Recompiling kernel on VanGogh cluster, and patching with web100 modification

6 may 2003, Freek Dijkstra

The installation is done in three parts:

1. Installing a new clean 2.4.26 kernel on vangogh0
2. Patching the kernel and installing it on vangogh0
3. Copying the kernel to the other nodes

---

A: Installing a new clean 2.x kernel on vangogh0

(Step 1) GRAB SOURCE

This step compiles a kernel from scratch, using externel sources. For Red Hat or Debian, you may prefer to download the sources using a specific rpm of deb package. See earlier howto's for that information.

(Variant A) Download fresh sources

# wget http://www.kernel.org/pub/linux/kernel/v2.4/linux-2.4.26.tar.bz2∞
# bunzip2 linux-2.4.26.tar.bz2
# tar xf linux-2.4.26.tar

(Variant B) Use Red Hat:

On Red Hat, we need the kernel-source package:
$ rpm -q kernel-source
package kernel-source is not installed
[Insert RedHat 7.3, CD 2]
# mount /dev/cdrom /mnt/cdrom
$ uname -r
2.4.18-3
[let's us the same kernel version]
# rpm -Uvh /mnt/cdrom/RedHat/RPMS/kernel-source-2.4.18-3.i386.rpm
[note: recommended package kernel-headers not found]

(Variant C) Use Debian:

You may want to consider installing a precompiled kernel, instead of compiling it yourself, in particular if you are not going to make any patch.

not written. go to http://packages.debian.org∞ to find sources.

(All Variants) Make sure the linux symbolic link points to the new source:

(Step 2) VERIFY THE PATCH WORKS

If you are going to patch the source, it is wise to first go through the whole procedure once without making the patch, so you know possible problems are not related to the patch. However, before you do that, make sure that the patch you have is indeed for the source you are working with.

Download patch:



Verify if patch works:



No problems should be found. Patch works flawlessly.

Now, revert the patch. Either using patch -r


Or by simply unpacking the sources again (see Step 1).
For example, for a downloaded tar file:

(Step 3) BACKUP THE CURRENT CONFIGURATION

If /boot is empty, it is most likely that it's contents is on a seperate partition which was automatically unmounted after the booting was finished. Use mount, for example "mount /dev/sda1 /boot", to mount the disk. (You can usually find the correct device in /etc/mtab or /etc/fstab)

Make sure the current kernel configuration is copied in /boot/. Here it is:
If it is not, make a copy:

(Step 4) CLEAN CURRENT INSTALL / CONFIG

This is not needed if the kernel was downloaded from scratch, and is only necessary if you alter an existing kernel.

However, it does not hurt, so here it is for sake of completeness:

(Step 5) CONFIGURE THE NEW KERNEL

For ease, you can start with an old configuration file (you did save it at step 2, did you?)

Now you can choose which of the tools you want to use to make changes to the .config file:

(a) On X:

(b) Curses:

(c) Vt100:

In the config it is important to check CPU type, SMP support (for dual processors), and enable loadable kernel support. Be sure to check these, especially if you are starting from scratch

Note: If you are going to build a complete module from scratch, make very sure that you build drivers for all your hardware as is listed in /etc/modules.conf. Otherwise, you will fail at step 9. If you are creating a derivate, based on your current system, there are most probably easy workarounds at step 9, because you have already compiled the modules earlier.

Optionally, you can manually edit the .config file by hand. For example, to set a parameter like CONFIG_NET_E1000=m (for the Intel 1Gbps network card)

(Step 6) MAKE DEPENDENCIES

(not needed for 2.6 kernels)

(Step 7) GIVE NEW KERNEL AN UNIQUE NAME

Edit Makefile

Edit the line starting with EXTRAVERSION, and change it to something unique. For example

(Step 8) MAKE

Note: you can simultaneous do the following two steps. Both will take a long time, so plan your lunch / night sleep / long break here. You can also combine the two commands into one: # make bzImage modules


Alternatively, if you want to keep the stdout:



Alternatively, if you want to keep the output:


Verify if the bzImage was correctly created


Inspect the outputs of both make's, and check if there have been any errors.
If there have been, stop and optionally start over.

(Step 9) INSTALL KERNEL AND MODULES

Install modules in /lib/modules (only if you enabled loadable module support, which you probably did in step 4):


Now you must copy the newly created kernel to the /boot/ directory. First check the current naming convention. It is either bzImage.plain.20031015 or vmlinuz-2.4.18-3plain.20031015 (starting with bzImage or vmlinuz)


If /boot/ is empty, it is probably a seperate partition which is not mounted. mount it first. See note at Step 3.

Install the kernel


Copy the config file (this is technically not necessary, but for organisational purposes it is a must):


Copy the System map file (this is required for some libraries):

(Step 10) CREATING AN INITIAL RAMDISK

This step is optional for some systems. However, it is highly recommended that you try it anyway, since you will detect missing modules in an early stage.

