[linux-lvm] Can't work normally after attaching disk volumes originally in a VG on another machine

Post by Gang He
6) attach disk2 to VM2(tb0307-nd2), the vg on VM2 looks abnormal.
tb0307-nd2:~ # pvs
WARNING: Device for PV JJOL4H-kc0j-jyTD-LDwl-71FZ-dHKM-YoFtNV not
found or rejected by a filter.
PV VG Fmt Attr PSize PFree
/dev/vdc vg2 lvm2 a-- 20.00g 20.00g
/dev/vdd vg1 lvm2 a-- 20.00g 20.00g
[unknown] vg1 lvm2 a-m 20.00g 20.00g

This is normal because /dev/vdd contains metadata for vg1 which includes
now missing disk /dev/vdc .... as the PV is no longer the same.

Post by Gang He
tb0307-nd2:~ # vgs
WARNING: Device for PV JJOL4H-kc0j-jyTD-LDwl-71FZ-dHKM-YoFtNV not
found or rejected by a filter.
VG #PV #LV #SN Attr VSize VFree
vg1 2 0 0 wz-pn- 39.99g 39.99g
vg2 1 0 0 wz--n- 20.00g 20.00g

This is normal because you haven't removed /dev/vdc from vg1 on
/dev/vdd, since it was detached while you operated on its vg.

Post by Gang He
7) reboot VM2, the result looks worse (vdc disk belongs to two vg).
tb0307-nd2:/mnt/shared # pvs
PV VG Fmt Attr PSize PFree
/dev/vdc vg1 lvm2 a-- 20.00g 0
/dev/vdc vg2 lvm2 a-- 20.00g 10.00g
/dev/vdd vg1 lvm2 a-- 20.00g 9.99g

When you removed vdd when it was not attached, the VG1 metadata on vdd
was not altered. The metadata resides on both disks, so you had
inconsistent metadata between both disks because you operated on the
shared volume group while one device was missing.

You also did not recreate PV on /dev/vdc so it has the same UUID as when
it was part of VG1, this is why VG1 when VDD is booted will still try to
include /dev/vdc because it was never removed from the volume group on
VDD.

So the state of affairs is:

/dev/vdc contains volume group info for VG2 and includes only /dev/vdc

/dev/vdd contains volume group info for VG1, and includes both /dev/vdc
and /dev/vdd by UUID for its PV, however, it is a bug that it should
include /dev/vdc even though the VG UUID is now different (and the name
as well).

Regardless, from vdd's perspective /dev/vdc is still part of VG1.

Gang He

2018-03-26 06:04:23 UTC

Hi Xen,

This is normal because /dev/vdd contains metadata for vg1 which includes
now missing disk /dev/vdc .... as the PV is no longer the same.

This is normal because you haven't removed /dev/vdc from vg1 on
/dev/vdd, since it was detached while you operated on its vg.

When you removed vdd when it was not attached, the VG1 metadata on vdd
was not altered. The metadata resides on both disks, so you had
inconsistent metadata between both disks because you operated on the
shared volume group while one device was missing.
You also did not recreate PV on /dev/vdc so it has the same UUID as when
it was part of VG1, this is why VG1 when VDD is booted will still try to
include /dev/vdc because it was never removed from the volume group on
VDD.
/dev/vdc contains volume group info for VG2 and includes only /dev/vdc
/dev/vdd contains volume group info for VG1, and includes both /dev/vdc
and /dev/vdd by UUID for its PV, however, it is a bug that it should
include /dev/vdc even though the VG UUID is now different (and the name
as well).

It looks like each PV includes a copy meta data for VG, but if some PV has changed (e.g. removed, or moved to another VG),
the remained PV should have a method to check the integrity when each startup (activated?), to avoid such inconsistent problem automatically.

