Root Drive Mirroring and LVM.

[linas at austin.ibm.com]

On Thu, Feb 05, 2004 at 10:07:32AM +1100, Neil Brown wrote:
> On Wednesday February 4, linas at austin.ibm.com wrote:
> > On Wed, Feb 04, 2004 at 05:32:01PM -0500, Mark Hahn wrote:
> > > > So what's wrong with autodetect, again?
> > >
> > > my guess is: in-kernel autodetect is the problem.
> > > out-of-kernel detection can be much smarter,
> > > and can be more easily tested/replaced.
> >
> > Hm, yes, that makes sense.
>
>
> Good :-)
>
> Just to flesh out my thoughts a bit more:
>
>  If the root filesystem is on an MD array, then I see the process of
>  assembling that md array as quite similar to the process of finding
>  the device for the root filesystem.
>
>  We don't expect the kernel, or anyone else, to scan all devices
>  looking for something that looks like a root filesystem, and loading
>  that.  Rather we tell the kernel or boot loader exactly where to find
>  the root filesystem.  And if the root filesystem moves, we get to
>  explicitly tell the boot loader where it is (root=/dev/hdc1 or
>  whatever).

Well, this is actually one of my more bitter complaints.  I'd much
much rather have some symbolic disk label, and have the kernel scan
*all* devices for that.  I can defend this for both low-end and high-end
machines.

On the low end i.e. PC's, PC servers: I've dealt (multiple times) with
disk and/or controller failure.  If you have more than 3 disks, it can
get very confusing as to which cable is which.  Add to that that some
BIOS'es allow you to enable/disable controllers, which cause devices
to be renumbered. E.g. I have a mobo with two 'plain vanilla' ide
connectors and two '100MHZ' 80-wire ide connectors.  Which one is
/dev/hda and which is not depends on the BIOS settings.  Worse:
if I plug in a 3rd party ide controller, then the numbering becomes
insane:

/dev/hda: mobo 33MHz controller
/dev/hdc: mobo 33MHz controller
/dev/hde: pci card controller
/dev/hdg: pci card controller
/dev/hdi: mobo 100MHz controller
/dev/hdk: mobo 100MHz controller

But if I disable the 33MHz mobo connector in bios:

/dev/hda: mobo 100MHz controller
/dev/hdc: mobo 100MHz controller
/dev/hde: pci card controller
/dev/hdg: pci card controller

Since some types of disk failures will lock up the kernel during boot,
you spend a lot of time plugging and unplugging disks and rebooting.
It gets confusing after a while.

To add to the confusion: If you RAID-1 mirror the root partition,
but then mount it as /dev/hdk (and not /dev/md0) during a rescue
operation (because rescue disks often don't have RAID on them),
and then edit /etc/fstab to try to match the new config...
you end up with two root partitions (each of the mirror pair) with
different /etc/fstab's, on different cables/controllers, ...
after fiddling with that, its a nightmare to figure out which
is which, what's mounted how and where.

I also had a similar experience on a machine with 30-odd scsi disks.
We installed a new kernel on /dev/sdp, rebooted ... and what confusion,
since /dev/sdp was now /dev/sdg.  Worse, since the bootloader was
unaware of this difference ... So then we made a change, and rebooted...
... again, much confusion till we worked it out.

As a cure-all, I started fiddling with ext2 disk labels.  However,
an ext2 disk label, when written on a RAID-1 device, will identify
*three* things: /dev/md0 (for example), and both mirror pairs:
/dev/hda1 and /dev/hde1 all have the same label.

rieserfs doesn't have ext2 labels....

I wanted to fool with other partition table schemes (non-DOS partition
tables) until I realized most PC rescue disks wouldn't be able to get
you out of that jam.

So then I thought about using LVM to label my disks (i.e. use LVM
for only one reason, and no other reason: to be able to assign logical
names, and not physical names).  But think about it... its scary,
LVM is not widespread, somewhat buggy, and standard debian/redhat/suse
rescue diskettes won't have it ... etc.


>  Assembling the root device should be handled the same way.  We tell
>  the boot loader/kernel where to expect it, but can over-ride that if
>  we need to:
>    md=0,/dev/hdc1,/dev/hde4
>
>  All other md arrays can, and so should, be assembled by code running
>  out of the root filesystem.

*If* you mounted the "right" root filesystem.  If you have multiple
copies around, and edit fstab on one but not the others, and then
recable and reboot ...

>  This could be some program that
>  assembles anything it finds after scanning all devices, or something
>  a bit more focused, but it should be controllable by the sysadmin.

At least once I managed to mount /usr as /var because of the confusion,
and upon reboot the init.d scripts spewed /var/lock crud into my poor
/usr filesystem ...

I want to know that /dev/mdwhatever is /var *before* I mount it,
not after.

>  It is true that in-kernel auto-detect can be controlled by fiddling
>  with partition types, but the problem is that it runs *before* the
>  root filesystem is mounted and so could conceivably confuse the
>  assembly of the root device (if e.g. you plugged in some other device
>  that also claimed to be part of /dev/md0, and it got scanned before
>  your real root device).

Well, yes, that too, I suppose.  But as you see, explicitly
specifying it at the boot prompt works only if you type in the
right thing ...

--linas


cc.ing ppc64 because although not an architecture issue, it is a
sysadmin issue on enterprise-class machines.

cc.ing hot-plug because this is a cold-plug issue.

Note you get similar crud for multiple ethernet cards ...

** Sent via the linuxppc64-dev mail list. See http://lists.linuxppc.org/