Some systems do NOT need an initial RAMdisk. For example, when using EFI and ELILO current installations do not use a ramdisk (though they might in the future). Best is to check in the bootloader config (/etc/lilo.conf, /etc/grub.conf or /etc/elilo.conf) for "init=" lines, or for initrd- or .img files in the /boot/ directory. If there are, you probably do need a initial ramdisk.

You can typically check whether this is the case if


This step may fail, typically with a message like:

Background: Mkinitrd makes a Initial Ramdisk (initrd) which is just a very small kernel with support for some drivers like SCSI and RAID. It is typically used in either as a first boot to access a real kernel residing on a SCSI or RAID device (it can just mount / and then kick off init), or it is used to create a boot floppy.
mkinitrd looks at /etc/modules.conf to see which modules are necessary for a mimimum. It will look into /lib/modules/ for the appropriate modules, and combine these into the initrd ramdisk.

If you get the error above, you should restart at step 4.
Either you can use xconfig, making sure that all modules are compiled as module this time. For example, you can find aic7xxx in SCSI Support->SCSI Low-level Drivers->Adaptec AIC7xxx support
Otherwise, you can manually edit the .config file, and find the required module, for example CONFIG_SCSI_AIC7XXX, and set the variable to m.
You can take the shortcut, leaving the kernel untouched, and only doing make modules and make modules_install before retrying step 9 again.

If all succeeded, copy the ramdisk:

(Step 11) CONFIGURING THE BOOTLOADER

LILO is a boot loader for i386 machines. Alternatively you can use the newer GRUB, which has a few advantages over LILO. For non-pentium machines, you need another bootloader anyway (for example, on the ppc, you can use yaboot, quik, BootX or OpenFirmware directly). It is very advisable to use the same bootloader as is currently already in use.

(Variant A) For LILO:

Edit /etc/lilo.conf
And add a section, like this:



(the initrd is only needed if you made an initial ramdisk)


Note that you MUST run the lilo command every time you create a new image, even if you don't edit lilo.conf! Otherwise, you will run into serious problems (lilo will store the location of the image file on the hard disk, and uses that at boot time).

(Variant B) For GRUB:

Edit /boot/grub/menu.lst or /boot/grub/grub.conf
And add a section, like this:



(the initrd is only needed if you made an initial ramdisk)
Adjust the default, and optionally the fallback lines as well.

(Check the boot device, typically /dev/sda or /dev/hda. Use df)

(Variant C) For ELILO:

For Elilo, as currently configured on the theo Itanium, the procedure is not ideal, but manageable:

First you do not need an initial ramdisk (though this is likely to change in forthcoming releases, be sure to check).
Second, you need to copy the vmlinuz files out of the /boot partition:

Make sure /boot is mounted

Copy (or move, but copy is safer) all files from /boot/ to /:


Now, you need to unmount /boot to run elilo.


Edit /etc/elilo.conf
(do not edit /boot/efi/debian/elilo.conf; this file will be overwritten later on!)
And add a section, like this:



Adjust the default line as well.

Run elilo to copy the kernel (and other files) from /vmlinuz-2.4.22-web100 to /boot/efi/debian/.


This should mount /boot, copy the files, and unmount /boot again.

(Step 12) REBOOT THE MACHINE

TROUBLESHOOTING

this means that the initrd kernel can't access the root system. You probably need to add more modules to your initrd (Step 10).
Typically, your ramdisk does not contain SCSI support. Check if aic7xxx is listed in /etc/modules.conf

---

B: Installing a patched 2.4.26-web100 kernel on vangogh0

Go through the above steps, only with a patched kernel:

1. Grab source (download variant)

2. Apply the patch

3. Backup the current configuration

already did so

4. Clean current install / config

5. Configure the new kernel

Just use the same configuration:


Change options as recommended in the Web100 kernel INSTALL readme file.

6. Make dependencies

7. Give new kernel an unique name

Edit Makefile

8. Make

Check logfiles: no errors?

9. Install kernel and modules

10. Create an initial ramdisk

11. Configure the bootloader (grub variant)

Add section:

12. Reboot

---

C: Copying the kernel to the other nodes

(Step 1) CONNECT TO DESTINATION NODE

# ssh root@vangogh1

(Step 2) COPY FILES

Should copy:

Note: the command bellow will follow symbolic links. In particular,
the build directory, which can be rather hugh (> 100 MB). As an
alternative, you may want to use tar to copy the files, which does
not traverse symbolic links.



Should copy a lot of files.

(Step 3) CONFIGURE THE BOOTLOADER

Note: The vangogh nodes used LILO. This has been GRUB.

(Variant A) For GRUB:

Edit /boot/grub/grub.conf
And add a section, like this:



Adjust the default, and optionally the fallback lines as well.

(Check the boot device, typically /dev/sda or /dev/hda. Use df)

(Step 4) REBOOT THE MACHINE

TROUBLESHOOTING

The "label / not found" error indicates that no hard disk with the label "/" could be found. Apparently, the hard disk does not have that meta-information in it's boot block. The easiest work-around is to specify the device instead of the label as startup parameter.
Thus, in grub.conf (or something simular in lilo.conf), change the line:
to the line: