add fake root support

Hi,

Attached is a patch for "fake" root functionality,
I'm (mostly) duplicating issue's description from my previous e-mail (for commit message if this is to be pushed):

===============
fakeroot use case -
* initramfs includes ext3 filesystem wrapped into squashfs (like ovirt-node-image has)
* that filesystem is used as a live root (initramfs is pulled from TFTP)
* kernel cmdline has CLIENTSTATE variable set to address of NFS mount. Then rc.sysinit will mount $CLIENTSTATE/$HOSTNAME (as a last resort case) and will use it as a STATE_MOUNT.

Thus, we have read-only root in RAM and all state files mounted from NFS. So server can operate totally diskless (opposed to ovirt-node which has to be set-up on a local disk or iSCSI target, both of which seem to be an overkill if one has 16+ servers). An adding new servers to a pool requires only correct DHCP server settings, nothing more (probably some files on a nfs server too, but that's another story).

What we need from dracut is to set up networking and pass ifcfg's to a system via /dev/.initramfs (for DHCP). But then we need to some-how "simulate" networked root because otherwise there is no chance to bring network up.

To achieve this we could pass "netroot=fake root=live:blablabla CLIENTSTATE=xxx.yyy.zzz.www:/mnt" via kernel's cmdline.
==============

Best,
Vladislav

new file mode 100755
index 0000000..1829b66
@@ -0,0 +1,8 @@
+#!/bin/bash
+# -*- mode: shell-script; indent-tabs-mode: nil; sh-basic-offset: 4; -*-
+# ex: ts=8 sw=4 sts=4 et filetype=sh
+
+# We depend on network modules being loaded
+[ "$1" = "-d" ] && echo network
+
+exit 0

new file mode 100755
index 0000000..79abdf7
@@ -0,0 +1,5 @@
+#!/bin/bash
+# -*- mode: shell-script; indent-tabs-mode: nil; sh-basic-offset: 4; -*-
+# ex: ts=8 sw=4 sts=4 et filetype=sh
+inst_hook cmdline 90 "$moddir/parse-fakeroot.sh"
+inst "$moddir/fakeroot" "/sbin/fakeroot"

new file mode 100755
index 0000000..9d22dd2
@@ -0,0 +1,23 @@
+#!/bin/sh
+# -*- mode: shell-script; indent-tabs-mode: nil; sh-basic-offset: 4; -*-
+# ex: ts=8 sw=4 sts=4 et filetype=sh
+#
+# Format:
+#       netroot=fake root=*
+#
+# both netroot=fake and root= should be set
+#
+# Note: this differs from other netroot handlers which usually replace netroot's value with
+# what root is set too.
+#
+
+# This script is sourced, so root should be set. But let's be paranoid
+[ -z "$root" ] && root=$(getarg root=)
+[ -z "$netroot" ] && netroot=$(getarg netroot=)
+
+if [ "${netroot%%:*}" = "fake" ] ; then
+    if [ -z "$root" ] ; then
+        echo "Warning: both root and netroot should be set for a 'fake' netroot"
+    fi
+fi
+

Sorry for being annoying, but... Any comments?

Best,
Vladislav

I like it, but I don't like the parameters...

Maybe we can express it like this:

netroot=fake:xxx.yyy.zzz.www:/mnt root=live:blablabla

and find another name for "fake".. maybe "container"?

netroot=container:nfs:xxx.yyy.zzz.www:/mnt root=live:blablabla

Which reminds me, that we should probably have the possibility to concatenate several stages.

root=nfs:xxx.yyy.zzz.www:/mnt|loop:test.img|live:blablabla

comments?

This can be used not only with nfs I think... nfs is just one use case which I needed for my project.

And information after fake: should be some-how handled in dracut, but I currently do not see any use-cases for that. Do you see them? iSCSI? Is it really needed so early if it is not for root?

And, I'd not mount not-root nfs from within dracut, we need to know mount point on a real system for that. Additionally, I've already had much fun with nfs4 and root user access to files. This requires rpcidmapd, and it has too many dependencies (and properly configured resolver).

I'd keep it generic, thus making it possible to start network-enabled node solely from PXE. It is much easier to exploit generic solutions (IMHO) for own needs.

Hmmm....
The main idea was to have networking configured without real networked root. Current dracut policy is to skip networking if no networked root is configured. That's why I called it 'fake' - I'm cheating with this policy. I agree that 'fake' is not a best name.
Container... then there should be some contents, but I do not see it...

netroot=force?
networking=force?
Later would require additional efforts, but sounds quite nice for me.

This is quite interesting. But this is from another story.

2010/11/10 Harald Hoyer <[email protected]>:

I like this idea, it would be really useful.

But wouldn't this require some kind of union mount? I ask this because I have read Vladislav's patch, but I've been able to figure out how it sets/links the writable overlay. I'm missing something?

And if that is the case, I suppose that dracut will have to support different kind of union mount (unionfs, aufs, ...) till Valerie's Union Mount gets pushed into the kernel. Or could something be archived only using COW?

Regards,

Jon Ander.

10.11.2010 18:09, Jon Ander Hernandez wrote:
...

BTW I have a patch which skips writable overlay in liveimg completely if 'readonlyroot' is set in cmdline.

Will send it later when I port it to a git.