ID_FS_TYPE can be much more than just ddf/imsm/linux raid member, so
do the proper checks.
This reverts certain changes from:
cf5891424e
Signed-off-by: Michal Soltys <soltys@ziu.info>
Reworked the flow of the rules file a bit, removed redundant tests, also
should be easier to follow. It's much shorter now as well, a bit more
similar to 90lvm script - both revolve around same concepts after all.
There's no reason to treat conf-assembled arrays differently from
incremental ones. Once we hit timeout in init's udev loop, we can use
common script (mdraid_start.sh) to try force inactive arrays
into degraded mode.
md-finished.sh was kind-of out of place - it didn't really wait for any
particular device(s) to show up, just watched if onetime mdadm scripts
are still in place. Furthermore, after moving mdraid_start to --timeout
initqueue, it didn't really have too much to watch at all, besides
mdadm_auto (and that served no purpose, as we do wait for concrete
devices).
Either way, with stock 64-md fixes, current version of 65-md*.rules does
the following:
- limits assembly to certain uuids, if specified
- watch for no ddf/imsm
- if mdadm.conf => setup onetime -As script, without forced --run option
- if !mdadm.conf => incrementally assemble
- for both cases, setup timeout script, run-forcing arrays as a last resort
Signed-off-by: Michal Soltys <soltys@ziu.info>
1) mdadm -As --auto=yes --run 2>&1 | vinfo (removed)
Currently such auto assembly will not complete or force-run partially
assembled arrays. It might assemble "concurrent" separate array and
force-run it, if possible (though the chances of suddenly showing
missing components in this scenario - a script run after udev timeout -
are pretty thin). See [1] for details. Also see #3 below.
2) mdadm -Is --run 2>&1 (removed)
This will only force-run native arrays - arrays in containers will not
be affected. See [1] for details. Also see #3 below.
3) mdadm -R run loop (implicitly handles #1 & #2)
This loop does everywthing that #1 & #2 are expected to do. Thus, the
above invocations are simply redundant and this is the most safe and
flexible option.
Also, it shouldn't be necessary to go under md/ directory, as those are
just symlinks to /dev/md[0-9]*.
Certain checks were changed to strict ones (array state, degraded state)
instead of relying on env tricks.
'cat' was added explicitly to installed programs (it has been used
implicitly in shutdown script either way)
4) mdmon bug
See [1] for details as well. In short - force-run arrays in containers
will not have mdmon started, so we do that manually.
5) stop/run queue magic
Also removed. mdadm -R will only cause change events to the array
itself, and they should not be an issue.
[1] http://article.gmane.org/gmane.linux.raid/35133
Signed-off-by: Michal Soltys <soltys@ziu.info>
Stop both arrays (first pass) and containers (second pass).
Loop only over /dev/md[0-9]*
Call cleanup script only once, make sure it's after crypt cleanup.
Signed-off-by: Michal Soltys <soltys@ziu.info>
Remove whole "start a container logic".
Containers once assembled, always remain in 'inactive' state.
Any attempt to run a container with mdadm -IR is a no-op, and any
attempt with just mdadm -R ends with an error.
Signed-off-by: Michal Soltys <soltys@ziu.info>
Currently shipped mdadm rules incrementally assemble all imsm and native
raids, and do so unconditionally. This causes few issues:
- fine-grained controls in 65-md* are shadowed - for example,
mdadm.conf's presence tests or uuid checks
- 90dmraid might also conflict with 90mdraid, if user prefers the former
to handle containers
- possibly other subtle issues
This patch adjusts the behaviour.
Signed-off-by: Michal Soltys <soltys@ziu.info>
Given that we boot into a modern Linux distribution with the "/run" toplevel
directory, we can easily mount move the whole /run directory to the real
root in the end and have the complete initramfs later on in
/run/initramfs. All log files and /run states are still accessible and
to save space /run/initramfs can be removed later on.
Because the kernel does not mount a tmpfs on /run prior to unpacking the
initramfs cpio image, we have to copy ourselves very early to a tmpfs
and mount it on /run.
Due to lazy umount the old initramfs binaries should
be removed in the end by switch_root.
This feature can be turned on with "--prefix".
We want all "/var/run" information to live in /dev/.run, until the real
root is mounted.
Therefore we mount a tmpfs on /dev/.run, which can/will be bind/move mounted
on /var/run later on.
When assembling containers + embedded arrays from mdadm.conf,
mdadm needs the /dev/md# node for the container to assemble the
arrays within the container. Stopping the udev exec queue, results in
this node not getting created and mdadm failing to online the
arrays within the container.
Not having stop / start udev exec-queue around "mdadm -As --run" should
be safe as the exact same command is run from rc.sysinit without
any queue locking.
This is a more sane solution, than ignoring subsequent "change" events.
The only danger is that we could loop, if a lvm scan triggers a broken
md partition, which triggers a broken PV and so on.
Better fix the scanning tools, not to emit change events for devices,
if no action was taken.
Copy /etc/mdadm.conf to initramfs (even for non-hostonly) if
mdadmconf="yes" is set in dracut.conf or --mdadmconf is specified on the
dracut command line.
This was done, because there seems _no_ sane way to autoassemble md raid
arrays.
also moved rd_NO_MD to an udev ENV
Since different distros may or may not use vol_id in udev, and blkid
is generally replacing vol_id, abstract them out into a function which
tries to use vol_id first and blkid second, on the assumption that
blkid can take over for vol_id if vol_id is no longer there.
--kernel-only
only install kernel drivers and firmware files
--no-kernel
do not install kernel drivers and firmware files
All kernel module related install commands moved from "install"
to "installkernel".
For "--kernel-only" all installkernel scripts of the specified
modules are used, regardless of any checks, so that all modules
which might be needed by any dracut generic image are in.
The basic idea is to create two images. One image with the kernel
modules and one without. So if the kernel changes, you only have
to replace one image.
Grub and the kernel can handle multiple images, so grub entry can
look like this:
title Fedora (2.6.29.5-191.fc11.i586)
root (hd0,0)
kernel /vmlinuz-2.6.29.5-191.fc11.i586 ro rhgb quiet
initrd /initrd-20090722.img
initrd /initrd-kernel-2.6.29.5-191.fc11.i586.img
initrd /initrd-config.img
initrd-20090722.img
the image provided by the initrd rpm
one old backup version is kept like with the kernel
initrd-kernel-2.6.29.5-191.fc11.i586.img
the image provided by the kernel rpm
initrd-config.img
optional image with local configuration files
LVM
rd_NO_LVM
disable LVM detection
rd_LVM_VG=<volume group name>
only activate the volume groups with the given name
crypto LUKS
rd_NO_LUKS
disable crypto LUKS detection
rd_LUKS_UUID=<luks uuid>
only activate the LUKS partitions with the given UUID
MD
rd_NO_MD
disable MD RAID detection
rd_MD_UUID=<md uuid>
only activate the raid sets with the given UUID
DMRAID
rd_NO_DM
disable DM RAID detection
rd_DM_UUID=<dmraid uuid>
only activate the raid sets with the given UUID
Intel BIOS raid is being shifted from dmraid to mdraid because mdraid offers
more features. So if an imsm metadata capable mdadm is present use mdraid
instead of dmraid for isw_raid_member's
This patch also adds code to mdraid_start.sh so that the raidsets
inside the imsm containers get started once udev is done probing
(doing this earlier leads to potentially degraded use of the sets and
an unwanted resync).
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.
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.
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.
Michal Soltys