Version 1.7, last updated: July 19, 1997 by Henrik Storner (storner@osiris.ping.dk)
This document explains how you can use the kerneld function in the Linux kernels. It describes
The latest released version of this document can be found at http://eolicom.olicom.dk/~storner/kerneld-mini-HOWTO.html. Between releases of the mini-HOWTO you can find updates on my unstructured list of changes at http://eolicom.olicom.dk/~storner/kern.html
If you see things in this document that are wrong, please send me a note about it. The following people have contributed to this mini-HOWTO at some point:
I much appreciate the encouragement and suggestions sent to me by readers of the mini-HOWTO.
kerneld is a feature introduced during the 1.3 development kernels by Bjorn Ekwall. It is included with all of the 2.0 and 2.1 kernels. It allows kernel "modules" - i.e. device drivers, network drivers, filesystems - to be loaded automatically when they are needed, rather than having to do it manually with modprobe or insmod.
And for the more amusing aspects, although these are not (yet ?) integrated with the standard kernel:
kerneld consists of two separate entities:
Both pieces must be working for the kerneld support to function - it is not enough that only one or the other has been setup.
There are some good reasons for using kerneld. The ones I will mention are mine - others may want to use it for other reasons.
Of course, there are also reasons why you may not want to use it - you may prefer to have just one kernel image file with all of your drivers built in. In that case, you are reading the wrong document.
The support in the Linux kernel was introduced with Linux 1.3.57. If you have an earlier kernel version, you will need to upgrade if you want the kerneld support. All of the major Linux ftp sites carry the kernel sources - I recommend that you upgrade to the latest stable kernel release, 2.0, now at patch-level 29:
ftp://sunsite.unc.edu/pub/Linux/kernel/v2.0/linux-2.0.29.tar.gz ftp://tsx-11.mit.edu/pub/linux/sources/system/v2.0/linux-2.0.29.tar.gz ftp://ftp.funet.fi/pub/Linux/PEOPLE/Linus/v2.0/linux-2.0.29.tar.gz
The user-level daemon is included with the modules-1.2.8 package, and with the newer modules-2.0 package. These are normally available from the same place as the kernel sources, but the official locations include:
ftp://sunsite.unc.edu/pub/Linux/kernel/v2.0/modules-2.0.0.tar.gz ftp://tsx-11.mit.edu/pub/linux/sources/sbin/modules-2.0.0.tar.gz ftp://ftp.funet.fi/pub/Linux/tools/modules-2.0.0.tar.gz
NOTE: If you want to try module-loading with the latest 2.1 development kernels, you should use the latest modutils- (NOT modules-) package. But see below about the problems with modules and 2.1 kernels.
First get the necessary parts: A suitable kernel and the latest modules-utilities. Then you should install the modules-utilities. Pretty simple - just unpack the sources and run make install. This compiles and installs the following programs in /sbin: genksysm, insmod, lsmod, modprobe, depmod, kerneld. I recommend that you add some lines to your startup-scripts to do some necessary setup whenever you boot Linux. Add the following lines to your /etc/rc.d/rc.S file (if you are running Slackware), or to /etc/rc.d/rc.sysinit (if you are running SysVinit, i.e. Debian, RedHat, Caldera):
# Start kerneld - this should happen very early in the
# boot process, certainly BEFORE you run fsck on filesystems
# that might need to have disk drivers autoloaded
if [ -x /sbin/kerneld ]
then
/sbin/kerneld
fi
# Your standard fsck commands go here
# And you mount command to mount the root fs read-write
# Update kernel-module dependencies file
# Your root-fs MUST be mounted read-write by now
if [ -x /sbin/depmod ]
then
/sbin/depmod -a
fi
The first part starts kerneld itself.
The second part calls 'depmod -a' at startup. The depmod program builds a list of all available modules and analyzes their inter-dependencies, so it knows if one module needs to have another loaded before it will itself load.
NOTE: Recent versions of kerneld as an option links with the GNU dbm library, libgdbm. If you enable this when building the module-utilities, kerneld will not start if libgdbm is not available which may well be the case if you have /usr on a separate partition and start kerneld before /usr is mounted. The recommended solution is to move libgdbm from /usr/lib to /lib, or link kerneld statically.
