[systemd-devel] Raid device issue

[Lennart Poettering]

On Wed, 07.12.11 21:39, Ken Bass (kbass at kenbass.com) wrote:

> I just upgraded from F14 to F16 so this is my first attempt at
> systemd and I think I found a bug/issue. (I'm running
> systemd-37-3.fc16.x86_64)
> 
> My /dev/md0 (not a bootable disk thank goodness) is encrypted with luks.
> 
> My crypttab contains:
> encrypt_storage  /dev/md0    /mnt/usbkey/storage.key
> cipher=aes-cbc-essiv:sha256,size=256,hash=sha1
> 
> systemd would fail with:
> [   26.576425] systemd-cryptsetup[951]: crypt_init() failed: Block
> device required
> 
> That tells me that systemd is trying to access /dev/md0 before it
> actually created the device. So,
> after much documentation reading I created a local  cryptsetup.target file:
> 
> [Unit]
> Description=Encrypted Volumes Wait
> Requires=mnt-usbkey.mount dev-md0.device
> After=mnt-usbkey.mount dev-md0.device
> 
> The mnt-usbkey.mount ensures that the USB flash key is loaded early.
> (Before cryptsetup.target). That is not germane to this discussion.
> 
> The problem is the most of the time my changes fail. Occasionally it
> would work, but mostly not. So after examining the logs and some
> debug, I think the problem is a race with systemd or something.
> 
> The problem is dev-md0.device becomes 'plugged' before the /dev/md0
> is actually created/ready. From the log:
> 
> [   26.136607] systemd[1]: dev-md0.device changed dead -> plugged
> [   26.136614] systemd[1]: Job dev-md0.device/start finished, result=done
> [   26.136634] systemd[1]: sys-devices-virtual-block-md0.device
> changed dead -> plugged
> [   26.136691] systemd[1]: About to execute:
> /lib/systemd/systemd-cryptsetup attach encrypt_storage /dev/md0
> /mnt/usbkey/storage.key
> cipher=aes-cbc-essiv:sha256,size=256,hash=sha1
> ...
> [   26.576425] systemd-cryptsetup[951]: crypt_init() failed: Block
> device required
> 
> However, the md0 is not really up. Because the above command only
> fails with /dev/md0 is missing. AND 500 milliseconds later...

Hmm, normally /dev/md0 should have been created by udev before systemd
gets the device plug event.

> If this means dev-md0.device is not accurate as an indicator of
> /dev/md0 being present, then I guess every other reference such as
> in my cryptsetup.target would need to be changed too.

Normally, the device names should be equivalent. It's highly recommended
though to use the /dev/by-xxx names, since they are stable, while
/dev/md0, /dev/md1 and so on might be enumerated in a different order on
different boots.
> 
> This whole issue seems odd and certainly not intuitive. I thought I
> was instructing systemd correcty, telling it to have the depending
> on dev-md0.device. I would have thought any 'uuid' /dev/disk would
> simply be an alias, but that does not appear to be the case.

Normally /dev/md0 should be a device node, while /dev/by-xxx sjould be
symlinks.

Kay, do you know anything about device node creation of md in context
with udev?

Lennart

-- 
Lennart Poettering - Red Hat, Inc.