Arch Linux

[ I've sent this to arch-general@ as well, but it still hasn't shown up, so I'm posting it here as well ]

Hello

For some time I've been using different hooks for raid assembly along with
few assorted changes. I planned to post it on bugs.archlinux.org, but
I'm not sure if it's a proper place for things like this one.

Anyway, for your consideration.

Advantages of this approach:

- only required modules are loaded
- single hook for both partitionable and non-partitionable raids
- no more tedious listing of components during boot, as
everything is assembled as per mdadm.conf
- if user insists on creating initramfs w/o mdadm.conf, it will
be created on the fly during the booting. User can check it out and adjust
the root parameter, should it be inappropriate (if booting fails)

Below, there're udev rules compatible with default (and recommended in man) mdadm behaviour. Current udev rules ignore partitionable raids completely, and get in a way of default mdadm behaviour (potentially in both raid cases).

# md block devices

## regular non-partitionable raid
KERNEL=="md[0-9]*", NAME="md/%n", SYMLINK="%k"

## partitionable raid, volume or partition
KERNEL=="md_d[0-9]*", PROGRAM="/bin/sh -c 'K=%k ; echo $${K#md_}'", NAME="md/%c", SYMLINK="%k"

Alternatively, when rules create nodes in /dev instead of /dev/md (in this case, mdadm itself doesn't create any extra links).

# md block devices (legacy)

## regular non-partitionable raid
KERNEL=="md[0-9]*", NAME="%k", SYMLINK="md/%n"

## partitionable raid, volume or partition
KERNEL=="md_d[0-9]*", PROGRAM="/bin/sh -c 'K=%k ; echo $${K#md_}'", NAME="%k", SYMLINK="md/%c"

Also, one could write a bit more complex rule, that would use helper to check mdadm.conf, and basing on the nodes there, choose the correct approach. OTOH, providing two files, i.e. 60-md-new.rules and 60-md-legacy.rules is just simpler though. Let the user choose his approach, and just stick to it, IMO. Btw - why are udev rules kept in a big, single file ? nn-something.rules scheme makes it easier for users to introduce their own changes, and impose proper order in which they are parsed.

Above rules work fine both during initramfs and rc.sysinit. They won't collide with custom naming if such is used in mdadm.conf, usually with 1.x superblocks - unless the names are deliberately chosen to cause conflicts.

Finally - initialization in rc.sysinit can be greately simplified:

# If necessary, find md devices and manually assemble RAID arrays
if grep -q ^ARRAY /etc/mdadm.conf 2>/dev/null ; then
status "Activating RAID arrays" /sbin/mdadm -As
fi

It's all that is required.
The previous comment 'udev won't create these md nodes, so we do it ourselves' wasn't really relevant, as it was mdadm's default job, unless explicitely blocked (which Arch didn't do).

#RAID2 hook for mkinitcpio

r2_am ()
{
modprobe -q "${1}" >/dev/null 2>&1
eval "${1}=1"
}

run_hook ()
{
## 1) discover which modules are actually needed
fil=/etc/mdadm.conf
if [ ! -f "${fil}" ] ; then
#If user didn't supply proper mdadm.conf, create one.
#Partitions in this case will be assembled as partitionable,
#user SHOULD provide proper mdadm.conf for their system and
#rebuild initramfs afterwards.
echo "CREATE mode=0660 owner=0 group=6 auto=part16 symlinks=yes" >"${fil}"
mdadm -Es >>"${fil}"
fi
while read dat ; do
#get raid type per device
lev="${dat##*level=}"
lev="${lev%% metadata*}"
#and load required module
case "${lev}" in
raid0 ) [ -z "${raid0}" ] && r2_am raid0 ;;
raid1 ) [ -z "${raid1}" ] && r2_am raid1 ;;
raid4 ) [ -z "${raid456}" ] && r2_am raid456 ;;
raid5 ) [ -z "${raid456}" ] && r2_am raid456 ;;
raid6 ) [ -z "${raid456}" ] && r2_am raid456 ;;
linear ) [ -z "${linear}" ] && r2_am linear ;;
multipath ) [ -z "${multipath}" ] && r2_am multipath ;;
faulty ) [ -z "${faulty}" ] && r2_am faulty ;;
esac
done <"${fil}"
## 2) modules ready, now assembly
# things to watch out for:
# raid-partitions require access to device, to make linux actually
# notice exisitng partitions. Without that, there will be no uevents
# right after (dis)assembly, and lvm using sysfs will not check them
# for lvm volumes (unless sysfs_scan = 0). The easiest thing to do
# that, is just to run mdadm -As twice.
# Neil knows about it, and according to him, current mdadm (2.6.4)
# doesn't do mentioned access in one case. It should get fixed in
# one of the future versions. If udev rules access device later
# while processing the rules, i.e. through vol_id helper,
# it would "self" trigger partitions uevents.
# Either way, it's safer to just do it explicitely.

/sbin/mdadm -As >/dev/null 2>&1
/sbin/mdadm -As >/dev/null 2>&1
}

#RAID2 install for mkinitcpio

install ()
{
MODULES=" $(checked_modules "drivers/md/*" | grep -v "dm-") "
BINARIES="mdadm"
FILES=""
SCRIPT="raid2"
add_dir "/dev/md"
add_file "/etc/mdadm.conf"
}

help ()
{
cat<<HELPEOF
This hook is responsible for raid initialisation. It supports both
partitionable and non-partitionable arrays.

Proper mdadm.conf SHOULD be included. If it's not, all raid devices
will be detected during run and assembled as partitionable devices.
User should rebuild initramfs with proper mdadm.conf.
HELPEOF
}


Regarding LVM - in its hook, there's no need for creating mapper/control.

--- lvm2_hook 2008-01-05 14:21:14.000000000 +0100
+++ lvm2 2008-01-05 14:21:30.000000000 +0100
@@ -3,13 +3,11 @@
{
/sbin/modprobe -q dm-mod >/dev/null 2>&1
if [ -e "/sys/class/misc/device-mapper" ]; then
- read dev_t < /sys/class/misc/device-mapper/dev
- /bin/mknod "/dev/mapper/control" c $(/bin/replace "${dev_t}" ':')

[ "${quiet}" = "y" ] && LVMQUIET=">/dev/null"

msg "Scanning logical volumes..."
- eval /bin/lvm vgscan --ignorelockingfailure $LVMQUIET
+ eval /bin/lvm vgscan --mknodes --ignorelockingfailure $LVMQUIET
msg "Activating logical volumes..."
eval /bin/lvm vgchange --ignorelockingfailure -ay $LVMQUIET
fi