Next, unpack the kernel sources, configure and build a kernel to your liking. If you have never done this before, you should definitely read the README file at the top level of the Linux sources. When you run make config to configure the kernel, you should pay attention to some questions that appear early on:
Enable loadable module support (CONFIG_MODULES) [Y/n/?] Y
You need to select the loadable module support, or there will be no modules for kerneld to load! Just say Yes.
Kernel daemon support (CONFIG_KERNELD) [Y/n/?] Y
This, of course, is also necessary. Then, a lot of the things in the kernel can be built as modules - you will see questions like
Normal floppy disk support (CONFIG_BLK_DEV_FD) [M/n/y/?]
where you can answer with an 'M' for 'Module'. Generally, only the drivers necessary for you to boot up your system - the harddisk driver, the driver for the root filesystem - should be built into the kernel; the rest can be built as modules.
When you have gone through the 'make config', run 'make dep', 'make clean', 'make zImage' or 'make zlilo', 'make modules' and 'make modules_install'.
Phew.
The 'make zImage' puts the new kernel image in the file arch/i386/boot/zImage. You will need to copy it where you keep your boot-image, or install it in LILO afterwards.
For more information about configuring, building and installing your own kernel, check out the Kernel-HOWTO posted regularly to comp.os.linux.answers, and available from sunsite.unc.edu in /pub/Linux/docs/HOWTO .
Now reboot with the new kernel. When the system comes back up, you can run a 'ps -ax', and you should see a line for kerneld:
PID TTY STAT TIME COMMAND
59 ? S 0:01 /sbin/kerneld
One of the nice things with kerneld is that once you have the kernel and the daemon installed, very little setup is needed. For a start, try using one of the drivers that you built as a module - it is more likely than not that it will work without further configuration. I build the floppy driver as a module, so I could put a DOS floppy in the drive and
osiris:~ $ mdir a:
Volume in drive A has no label
Volume Serial Number is 2E2B-1102
Directory for A:/
binuti~1 gz 1942 02-14-1996 11:35a binutils-2.6.0.6-2.6.0.7.diff.gz
libc-5~1 gz 24747 02-14-1996 11:35a libc-5.3.4-5.3.5.diff.gz
2 file(s) 26689 bytes
So the floppy driver works - it gets loaded automatically by kerneld when I try to use the floppy disk.
To see that the floppy module is indeed loaded, you can run /sbin/lsmod which lists all currently loaded modules:
osiris:~ $ /sbin/lsmod Module: #pages: Used by: floppy 11 0 (autoclean)
The "(autoclean)" means that the module will automatically be removed by kerneld when it has not been used for more than one minute. So the 11 pages of memory (= 44kB, one page is 4 kB) will only be used while I access the floppy drive - if I don't use the floppy for more than a minute, they are freed. Quite nice, if you are short of memory for your applications!
Although kerneld comes with builtin knowledge about the most common types of modules, there are situations where kerneld will not know how to handle a request from the kernel. This is the case with things like CD-ROM drivers or network drivers, where there are more than one possible module that can be loaded.
The requests that the kerneld daemon gets from the kernel is for one of the following items:
kerneld determines what module should be loaded by scanning the configuration file /etc/conf.modules There are two kinds of entries in this file: Paths (where the module-files are located), and aliases (what module should be loaded). If you don't have this file already, you could create it by running
/sbin/modprobe -c | grep -v '^path' >/etc/conf.modules
If you want to add yet another "path" directive to the default paths, you must include all the "default" paths as well, since a path directive in /etc/conf.modules will replace all the ones that modprobe knows by default!
Normally you don't want to add any paths by your own, since the built-in set should take care of all "normal" setups (and then some...), I promise!
On the other hand, if you just want to add an alias or an option directive, your new entries in /etc/conf.modules will be _added_ to the ones that modprobe already knows. If you should _redefine_ an alias or an option, your new entries in /etc/conf.modules will override the built-in ones.
