In the long run, it's desirable to be able to drop dracut's copy of
multipathd.service and use the upstream one from multipath-tools instead.
This patch makes a step in that direction.
With these changes, the only remaining difference is the support for
rd.multipath=0 and rd_NO_MULTIPATH, which must obviously be ignored in the
upstream unit.
The modifications in this patch are minor and will have no effect in the
initramfs.
The dependency of multipathd on "udev settle" has recently been removed in
c9689b6 ("multipathd: Remove dependency on systemd-udev-settle.service").
But this dependency has never been necessary in the initramfs environment. It
was only required after switching root, because multipathd would potentially
tear down valid multipath maps after switching from initrd to root FS. This can
happen because dm devices "survive" the root FS switch in the udev data
base (they have the "db_persist" flag set), whereas their component devices
(SCSI etc) do not. But this can only happen after initrd-udevadm-cleanup-db.service
has been run, which happens after initrd processing.
The only dependency that's really needed is that on
systemd-udevd-kernel.socket, because multipathd depends on uevents for
devices being delivered via systemd-udevd.
This removes the 'ExecStop=' field from `multipathd.service`.
Sometimes CI runs do encounter a failure related to this
service in initrd, which seems to be stemming from a socket
I/O race between the client and the server on shutdown.
It looks like the client (`multipathd shutdown`) can lose the race,
hit an I/O error, and cause the whole unit to fail (even if the server
managed to shutdown properly already).
Notably, the upstream unit does not have such stop command
as the daemon can already perform a graceful exit through
its signal handler.
As such, this commit partially re-aligns the two units,
trying to sidestep any of the existing races.
Refs:
* https://github.com/coreos/fedora-coreos-tracker/issues/803
* https://github.com/opensvc/multipath-tools/blob/0.8.7/multipathd/multipathd.service
All device-mapper based devices, including device-mapper-multipath,
do reuse the udev db from the initramfs after switching to the root fs.
Therefore device-mapper devices have to be correctly initialized before
the udev daemon is stopped, to have the correct entries in the udev db.
See also https://bugzilla.redhat.com/show_bug.cgi?id=1949076
local-fs-pre.target serves as a separator between the code for
detecting block devices and systemd's fsck/mount logic. This
patch ensures that multipathd is started before local-fs-pre.target
in the initrd. By adding a "Wants=" line for local-fs-pre.target,
it makes sure that this target is started at all.
References: bsc#1006118
Signed-off-by: Martin Wilck <mwilck@suse.de>
===================================================================
SLES11 provided a kernel commandline option 'multipath=off',
so dracut should be parsing the option, too.
References: bsc#1001691
Signed-off-by: Hannes Reinecke <hare@suse.com>
We need to wait until udev has processed all events, otherwise we'll
risk of misdetecting devices. This might cause a temporary interruption
during which multipath removes a device-mapper device, which then
causes a booting failure.
References: bsc#986734
Signed-off-by: Hannes Reinecke <hare@suse.com>
References: bsc#930019
If rootfs is on multipath, but platform does not have an /etc/multipath.conf
file which is not urgently needed, they system will not boot, due to:
multipathd is not started and rootfs and swap are not found:
systemctl status multipathd.service
* multipathd.service - Device-Mapper Multipath Device Controller
Loaded: loaded (/usr/lib/systemd/system/multipathd.service; disabled; vendor preset: enabled)
Active: inactive (dead)
Condition: start condition failed at Thu 2015-05-07 11:49:11 CEST; 7min ago
ConditionPathExists=/etc/multipath.conf was not met
and exit to dracut shell.
The multipathd package might install a service and a socket
file. Using the original service file from the installed
system without the socket file triggers a bug in systemd,
causing systemd to crash.
As we don't actually need to socket file in the initrd we
should be installing our own service file which does not
reference the socket file at all.
References: bnc#871610
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Thomas Renninger <trenn@suse.de>
Martin Wilck