary, just guessing as it seems
# to be the cleanest way to get our nVidia PCI IDs listed.
#
rm("x_initrd/modules/pci.ids")
xsys("wget -q -O x_initrd/modules/pci.ids http://pciids.sourceforge.net/pci.ids";)
# Add some PCI aliases - total guesswork based on existing entries
#
File.open("x_initrd/modules/modules.alias", "a") do |fh|
fh.print <<__
alias pci:v000010DEd000003efsv*sd*bc*sc*i* forcedeth
alias pci:v000010DEd000003f6sv*sd*bc*sc*i* sata_nv
__
end
# Make a new modules.cgz archive
#
Dir.mkdir_and_work_in("x_modules") do
xsys("find . | cut -f2- -d/ | cpio -o -H newc |
gzip >../x_initrd/modules/modules.cgz")
end
# Make a new initrd.img, move it to the right place
#
Dir.mkdir_and_work_in("x_initrd") do
xsys("find . | cut -f2- -d/ | cpio -o -H newc | gzip >../initrd.img")
end
mv("initrd.img", initrd_out)