Thanks
Gang

Post by Xen
Regardless, from vdd's perspective /dev/vdc is still part of VG1.
_______________________________________________
linux-lvm mailing list
https://www.redhat.com/mailman/listinfo/linux-lvm
read the LVM HOW-TO at http://tldp.org/HOWTO/LVM-HOWTO/

Fran Garcia

2018-03-26 10:23:40 UTC

Post by Gang He
It looks like each PV includes a copy meta data for VG, but if some PV has changed (e.g. removed, or moved to another VG),
the remained PV should have a method to check the integrity when each startup (activated?), to avoid such inconsistent problem automatically.

Your workflow is strange. What are you trying to accomplish here?

Your steps in 5 should be:

vgreduce vg01 /dev/vdc /dev/vdc
pvremove /dev/vdc /dev/vdd

That way you ensure there's no leftover metadata in the PVs (specially
if you need to attach those disks to a different system)

Again a usecase to understand your workflow would be beneficial...

Cheers

fran

Gang He

2018-03-27 05:55:57 UTC

Hi Fran,

Post by Gang He
It looks like each PV includes a copy meta data for VG, but if some PV has

changed (e.g. removed, or moved to another VG),

Post by Gang He
the remained PV should have a method to check the integrity when each

startup (activated?), to avoid such inconsistent problem automatically.
Your workflow is strange. What are you trying to accomplish here?

I just reproduced a problem from the customer, since they did virtual disk migration from one virtual machine to another one.
According to your comments, this does not look like a LVM code problem,
the problem can be considered as LVM administer misoperation?

Thanks
Gang

Post by Gang He
vgreduce vg01 /dev/vdc /dev/vdc
pvremove /dev/vdc /dev/vdd
That way you ensure there's no leftover metadata in the PVs (specially
if you need to attach those disks to a different system)
Again a usecase to understand your workflow would be beneficial...
Cheers
fran
_______________________________________________
linux-lvm mailing list
https://www.redhat.com/mailman/listinfo/linux-lvm
read the LVM HOW-TO at http://tldp.org/HOWTO/LVM-HOWTO/

Zdenek Kabelac

2018-03-27 08:32:45 UTC

Post by Gang He
Hi Fran,

Post by Gang He
It looks like each PV includes a copy meta data for VG, but if some PV has

changed (e.g. removed, or moved to another VG),

Post by Gang He
the remained PV should have a method to check the integrity when each

startup (activated?), to avoid such inconsistent problem automatically.
Your workflow is strange. What are you trying to accomplish here?

Ahh, so welcome Eric's replacement :)

Yes - this use scenario was improper usage of lvm2 - and lvm2 has cached the
user before he could ruing his data any further...

Regards

Zdenek

Gang He

2018-03-28 02:09:01 UTC

Hi Zdenek,

Post by Gang He
Hi Fran,

Post by Gang He
It looks like each PV includes a copy meta data for VG, but if some PV has

changed (e.g. removed, or moved to another VG),

Post by Gang He
the remained PV should have a method to check the integrity when each

startup (activated?), to avoid such inconsistent problem automatically.
Your workflow is strange. What are you trying to accomplish here?

I just reproduced a problem from the customer, since they did virtual disk

migration from one virtual machine to another one.

Post by Gang He
According to your comments, this does not look like a LVM code problem,
the problem can be considered as LVM administer misoperation?
Thanks
Gang

Ahh, so welcome Eric's replacement :)

Yes, thank for your inputs.

Post by Gang He
Yes - this use scenario was improper usage of lvm2 - and lvm2 has cached the
user before he could ruing his data any further...
Regards
Zdenek

Xen

2018-03-27 09:12:22 UTC

Post by Gang He
I just reproduced a problem from the customer, since they did virtual
disk migration from one virtual machine to another one.
According to your comments, this does not look like a LVM code problem,
the problem can be considered as LVM administer misoperation?

Counterintuitively, you must remove the PV from the VG before you remove
the (physical) disk from the system.

Yes that is something you can often forget doing, but as it stands
resolving the situation often becomes a lot harder when you do it in
reverse.

Ie. removing the disk first and then removing the PV from the VG is a
lot harder.

