Re: [PATCH 0/6] Assorted patches relating to mdmon

On Mon, 27 Feb 2023 11:13:07 +1100
NeilBrown [off-list ref] wrote:

mdmon is a root-storage daemon is the sense defined by systemd
documentation, but it does not follow the practice that systemd
recommends.  Specifically it is run from the root filesystem when
possible.  The instance started in the initrd hands-over to a root-fs
based instance, which then hands-over to an initrd-based instance
started by dracut at shutdown.

Part of the reason that we ignore systemd advise is that mdmon needs
access to the filesystem - specifically /dev and /sys - which is not
available in the initrd context after switchroot.  We could possibly
change mdmon to work in the systemd-preferred way by splitting mdmon
into two processes instead of just having 2 threads.  The "monitor"
process could running entirely in the initrd context, the "manager"
process could safely run in the root-fs context, passing newly opened
file descriptors to the monitor over a unix-domain socket.

But we aren't there yet and may never be.

For now, mdmon doesn't work correctly.  There is no mechanism to ensure
a new instance starts after switchroot.  Until recently the initrd
instance of the systemd mdmon unit would be stopped at switchroot time
because udev would temporarily forget about md devices.  This would
allow the "udevadm trigger" process to start a new instance.  udev was
recently fixed:

Commit: 7ec624147a41 ("udevadm: cleanup-db: don't delete information for kept
db entries")

so now the attempt to start mdmon via "udevadm trigger" does nothing as
mdmon already has an active unit.

The net result is that mdmon continues running in the initrd mount
namespace and so cannot access new devices.  Adding a device to a root
md array that depends on mdmon will no longer work.

We want the initrd instance of mdmon to continue running until the
root-fs based instance starts, and that really requires we have two
different systemd units.  This series achieves this in the final patch by
using a different instance name inside or initrd and outside.
"initrd-mdfoo" and "mdfoo".

Other patches in the series are mostly clean-ups and minor improvements
in related code.

NeilBrown

Hi Jes,
The problem descried by Neil is critical for IMSM. I will test the patchset
ASAP.
Additionally, it resolves "KillMode=none" problem so we will be able to finally
drop it.

I will be back with results soon.

Thanks,
Mariusz