If you run '/sbin/modprobe -c', you will get a listing of the modules that kerneld knows about, and what requests they correspond to. For instance, the request that ends up loading the floppy driver is for the block-device that has major number 2:
osiris:~ $ /sbin/modprobe -c | grep floppy alias block-major-2 floppy
Why block-major-2 ? Because the floppy devices /dev/fd* use major device 2 and are block devices:
osiris:~ $ ls -l /dev/fd0 /dev/fd1 brw-rw-rw- 1 root root 2, 0 Mar 3 1995 /dev/fd0 brw-r--r-- 1 root root 2, 1 Mar 3 1995 /dev/fd1
Character devices are dealt with in a similar way. E.g. the ftape floppy tape driver sits on major-device 27:
osiris:~ $ ls -lL /dev/ftape crw-rw---- 1 root disk 27, 0 Jul 18 1994 /dev/ftape
However, kerneld does not by default know about the ftape driver - it is not listed in the output from '/sbin/modprobe -c'.
So to setup kerneld to load the ftape driver, I must add a line to the kerneld configuration file, /etc/conf.modules:
alias char-major-27 ftape
You can also use the device name instead of the 'char-major-xxx' / 'block-major-yyy' setup. This is especially useful for network drivers. E.g. a driver for an ne2000 netcard acting as eth0 would be loaded with
alias eth0 ne
If you need to pass some options to the driver - e.g. to tell the module about what IRQ the netcard is using, you add an 'options' line:
options ne irq=5
This will cause kerneld to load the NE2000 driver with the command
/sbin/modprobe ne irq=5
Of course, the actual options available are specific to the module you are loading.
Binary formats are handled in a similar way. Whenever you try to run a program that the kernel does not know how to load, kerneld gets a request for "binfmt-xxx", where xxx is a number determined from the first few bytes of the executable. So, the kerneld configuration to support loading the binfmt_aout module for ZMAGIC (a.out) executables is
alias binfmt-267 binfmt_aout
since the magic number (see /etc/magic) for ZMAGIC files is 267. (If you check /etc/magic, you will see the number 0413, but /etc/magic uses octal numbers where kerneld uses decimal, and octal 413 = decimal 267). There are actually three slightly different variants of a.out executables (NMAGIC, QMAGIC and ZMAGIC), so for full support of the binfmt_aout module we need
alias binfmt-264 binfmt_aout # pure executable (NMAGIC) alias binfmt-267 binfmt_aout # demand-paged executable (ZMAGIC) alias binfmt-204 binfmt_aout # demand-paged executable (QMAGIC)
a.out, Java and iBCS binary formats are recognized automatically by kerneld, without any configuration.
Line disciplines are requested with "tty-ldisc-x", with 'x' being usually 1 (for SLIP) or 3 (for PPP). Both of these are known by kerneld automatically.
Speaking of ppp, if you want kerneld to load the bsd_comp data compression module for ppp, then you must add the following two lines to your /etc/conf.modules:
alias tty-ldisc-3 bsd_comp alias ppp0 bsd_comp
Some network protocols can be loaded as modules as well. The kernel asks kerneld for a protocol family (e.g. IPX) with a request for "net-pf-X" where X is a number indicating what family is wanted. E.g. net-pf-3 is AX.25, net-pf-4 is IPX and net-pf-5 is AppleTalk. (These numbers are determined by the AF_AX25, AF_IPX etc. definitions in the linux source file include/linux/socket.h). So to autoload the IPX module, you would need an entr like this in /etc/conf.modules:
alias net-pf-4 ipx
See also the section below on common problems for information about how you can avoid some annoying boot-time messages related to undefined protocol families.
kerneld requests for filesystems are simply the name of the filesystem type. A common use of this would be to load the isofs module for CD-ROM filesystems, i.e. filesystems of type "iso9660":
alias iso9660 isofs
Some devices require a bit on configuration beyond the normal aliasing of e.g. a device to a module.
