Bug 117575 - rhgb breaks root-on-LVM fsck (was: [device-mapper] block device numbers are unstable)
Summary: rhgb breaks root-on-LVM fsck (was: [device-mapper] block device numbers are u...
Keywords:
Status: CLOSED NEXTRELEASE
Alias: None
Product: Fedora
Classification: Fedora
Component: rhgb
Version: rawhide
Hardware: i386
OS: Linux
medium
medium
Target Milestone: ---
Assignee: Daniel Veillard
QA Contact: Mike McLean
URL:
Whiteboard:
: 116573 (view as bug list)
Depends On:
Blocks: FC3Target
TreeView+ depends on / blocked
 
Reported: 2004-03-05 14:59 UTC by Alexandre Oliva
Modified: 2007-11-30 22:10 UTC (History)
5 users (show)

Fixed In Version:
Doc Type: Bug Fix
Doc Text:
Clone Of:
Environment:
Last Closed: 2004-09-09 12:44:17 UTC
Type: ---
Embargoed:


Attachments (Terms of Use)
lvm-on-root rc.sysint fix (1.72 KB, patch)
2004-03-05 18:41 UTC, Stephen Tweedie
no flags Details | Diff
Patched rc.sysinit file (23.21 KB, text/plain)
2004-03-14 20:00 UTC, Th0ma7
no flags Details
lvm-on-root rc.sysinit fix for FC3test1 (2.19 KB, patch)
2004-07-18 02:40 UTC, Alexandre Oliva
no flags Details | Diff

Description Alexandre Oliva 2004-03-05 14:59:35 UTC
Description of problem:
I lost a hard disk that contained some RAID 1 and RAID 5 members that
formed raid1 and raid5 volume groups, as well as some partitions that
didn't have any form of redundancy forming yet another volume group,
that contained only unimportant data.

It turned out that, even though I removed the unimportant filesystem
from /etc/fstab, the system still wouldn't boot, device the
device-mapper block device associated with the root filesystem (on
LVM, in one of the RAID 1 devices) needed to change, so the initial
fsck of the root filesystem failed.  I had to enter the root password,
remount the root filesystem read-write, remove the device-mapper block
devices, run lvm vgscan --mknodes, and then reboot.

Am I the only one who finds the need for these contortions to get a
system to boot up due to the loss of an unrelated hard disk a Very Bad
Thing (TM)?  Is there any hope for device mapper to try to come up
with more stable device numbers?

Version-Release number of selected component (if applicable):
device-mapper-1.00.07-3 kernel-2.6.3-1.118

Comment 1 Stephen Tweedie 2004-03-05 16:35:32 UTC
Known problem.  I've been using the scriptlet:

# If we're running with LVM and with no devfs, we need to populate
/dev/mapper
# now to pick up dynamic major/minor numbers.
dm_minor=`LC_ALL=C fgrep device-mapper /proc/misc | LC_ALL=C awk
'{print $1}'`
if [ "x"$dm_minor != "x" -a ! -e /dev/.devfsd ] ; then
   mkdir -p /dev/mapper
   mount -t tmpfs -o context=system_u:object_r:fixed_disk_device_t
tmpfs /dev/mapper/
   mknod --mode 0600 /dev/mapper/control c 10 $dm_minor
   # We can't lock the lvm config files if root is still r/o
   lvm vgscan --ignorelockingfailure --mknodes
fi

in rc.sysinit as a temporary workaround, but that has nasty
interactions with SELinux.  Currently, the plan to fix this is to use
the initrd's root device node for the initial fsck instead, and only
to unmount initrd once that has completed.

Comment 2 Stephen Tweedie 2004-03-05 18:41:01 UTC
Created attachment 98326 [details]
lvm-on-root rc.sysint fix

Defer initrd unload until after root fs check;
Use the /initrd/dev/* root device node if available.

Comment 3 Stephen Tweedie 2004-03-05 18:42:22 UTC
This patch fixes things for me --- Bill, can you give it a check and
merge it if it looks sane?


Comment 4 Stephen Tweedie 2004-03-08 14:52:42 UTC
*** Bug 116573 has been marked as a duplicate of this bug. ***

Comment 5 Th0ma7 2004-03-14 19:59:08 UTC
This patch ain't working for me...

My system does boot with kernel 2.6.1-1.65smp but takes about 5 mins
to pass the "Setting up local disks"...

