TODO: /etc/passwd and /etc/group are not removed yet due to 90mdraid.
dledford said he'll go in and clean this up since he has the hardware
to actually test the mdmon stuff.
init now has the following points to inject scripts:
/cmdline/*.sh
scripts for command line parsing
/pre-udev/*.sh
scripts to run before udev is started
/pre-trigger/*.sh
scripts to run before the main udev trigger is pulled
/initqueue/*.sh
runs in parallel to the udev trigger
Udev events can add scripts here with /sbin/initqueue.
If /sbin/initqueue is called with the "--onetime" option, the script
will be removed after it was run.
If /initqueue/work is created and udev >= 143 then this loop can
process the jobs in parallel to the udevtrigger.
If the udev queue is empty and no root device is found or no root
filesystem was mounted, the user will be dropped to a shell after
a timeout.
Scripts can remove themselves from the initqueue by "rm $job".
/pre-mount/*.sh
scripts to run before the root filesystem is mounted
NFS is an exception, because it has no device node to be created
and mounts in the udev events
/mount/*.sh
scripts to mount the root filesystem
NFS is an exception, because it has no device node to be created
and mounts in the udev events
If the udev queue is empty and no root device is found or no root
filesystem was mounted, the user will be dropped to a shell after
a timeout.
/pre-pivot/*.sh
scripts to run before the real init is executed and the initramfs
disappears
All processes started before should be killed here.
The behaviour of the dmraid module demonstrates how to use the new
mechanism. If it detects a device which is part of a raidmember from a
udev rule, it installs a job to scan for dmraid devices, if the udev
queue is empty. After a scan, it removes itsself from the queue.
it solves the following case:
root=/dev/nfs nfsroot=server:/path
- the server could be reachable on any interface
- any interface can get an IP by dhcp
- only one IP is allowed to mount the root
Udev rules set a /dev/root symlink to the real root and add
a mount script to /mount/. This enables the proper use of pre-mount
scripts and prevents mount being killed by a udev timeout.
Jobs are no longer handled inside the udev events.
/sbin/initqueue is called with the commands to queue.
init will work on these jobs sequentially, so that we prevent jobs
from being killed by udev timeouts.
This serialization also prevents some problems introduced by
the udev event parallelization.
Talked with Debian nbdroot author and he agreed we shouldn't support their obsolete syntax in dracut.
Their root=/dev/nbd[0-9] is no longer needed. Also their syntax was very standardized requiring
a separate boot= parameter.
- Remove root=nfs nfsroot=...
- Remove root=nfs4 nfsroot=...
- Remove root=/dev/nfs4 nfsroot=...
- Legacy nfsroot= without root=/dev/nfs should be unsupported according to the nfsroot.txt.
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=blob;f=Documentation/filesystems/nfsroot.txt
45 root=/dev/nfs
46
47 This is necessary to enable the pseudo-NFS-device. Note that it's not a
48 real device but just a synonym to tell the kernel to use NFS instead of
49 a real device.
Pretty much everyone uses it, and things break of userspace
does not make it because of a case of mistaken identities.
/sbin/start_udev on FC11, I am looking at you.
The Bug causing dhclient to recall BIND has been identified:
dhclient-script runs with #!/bin/sh -e, causing setup_interface
to fail if no hostname is provided by dhcp as it is the last statement
in that function.
This introduces a new cmdline argument bootdev, to support the case
where multiple nics need to be up before the netroot handler is called.
Cases involved might be bonding, iscsi multipathing, bonding, ...
This argument is required to decide which interface is the primary to
use for dhcp root-path, default gw, etc.
When multiple ip= items are present on the cmdline, the ip= parser
now enforces the presence of <dev> further demands that the new argument
bootdev contains the name of the primary interface. Configurtion if of
course still delegated to netroot but in is enhance to ensure that netroot
"waits" for all required interfaces to be up.
Example: root=dhcp ip=eth0:dhcp ip=client-ip:::netmask::eth1:off bootdev=eth0
First, the ip= cmdline parser ensures that all ip items contain a <dev> then
checks the ip items and checks as well that an ip= item for the given bootdev
was found.
When the first netroot starts, probably for eth1, it checks wheter interface
configuration for all interfaces is available. If not it exits. The second
start of netroot (eth0, which was a bit delayed because of dhcp) sees that
all interfaces are present, configures them and continues.
Instead of configuring devices when they are ready, ifup and
dhclient-script write script files which are then sourced by
netroot.
This solves the problem of having multiple interfaces and not knowing
which one to use for dhcp or default route. This way, netroot (which
is serialized anyway) configures the interface before calling the root
handler and deconfigures it if the handler failed.