Hardware devices are usually identified through their major device numbers, e.g. ftape is char-major-27. However, if you look through the entries in /dev for char major 10, you will see that this is a bunch of very different devices, including
Obviously, these devices are controlled by several different modules, not a single one. Therefore, the kerneld configuration for these misc. devices use the major number and the minor number:
alias char-major-10-1 psaux # For PS/2 mouse
alias char-major-10-130 wdt # For WDT watchdog
You need a kernel version 1.3.82 or later to use this; earlier versions
do not pass the minor number to kerneld, making it impossible for kerneld
to figure out which of the misc. device modules to load.
Drivers for SCSI devices consist of a driver for the SCSI host adapter (e.g. an Adaptec 1542), and a driver for the type of SCSI device you use, e.g. a hard disk, a CD-ROM or a tape-drive. All of these can be loaded as modules. However, when you want to access e.g. the CD-ROM drive that is connected to the Adaptec card, the kernel and kerneld only knows that it needs to load the sr_mod module in order to support SCSI CD-ROM's - it does not know what SCSI controller the CD-ROM is connected to, and hence does not know what module to load to support the SCSI controller.
To resolve this, you can add an entry for the SCSI driver module to your /etc/conf.modules that tells kerneld which of the many possible SCSI controller modules it should load:
alias scd0 sr_mod # sr_mod for SCSI CD-ROM's ...
alias scsi_hostadapter aha1542 # ... need the Adaptec driver
This only works with kernel version 1.3.82 or later.
This works if you have only one SCSI controller. If you have more than one, things become a little more difficult.
In general, you cannot have kerneld load a driver for a SCSI host adapter, if a driver for another host adapter is already installed - you must either build both drivers into your kernel (not as modules), or load the modules manually.
There is a way that you can have kerneld load multiple SCSI drivers. James Tsiao came up with this idea:
You can easily have kerneld load the second scsi driver by setting up
the dependency in your modules.dep by hand. You just need an entry like:
/lib/modules/2.0.30/scsi/st.o: /lib/modules/2.0.30/scsi/aha1542.o
To have kerneld load the aha1542.o before it loads st.o. My machine
at home is set up almost exactly like the setup above, and it works
fine for all my secondary scsi devices, including tape, cd-rom, and
generic scsi devices. The drawback is that 'depmod -a' can't
autodetect these dependencies, so the user needs to add them by hand,
and not run 'depmod -a' on boot up. But once it is set up, kerneld
will autoload the aha1542.o just fine.
You should be aware, that this technique only works if you have different kinds of SCSI devices on the two controllers - e.g. hard disks on one controller, and cd-rom drives, tapes or generic SCSI devices on another.
Sometimes, just loading the module is not enough to get things working. For instance, if you have your soundcard compiled as a module, it is often convenient to set a certain volume level. Only problem is, the setting vanishes the next time the module is loaded. Here is a neat trick from Ben Galliart (bgallia@luc.edu):
The final solution required installing the setmix-0.1 package
( ftp://sunsite.unc.edu/pub/Linux/apps/sound/mixers/setmix-0.1.tar.gz )
And then adding the following lines to my /etc/conf.modules :
post-install sound /usr/local/bin/setmix -f /etc/volume.conf
What this does is that after the sound module is loaded, kerneld runs the command indicated by the 'post-install sound' entry. So the sound module gets configured with the command '/usr/local/bin/setmix -f /etc/volume.conf'.
This may be useful for other modules as well, e.g. the lp module can be configured with the tunelp program by adding
post-install lp tunelp <options>
For kerneld to recognize these options, you will need a version of kerneld that is 1.3.69f or later.
NOTE: An earlier version of this mini-HOWTO mentioned a "pre-remove" option, that might be used to run a command just before kerneld removed a module. However, this has never worked and its use is therefore discouraged - most likely, this option will disappear in a future kerneld release. The whole issue of module "settings" is undergoing some change at the moment, and may look different on your system by the time you read this.
If you have tried everything, and just cannot figure out what the kernel is asking kerneld to do, there is a way of seeing the requests that kerneld receives, and hence to figure out what should go into /etc/conf.modules: The kdstat utility.