My system disk looks like this:
- 1st, Win2k3 server, 9gig
- 2nd, /boot, 256mb
- 3rd, vg00:
   - /, 4096 (lvol1)
   - /var, 2048 (lvol2)
   - /tmp, 1024 (lvol3)

I've tried kernel 2.6.3-2.1.242smp, 246smp, 254smp and I always get this:

insmod: error inserting
'/lib/modules/2.6.3-2.1.246smp/kernel/drivers/acpi/toshiba_acpi.ko':
-1 No such device
Initialising USB Controller (uhci-hcd): OK
Mounting USB filesystem: OK
Checking root filesyste
fsck.ext3: /dev/vg00/lvol0:
The superblock could not be read or does not describe a correct ext2
filesystem.  If the device is valid and it really contains an ext2
filesystem (and not swap or ufs or something else), then the
superblock is corrupt, and you might try running e2fsck with an
alternative superblock:
   e2fsck -b 8193 <device>
Invalid argument while trying to open /dev/vg00/lvol0    FAILED

*** An error occured during the filesystem chkeck.
*** Dropping you to a shell; the system will reboot
*** when you leave the shell.

Give root password for maintenance
(or type Control-D to continue): cat:
/sys//devices/pci0000:00/0000:00:04.3/usb2/2-1/2-1:1.0/host1/1:0:0:0/type:No
such file or directory
Unable to open /etc/fstab for writing: read-only file system
cat: /sys//class/usb/lp0/bNumConfigurations: No such file or directory
/etc/hotplug/usb.agent: line 144: [: too many arguments


---- after typing root passwd:
(repair filesystem) 1# df
filesystem     1k-block   used   available   use%   mounted on
/dev/mapper/vg00-lvol0   4128448   3292616   626120    85% /
none                     4128448   3292616   626120    85% /dev/shm
(repair filesystem) 2#


I've tried the patch and attached my actual /etc/rc.d/rc.sysinit file

Comment 6 Th0ma7 2004-03-14 20:00:22 UTC
Created attachment 98529 [details]
Patched rc.sysinit file

Comment 7 Stephen Tweedie 2004-03-15 11:11:50 UTC
What is specified as the root filesystem in your /etc/fstab and
/etc/grub.conf files?

Comment 8 Th0ma7 2004-03-15 21:26:05 UTC
/boot/grub/grub.conf (partial):
title Fedora Core (2.6.3-2.1.253smp)
        root (hd0,1)
        kernel /vmlinuz-2.6.3-2.1.253smp ro root=/dev/vg00/lvol0 rhgb
        initrd /initrd-2.6.3-2.1.253smp.img
title MS Windows Server 2003
        rootnoverify (hd0,0)
        chainloader +1

Maybie it's new in kernel 2.6 but my LV is seen has
/dev/mapper/vg00-lvolX instead of /dev/vg00/lvolX ... and the error
I'm getting at boot time with any updated kernel is:
fsck.ext3: /dev/vg00/lvol0: blabla..



Comment 9 Bill Nottingham 2004-03-16 23:34:42 UTC
Added in CVS, will be in 7.47-1.

Comment 10 Alexandre Oliva 2004-03-19 03:50:29 UTC
Seems to have fixed the problem for me.  It could use a cosmetic
improvement, though: ATM, it prints `/initrd//dev/int/fc2test: ...' in
the fsck message, instead of the FS label.  I don't particularly care
about not getting the label, but the double slash looks ugly, and
shouldn't be hard to fix.

Comment 11 Th0ma7 2004-03-19 23:57:21 UTC
maybie my problem is a bit different?

Nibody has a clue?

Comment 12 Stephen Tweedie 2004-03-20 00:21:55 UTC
Can you try with the current initscripts package?

Comment 13 Th0ma7 2004-03-20 15:48:27 UTC
still the same problem.... really long to boot with 2.6.1 kernel and
unable to boot with any updated kernel..

Comment 14 Alexandre Oliva 2004-03-20 17:23:17 UTC
What's the fsck error message you get now?  At least the pathname to
the root device should have changed after sct's patch.

Comment 15 Th0ma7 2004-03-20 23:34:45 UTC
in fact, the first comment I've made about my fsck error was with
sct's patch.

fsck.ext3: /dev/vg00/lvol0:
The superblock could not be read or does not describe a correct ext2
filesystem.  If the device is valid and it really contains an ext2
filesystem (and not swap or ufs or something else), then the
superblock is corrupt, and you might try running e2fsck with an
alternative superblock:
   e2fsck -b 8193 <device>
Invalid argument while trying to open /dev/vg00/lvol0    FAILED

*** An error occured during the filesystem chkeck.
*** Dropping you to a shell; the system will reboot
*** when you leave the shell.