Zdenek Kabelac

2018-03-27 10:22:28 UTC

Counterintuitively, you must remove the PV from the VG before you remove the
(physical) disk from the system.
Yes that is something you can often forget doing, but as it stands resolving
the situation often becomes a lot harder when you do it in reverse.
Ie. removing the disk first and then removing the PV from the VG is a lot harder.

IMO 'vgreduce --removemissing' doesn't look to me like a real rocket science.

Of course if the space on a missing PV has been in-use with some LVs - it gets
a bit more complex - but still as long as you don't care about data adding
'--force' option should usually help (although there can be some room for
improvement - but every coding takes its time - and there are not to many
hands working on this...)

Regards

Zdenek

Xen

2018-03-27 10:27:42 UTC

Post by Zdenek Kabelac
IMO 'vgreduce --removemissing' doesn't look to me like a real rocket science.

Yeah I don't wanna get into it.

--force didn't work very well when the missing PV was a cache PV as it
removed the entire cached volume including the origin.

Zdenek Kabelac

2018-03-27 22:17:15 UTC

IMO 'vgreduce --removemissing' doesn't look to me like a real rocket science.

Yeah I don't wanna get into it.
--force didn't work very well when the missing PV was a cache PV as it removed
the entire cached volume including the origin.

Hi

This is why users do open BZ if they would like to see some enhancement.

Normally cache is integral part of a volume - so it's partially missing -
whole volume is considered to be garbage.

But in 'writethrough' mode there could be likely possible better recovery.

Of course this case needs usability without --force.

So please open RFE BZ for this case.

Regards

Zdenek

Xen

2018-03-28 10:08:25 UTC

Post by Zdenek Kabelac
Hi
This is why users do open BZ if they would like to see some
enhancement.
Normally cache is integral part of a volume - so it's partially
missing - whole volume is considered to be garbage.
But in 'writethrough' mode there could be likely possible better recovery.
Of course this case needs usability without --force.
So please open RFE BZ for this case.

It goes into the mess I usually get myself into; if you "dd copy" the
disk containing the origin volume before uncaching it, and then go to
some live session where you only have the new backup copy, but you want
to clean up its LVM,

then you now must fix the VGs in isolation of the cache; I suppose this
is just the wrong order of doing things, but as part of a backup you
don't really want to uncache first, as that requires more work to get it
back to normal after.

So you end up in a situation where the new origin copy has a reference
to the cache disk --- all of this assumes writethrough mode --- and you
need to clear that reference.

However, you cannot, or should not, attach the cache disk again; it
might get effected, and you don't want that, you want it to remain in
its pristine state.

Therefore, you are now left with the task of removing the cache from the
VG, because you cannot actually run vgimportclone while the cache disk
is missing.

The obvious solution is to *also* clone the cache disk and then run
operations on the combined set, but this might not be possible.

Therefore, all that was left was:

vgreduce --remove-missing --force
cd /etc/lvm/archive
cp <latest backup> /etc/lvm/backup/<vg>
cd /etc/lvm/backup
vi <vg>
" remove cache PV, and change origin to regular linear volume, and add
" the visible tag

vgcfgrestore <vg>

# presto, origin is restored as regular volume without the cache

vgimportclone -i <vg> <bla>

# now have distinct volume group, VG UUID and PV UUID

So the problem is making dd backups of origin, perhaps dd backups should
be avoided, but for some purposes (such as system migration) file copies
are just

more work

in general, and can complicate things as well, for instance if there are
NTFS partitions or whatnot.

And disk images can be nice to have, in any case.

This was the use case basically.

Zdenek Kabelac

2018-03-26 10:46:01 UTC

Post by Gang He
Hi Xen,

This is normal because /dev/vdd contains metadata for vg1 which includes
now missing disk /dev/vdc .... as the PV is no longer the same.

This is normal because you haven't removed /dev/vdc from vg1 on
/dev/vdd, since it was detached while you operated on its vg.