This nifty little program comes with the modules-package, but it is not compiled or installed by default. To build it:
cd /usr/src/modules-2.0.0/kerneld make kdstat
Then, to make kerneld display information about what it is doing, run
kdstat debug
and kerneld will start spewing messages on the console about what it is doing. If you then try and run the command that you want to use, you will see the kerneld requests; these can be put into /etc/conf.modules and aliased to the module needed to get the job done.
To turn off the debugging, run '/sbin/kdstat nodebug' .
I knew you would ask about how to setup the screensaver module ...
The 'kerneld/GOODIES' directory in modules-package has a couple of kernel patches for screensaver- and consolebeep-support in kerneld; these are not yet part of the official kernel. So you will need to install the kernel-patches and rebuild the kernel.
To install a patch, you use the "patch" command:
cd /usr/src/linux patch -s -p1 </usr/src/modules-2.0.0/kerneld/GOODIES/blanker_patch
Then rebuild and install the new kernel.
When the screensaver triggers, kerneld will run the command "/sbin/screenblanker" - this may be a shell script that runs your favourite screensaver.
When the kernel wants to unblank the screen, it sends a SIGQUIT signal to the process running /sbin/screenblanker. Your shell script or screensaver should trap this, and terminate. Remember to restore the screen to the original text mode!
Around kernel version 1.3.80, the networking code was changed to allow
loading protocol families (e.g. IPX, AX.25 and AppleTalk) as modules. This
caused the addition of a new kerneld request: net-pf-X, where X is a number
identifying the protocol (see /usr/src/linux/include/linux/socket.h for
the meaning of the various numbers).
Unfortunately, ifconfig accidentally triggers these messages, so a lot
of people get a couple of messages logged when the system boots and runs
ifconfig to setup the loopback device. The messages are harmless, and you
can disable them by adding the lines
alias net-pf-3 off # Forget AX.25
alias net-pf-4 off # Forget IPX
alias net-pf-5 off # Forget AppleTalk
to /etc/conf.modules. Of course, if you do use IPX as a module, you
should not add a line to disable IPX.
There have been a couple of reports of this. It seems to be an unfortunate interaction between kerneld and the tkPPP script that is used on some systems to setup and monitor the PPP connection - the script apparently runs loops while running ifconfig. This triggers kerneld, to look for the net-pf-X modules (see above), keeping the system load high and possibly pouring lots of "Cannot locate module for net-pf-X" messages into the system log. There is no known workaround, other than not use tkPPP, or change it to use some other way of monitoring the connection.
Add an entry for the SCSI hostadapter to your /etc/conf.modules. See the description of the scsi_hostadapter entry above.
This is a bug in the module-utilities, that show up only with binutils 2.6.0.9 and later, and it is also documented in the releasenote for the binutils. So read that. Or fetch an upgrade to the module-utilities that fix this, e.g. modules-2.0.0.
The settings for a module are stored inside the module itself when it is loaded. So when kerneld auto-unloads a module, any settings you have made are forgotten, and the next time the module loads it reverts to the default settings.
You can tell kerneld to configure a module by running a program after the module has been auto-loaded. See the section above on the 'post-install' entry.
You cannot. None of the dosemu versions - official or development versions - support loading the dosemu modules through kerneld. However, if you are running kernel 2.0.26 or later, you do not need the special dosemu modules any longer - just upgrade dosemu to 0.66.1.
When the kernel sends a request off to to kerneld, it expects to receive an acknowledgement back within one second. If kerneld does not send this acknowledgement, this message is logged. The request is retransmitted, and should get through eventually.
This usually happens on systems with a very high load - since kerneld is a user-mode proces, it is scheduled just like any other proces on the system. At times of high load, it may not get to run in time to send back the acknowledgement before the kernel times out.
If this happens even when the load is light, try restarting kerneld. (Kill the kerneld proces, and start it again with the command /usr/sbin/kerneld. If the problem persists, you should mail a bug report to linux-kernel@vger.rutgers.edu, but please make sure that your version of the kernel and kerneld is up-to-date before posting about the problem.