Comment 16 Alexandre Oliva 2004-03-21 03:12:35 UTC
Did you install sct's patch or update the package, like he suggested?
 I suppose you did the former, since the latter has the effect of
prepending /initrd/ to the root block device.  It is the root
filesystem it's trying to fsck, right?

Make sure you initrd.img actually has code to enable root on lvm2.

Comment 17 Th0ma7 2004-03-21 17:52:46 UTC
I did install sct's patch... and with the usual up2date I have'nt
found any new initscripts package.

It is the root filesystem on witch fsck is trying to run.

On 2.6 kernel, how can I know if lvm2 is enabled within the initrd image?
Do I simply call "mkinitrd NewImageFile 2.6.3kernel-smp-what-so-ever"
or do I have to add a param like "--with=lvm2"

I've seen at boot time "vg00 activated has lvm2" or something like that.

Comment 18 Th0ma7 2004-03-21 18:09:04 UTC
I've just been able to boot on 2.6.3 kernel.. finally.. The problem is
that, when installing the newer kernel (2.6.1 to 2.6.3-xxx) it did'nt
change the grub.conf file properly.  It seems that the device has to
be "/dev/vg00-lvol0" instead of "/dev/vg00/lvol0". So I just edited my
line :
kernel /vmlinuz-2.6.3-2.1.253.2.1smp ro root=/dev/vg00/lvol0 rhgb
for
kernel /vmlinuz-2.6.3-2.1.253.2.1smp ro root=/dev/vg00-lvol0

and it worked (I removed the rhgb just to actually see all the output).

It still take about 3 to 5 minutes to pass the "Setting up logical
Volume Management" section... I get the error message "/dev/cdrom1:
open failed: read-only file system"

How can I remove /dev/cdrom1 from the lvm check at boot time?

Comment 19 Alexandre Oliva 2004-03-22 00:08:09 UTC
Create /etc/lvm/lvm.conf with an entry such as:

devices {
    filter = [ "a|/dev/md.*|", "r|.*|" ]
}

The above will only look for physical volumes in raid block devices;
adjust to suit your needs.  You may want to have the `r' entry first,
and a catch-all for `a'.

After you're happy with the performance of lvm vgscan on the booted
system, re-run mkinitrd and it will copy lvm.conf into the initrd img.

Comment 20 Th0ma7 2004-03-22 02:31:48 UTC
The lvm.conf proposed work really nicely.  It now boots-up in a few
seconds like it used to with other redhat builds.   And with the
correction to the grub configurations it fixed the 2.6.3 upgrade from
2.6.1.

Everything is now fine.. thnx a lot!

Comment 21 Stephen Tweedie 2004-03-22 15:44:51 UTC
"/dev/vg00-lvol0" looks wrong: I'd expect it to be
"/dev/mapper/dev/vg00-lvol0".  Are you sure it's the former?

Comment 22 Stephen Tweedie 2004-03-22 15:45:31 UTC
Also, are you running in SELinux enforcing mode or not?

Comment 23 Th0ma7 2004-03-22 23:24:22 UTC
I've jsut tried /dev/mapper/vg00-lvol0 and it worked properly with
latest 2.6.3 kernel instead of /dev/vg00-lvol0 (wich was working too:).

When I installed the system, I created the vg00 and it's LV with the
former installer of Core2-test1.  It created a grub.conf file with
/dev/vg00/lvol0 has / filesystem.  With 2.6.3 it seems I needed to
call /dev/mapper/vg00-lvol0 instead but the kernel upgrade did not
change the grub file automatically.. instead it simply cloned the last
entry (from 2.6.1 kernel).

And about SELinux enforcing mode, I don't know... I would need to read
more about it.

note: I have not tried /dev/mapper/dev/vg00-lvol0 ... It did'nt match
with the usual df output.  But if you insist, I can try it.

Comment 24 Stephen Tweedie 2004-03-23 16:03:28 UTC
"/dev/mapper/dev/vg00-lvol0" was a typo, it should indeed be
"/dev/mapper/vg00-lvol0" as you tried.  If it's working for you now, I
can only imagine that it's an initrd issue, as nothing relevant has
changed in the kernel itself.  But as long as it's working, I'll close
it for now.  Please reopen if you find it is not properly fixed.

Comment 25 Alexandre Oliva 2004-03-23 16:37:22 UTC
Err...  I had reopened this because of hte cosmetic issue pointed out
in https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=117575#c10

Should I file a separate bugzilla for it?

Comment 26 Bill Nottingham 2004-03-23 17:23:46 UTC
Yes, please do.

Comment 27 Alexandre Oliva 2004-07-15 23:20:22 UTC
FC3test1 fails to boot again if the root device-mapper device changes.
 /initrd/dev does not exist any more, so once again we're trying to
fsck the outdated device-mapper device in the read-only root filesystem.