When you removed vdd when it was not attached, the VG1 metadata on vdd
was not altered. The metadata resides on both disks, so you had
inconsistent metadata between both disks because you operated on the
shared volume group while one device was missing.
You also did not recreate PV on /dev/vdc so it has the same UUID as when
it was part of VG1, this is why VG1 when VDD is booted will still try to
include /dev/vdc because it was never removed from the volume group on
VDD.
/dev/vdc contains volume group info for VG2 and includes only /dev/vdc
/dev/vdd contains volume group info for VG1, and includes both /dev/vdc
and /dev/vdd by UUID for its PV, however, it is a bug that it should
include /dev/vdc even though the VG UUID is now different (and the name
as well).

Hi

I'm not really sure what are you trying to achieve - are you 'validating' that
you cannot foolish lvm2 too easily or something else ?

Simple case is when you have a VG on 2 PV disks. Both PV hold full metadata
for a VG. There are numerous other case - i.e. you can have 1000PVs in single
VG then any update of metadata would require to update 1000 disks - for this
case you can select lower number metadata copies - i.e. randomly or
user-selected PVs only hold VG metadata and rest of PV are without metadata.
The less metadata copies - the less secure it is, but update is faster...

There are no metadata for use stored in your filesystem - VG metadata are
always recorded in PV metadata area.

1.) So when you 1st. remove device and then you run 'pvremove' on this missing
disk, it's kind of pointless operation.

2.) lvm2 command will not let you 'easily' remove PV which is in use by some
LV in your VG

3.) lvm2 supports 2 commands:
'vgreduce --removemissing' (try to make consistent VG when PV is lost)
'vgextend --restoeremissing' (restore missing PV back into VG)

Regards

Zdenek

Gang He

2018-05-18 04:56:59 UTC

Hi Zdenek and Guys,

As you know, we prefer to consider this problem as mis-operation.
But, the customer want to know if we should add a warning in this case?
Any comments for this customer's request?

Thanks
Gang

Post by Gang He
Hi Xen,

This is normal because /dev/vdd contains metadata for vg1 which includes
now missing disk /dev/vdc .... as the PV is no longer the same.

This is normal because you haven't removed /dev/vdc from vg1 on
/dev/vdd, since it was detached while you operated on its vg.

When you removed vdd when it was not attached, the VG1 metadata on vdd
was not altered. The metadata resides on both disks, so you had
inconsistent metadata between both disks because you operated on the
shared volume group while one device was missing.
You also did not recreate PV on /dev/vdc so it has the same UUID as when
it was part of VG1, this is why VG1 when VDD is booted will still try to
include /dev/vdc because it was never removed from the volume group on
VDD.
/dev/vdc contains volume group info for VG2 and includes only /dev/vdc
/dev/vdd contains volume group info for VG1, and includes both /dev/vdc
and /dev/vdd by UUID for its PV, however, it is a bug that it should
include /dev/vdc even though the VG UUID is now different (and the name
as well).

Fran Garcia

2018-03-26 10:17:49 UTC

This is normal because /dev/vdd contains metadata for vg1 which includes
now missing disk /dev/vdc .... as the PV is no longer the same.

Xen

2018-03-26 11:33:22 UTC

Post by Gang He
It looks like each PV includes a copy meta data for VG, but if some PV
has changed (e.g. removed, or moved to another VG),
the remained PV should have a method to check the integrity when each
startup (activated?), to avoid such inconsistent problem
automatically.

Well I'm not really interested in "what should be", I just answered your
question.

Also automatic resolution of differences turns into an automatic (or
bidirectional) "sync" problem which generally depends on consistent time
AND an agreed-upon strategy for merging conflicts which I doubt is truly
a realistic proposal because you are talking about data loss here.

I'm sorry, I don't think (I'm not an LVM developer here) what you ask
for is theoretically [im]possible.

At best you could have a "best merge" strategy with all associated
risks, and I don't think you should even go that route, honestly.

It's more something for an individual administrator to solve with
appropriate tools.

Nir Soffer

2018-10-16 20:59:45 UTC