There has been a number of reports that the mount(8) command does not wait for kerneld to load the filesystem module. lsmod does show that kerneld loads the module, and if you repeat the mount command immediately it will succeed. This appears to be a bug in the module-utilities version 1.3.69f that affects some Debian users - it can be fixed by getting a later version of the module-utilities.
You need to compile the ncpfs utilities with -DHAVE_KERNELD. See the ncpfs Makefile.
You are using an older version of the smbmount utilities. Get the latest version (0.10 or later) from ftp://tsx-11.mit.edu/pub/linux/filesystems/smbfs/
You cannot modularize everything: The kernel must have enough drivers built in for it to be able to mount your root filesystem, and run the necessary programs to start kerneld. So you cannot modularize
[Actually, this is not true. Late 1.3.x and all 2.x kernels support the use of an initial ram-disk that is loaded by LILO or LOADLIN, and it is possible to load modules from this "disk" very early in the boot process. How to do it is described in the Documentation/initrd.txt file that comes with the kernel source-files.]
Newer versions of kerneld need the GNU dbm library, libgdbm.so, to run. Most installations have this file in /usr/lib, but you are probably starting kerneld before the /usr filesystem is mounted. One symptom of this is that kerneld will not start during boot-up (from your rc-scripts), but runs fine if you start it by hand after that system is up. The solution is to either move the kerneld startup to after your /usr is mounted, or move the gdbm library to your root filesystem, e.g. to /lib.
The Slackware installation (possibly others) builds a default /etc/rc.d/rc.modules which does an explicit modprobe on a variety of modules. Exactly which modules get modprobed depends on the original kernel's configuration. You have probably reconfigured your kernel to exclude one or more of the modules that is getting modprobed in rc.modules, thus, the error message(s). Update your rc.modules by commenting out any modules you no longer use, or remove the rc.modules entirely and let kerneld load the modules when they are needed.
You probably reconfigured/rebuilt your kernel and excluded some modules. You've got some old modules that you no longer use hanging around in the /lib/modules directory. The easiest fix is to delete your /lib/modules/x.y.z directory and do a 'make modules_install' from the kernel source directory again. Note that this problem only occurs when reconfiguring your kernel without changing versions. If you see this error when moving to a newer kernel version you've got some other problem.
Linux 2.1 is the current development kernel. As such, it should be expected that things break from time to time. One of the things that has changed significantly is the way modules are handled, and where the kernel and modules are loaded into memory. Also, Richard Henderson is now in charge of the module kernel development.
In brief, if you want to use modules with a 2.1 kernel, you must
I would recommend using at least kernel 2.1.29, if you want to use modules with a 2.1 kernel.
kerneld originally had some support for establishing dial-up network connections on demand; trying to send packets to a network without being connected would cause kerneld to run the /sbin/request_route script to setup a PPP or SLIP connection.
This turned out to be a bad idea. Alan Cox of Linux networking fame wrote on the linux-kernel mailing list, that
The request-route stuff is obsolete, broken and not required [...] Its also removed from 2.1.x trees.Instead of using the request-route script and kerneld, I whole heartedly advise that you install Eric Schenk's diald package, available from http://www.dna.lth.se/~erics/diald.html
This document is Copyright (c) Henrik Storner, 1996, 1997.
Unless otherwise stated, Linux HOWTO documents are copyrighted by their respective authors. Linux HOWTO documents may be reproduced and distributed in whole or in part, in any medium physical or electronic, as long as this copyright notice is retained on all copies. Commercial redistribution is allowed and encouraged; however, the author would like to be notified of any such distributions.
All translations, derivative works, or aggregate works incorporating any Linux HOWTO documents must be covered under this copyright notice. That is, you may not produce a derivative work from a HOWTO and impose additional restrictions on its distribution. Exceptions to these rules may be granted under certain conditions; please contact the Linux HOWTO coordinator at the address given below.
In short, we wish to promote dissemination of this information through as many channels as possible. However, we do wish to retain copyright on the HOWTO documents, and would like to be notified of any plans to redistribute the HOWTOs.
If you have questions, please contact Tim Bynum, the Linux HOWTO coordinator, at linux-howto@sunsite.unc.edu via email.