Comment 28 Michael Breuer 2004-07-16 14:40:50 UTC
FYI - I believe I'm having the same issue.  I have no issue booting
2.6.3, but cannot boot with any kernel > 2.6.3 - same error as comment
#15, above.

When I enter lvm commands by hand after being dropped into the shell,
I get the locking type 1 errors. When I try --ignorelocking failures,
the lvm commands succeed, but all logical partitions are reported as
invalid and will not mount.  I've tried both "stable" and "latest" dm
and lvm packages.  All of the above works with 2.6.3, but not 2.6.4,
2.6.5 or 2.6.7 (didn't try 2.6.6).  All kernels built from the 2.6.3
.config (make oldconfig).  Also tried with and without selinux support
in the kernel.  I also hand-checked the initrd files created by
mkinitrd vs. the working 2.6.3 - only difference are the kernel modules.

Comment 29 Alexandre Oliva 2004-07-18 02:40:47 UTC
Created attachment 102011 [details]
lvm-on-root rc.sysinit fix for FC3test1

rhgb really abuses /initrd.  It mounts a ramfs atop of the original /initrd,
hiding the device nodes we actually needed.  Which means we end up not
umounting the original /initrd at all (or at least not as early as we could).

This patch, that depends on the directory /dev/initrd-dev to exist (I just made
it up, and I'm open to suggestions), binds /initrd/dev to this new dir, umounts
/initrd early before rhgb starts and mounts something over it, and then umounts
/dev/initrd-dev after fscking the root filesystem.  How does this look?

Comment 30 Alexandre Oliva 2004-07-18 20:03:15 UTC
Ugh, I made a mistake while editing the patch file I posted.  The
`elif' line added in the first hunk is missing a `; then' at the end.
 Sorry about that.

Comment 31 Alexandre Oliva 2004-08-11 21:58:09 UTC
I'm more and more convinced that we shouldn't have to change
rc.sysinit because of rhgb's insistence in abusing /initrd.  I'm
changing this bug to rhgb.

Comment 32 Daniel Veillard 2004-08-11 22:36:36 UTC
Do I understand correctly that you would want rhgb to create
its ramfs in a different location than /initrd ? if yes do
you have suggestion of where it should mount it, any name can
potentially be used for something else. The point I guess was to
avoid stepping on the users/administrator toes because /initrd
is somewhat already reaserved...

Daniel

Comment 33 Alexandre Oliva 2004-08-12 06:57:44 UTC
The problem is precisely that /initrd is reserved.  It is still in use
at the time rhgb starts (if it starts early).

I have suggested /tmp/rhgb, /var/rhgb, or /var/tmp/rhgb.  If it
doesn't exist at the time we attempt to early-start rhgb (before the
root fsck), we wait until root is remounted rw and then we can create
whatever directory we need and proceed.  It would be just a small
delay for rhgb in the unlikely case of the directory we need not being
there.

The only problem is if /tmp or /var are mounted as separate
filesystems: then umounting the rhgb-mounted filesystem may be a bit
tricky.

Another option, if rhgb umounts /initrd twice as it appears to me that
it does (because when it starts it mounts something else atop of
/initrd, but after it completes, I can't see the original initrd
mounted any more), it might as well use /initrd/tmp/rhgb.

I agree it's not a simple problem, but we can't have two different
programs competing to decide what's mounted in /initrd, and clearly
what rhgb is mounting on /initrd is not an initrd image, so I think
it's the one that must give.

Comment 34 Alexandre Oliva 2004-08-16 02:58:22 UTC
Ugh.  This problem is present in FC2 as well.  As soon as I upgraded a
remote box to kernel-2.6.8-1.520, it wouldn't reboot, because the
device number associated with the logical volume holding the root
device, for some reason, changed.

Comment 35 Daniel Veillard 2004-09-09 12:44:17 UTC
I have moved rhgb out of /initrd into /etc/rhgb , this should 
fix this problem. This is commited in CVS, there is test (S)RPMS
available under http://people.redhat.com/veillard/testing/ for
FC2 and Rawhide, version 0.12.4 .

  I hope this resolve the issue, it really should.

Daniel

Comment 36 Alexandre Oliva 2004-09-10 06:29:30 UTC
Thanks for the change.  /etc/rhgb looks like an excellent solution! 
FWIW, udev in initrd (and possibly even in sysinit) already offers a
solution for the problem of fscking the root filesystem, since then
/dev is writable, but for people who choose to disable udev, it would
still be a problem.


Note You need to log in before you can comment on or make changes to this bug.