Example: root=nfs:server:path and ip=dhcp with eth0 and eth1 receiving
a dhcp reply, but eth0 is the correct one to use.
Assuming eth1 is the first to receive the dhcp-reply, netroot starts
and configures eth1. nfsroot is run but fails, so eth1 is deconfigured. If
eth0 has received a dhcp-reply (or not, then we wait) the other locked
netroot process starts and tries with eth0 and succeeds.
Currently network configuration is launched by udev as soon as it
discovers a driver. This isn't such a good idea since we don't know
if network configuration is actually required. Change this by writing
the udev rules on the fly if required.
As with other netroot boot arguments ip= lines should be parsed and
checked. Having this has the advantage that other cmdline scripts
can tell the ip= parser if dhcp or ip=:srv:... is needed, like
parse-root.sh (renamed dhcp-fallback) or parse-nfsroot do.
The nfs test-suite has one check which uses static ip lines, but
the autoconf param is wrong. Fixed this as well.
This introduces detailed cmdline parsing, warning or aborting if the
cmdline does not contain arguments according to the spec.
Makeing sure the parsers don't just call getarg for netroot et al,
allows their reuse inside netroot to analyse dhcp root-path as well.
Hence we can get rid of the current netroot hooks. The hook itself stays
in order to add further modules which should run before netroot handlers
are called.
This has one drawback: nfsroot needs some more logic to handle nfs
specific data inside dhcp root-path.
The parsers have been writting according to current discussions about
cmdline arguments. This lead to the "discovery" that some test-cases
violate the spec. These tests have not been removed, but change to
"must fail".
When die()'ing inside scripts that are udev run, a simple echo isn't
good, udev usually swallows that. Routing messages to /dev/kmsg
ensures they appear on screen and in addition are easily saved through
dmesg.
Printing out an errormessage and quitting is easier if there's a function
to call. If the script calling die has exit trapped, like in init, the
emergency_shell is automatically called.
Multiline conditionals ( && { ... } ) should use if, function
declarations go to the top of the file, add/update comments and remove
some unnecessary clutter.
Saveing ip= options in ifup makes the information available for
later use. This solves the problem how to write static ip configuration
in ifcfg files.
This is mostly about style: Doing stuff after a successful mount
should go into pre-pivot.
In addition this corrects the case where the used netif is not eth0
base/init uses [ -d "$NEWROOT/proc" ] to check if root is mounted
or not. Netroot should do the same. In addition this gets rid of
another file in /tmp, since /tmp/netroot.done isn't needed anymore.
Having two files which are used more or less at the same time isn't
that easy to handle and read. Instead, move the contents of netroot.info
to root.info and only use root.info. This is cleaner and makes debugging
easier.
This is probably not necessary, but paranoia dictates that the actual
netroot handlers should check if all three required arguments (netif,
root, NEWROOT) are there and useable.
rpc.statd is only needed for NFSv4, because it's required for
locking. We don't support locking for NFSv2/3, so start it only
if we are mount from NFSv4.
This patch enhances nfsroot and the nfs test-suite with compatibility
for debian based distros. This is mainly the difference of using
portmap instead of rpcbind and the missing file /etc/netconfig
With this change, we can now use LUKS and LVM over NBD. There are
some decisions to be made regarding where we should get the fstype
and fsoptions from (DHCP root vs rootfstype= etc), but the basic
functionality is there.
This adds support for a command line option netroot=, which is currently
equivalent to root=. This will allow us to break out handling in NBD and
iSCSI to support constructs such as "root=LABEL=/ netroot=dhcp" to make
use of our block device handling with network attached devices.
iSCSI has not been changed in this patch as I don't currently have a way
to test it.
This implements a default path of /tftpboot/%s if no path is provided,
and adds host name substitution for %s, with a fall back to the IP address,
as provided by the kernel's nfsroot handling.
The test suite is updated to test this functionality.
Probe for NFS and NBD capability before trying to load their modules
in case they are built into the kernel. Ugly use of flag files, but
avoids the need for grep to be on the image.
Parse and convert commandline options in the cmdline hook, and
fill in the missing pieces in the netroot hook. This also allows
root=dhcp to work as expected.
/sbin/netroot is a jumping off point to allow various network
root devices to share infrastructure. It will loop over scriptlets
in the netroot handler, looking for a handler to run for this type
of netroot. Handlers can do choose to act based on command line
options to the kernel, or via DHCP options received on this interface.
They should massage root= into a form suitable for their handler.
I noticed when doing the block root= parser that resume
didn't support labeled resume devices. Fix that up and
get rid of a dependency on /bin/echo in the process.
Now init checks that we understood the root= option, and can
act upon it, presenting and error if not. Also, clean up the
generation of udev rules such that we don't require /bin/echo.
First, add a check script to 99base to ensure that it will load its
prerequisites.
Second, disable the udev magic dracut normally uses when generating
test images -- it was causing random failures when creating the test
root filesystem, presumably due to race conditions between the
rootfs creation scripts and udev.
Third, consolidate the rootfs creation scripts into one script.
This gets rid of the need to have potentially security-sensitive files
like /etc/passwd and /etc/group in the initramfs, and makes the initramfs
a bit smaller. We trust the final userspace to do The Right Thing when
assigning user/group ownership to devices.
This also eliminates --skip-missing. Check scripts should now check
to ensure that any files and settings they will copy from the host
system actually exist when called without arguments.
The check scripts are also updated to not try to source dracut-functions
which(1) is a perfectly good way of checking if a command is on the path.
Work-in-progress to mount NFS root from udev, using the DHCP options and
allowing an override from the kernel command line. Supports NFSv4 w/locking
and NFSv2/v3 with nolock.
This is just an RFC patch, you'll note all the XXX markings where I've been
thinking about how to best do things. In addition, we really want to retry
DHCP and NFS mounting in a loop, in case the servers are down or overloaded
like when booting a full diskless cluster.
This changes the network config to run dhclient from within udev, and
allows multiple dhclients to run at once. It also sets the hostname and
adds a kernel command line option 'netdebug' to help debug network issues.
There are several snippits of shell code that are useful to multiple
scripts. Start by moving getarg() into the library.
Signed-off-by: David Dillow <dave@thedillows.org>
If we purport to test booting to an md raid, we may as well test it.
We do not need the plymouth module and the normal crypt module to include
duplicate functionality.
Otherwise it would succeed silently when we KNOW it would fail, which is not cool.
Usually this means you didn't build switch_root in your dracut checkout and you used dracut -l.
(This currently cannot be done in a check script, because of the current design of check.
Andreas intends on discussing a redesign on the list.)
Prevent messages such as the following when booting dracut:
++ mknod /dev/null c 1 3
mknod: `/dev/null': File exists
The plymouth-start.sh script tries to create device nodes it needs, but which
have already been created by the /init script.
This patch makes the mknod calls conditional.
diffstat dracut-plymouth-prevent-mknod-errormsg.patch
plymouth-start.sh | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
With the /check infrastucture in place, dmraid is missing a check file.
Patch is built on top of Harald Hoyers merge tree
(git://git.surfsite.org/pub/git/dracut.git)
Shift some required functionality from switch_root.c to init.
String manipulation and filtering, and testing for the existence of init
is easier in the init script than in C.
The reason this works is that inst will refuse to overwrite a file that
already exists. We jsut make plymouth's code to ask for a password have the
same name as 90crypt's password asking script, and it will override 90crypt's
script while leaving everything else intact.
Followup: Harald Hoyers "Merge" repository boots on ub8.10 with a dracut.conf similar to the following one and the attached patch applied
> cat dracut.conf
modules="ata_piix sd_mod sg"
dracutmodules="dash kernel-modules debug terminfo udev-rules.ub810 base"
The switch_root shell script did not work with bash-4.0-2, because
"exec" gets the real path of the executable which is then
"/sysroot/lib/ld-linux.so.2" instread of "./lib/ld-linux.so.2".
Also the required chroot binary might live in /usr/bin, which can
be mounted later.
Here is the switch_root code from nash, which can be stripped down
further, but which works.
>From f1b1e4f8694104f007e4483c36b0bf40a5760167 Mon Sep 17 00:00:00 2001
From: Harald Hoyer <harald-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
Date: Wed, 4 Mar 2009 13:41:07 +0100
Subject: [PATCH] replace switch_root shell script with binary
The switch_root shell script did not work with bash-4.0-2, because
"exec" gets the real path of the executable which is then
"/sysroot/lib/ld-linux.so.2" instead of "./lib/ld-linux.so.2".
Also the required chroot binary might live in /usr/bin, which can
be mounted later.
Here is the switch_root code from nash, which can be stripped down
further, but which works.
They are not needed to boot with dracut, so away they go.
Also rewrite inst_rules to automatically search likely rule locations when
installing udev rules.
All modules now live in a subdirectory /modules.d. Each module is
self-contained -- it should contain everything that it will install
on the initrd. All modules must include a script named install, which
will be sourced by dracut. This script should do the actual file installation
using inst methods. The moddir variable will point to the full path to the
module dracut is currently working on.
As usual, this patch series applies on top of my previous patch series,
and is available as normalize-modules from http://git.fnordovax.org/dracut.
Harald Hoyer