5

I had problem with my system (faulty internal power cable). When I got the system back up and running, arrays rebuilding, etc, I seem to have a situation where the pvs command (and vgs and lvs) reports No device found for PV <UUID> but the logical volume which is on the supposedly missing physical volume can be successfully mounted as their DM devices exist and are mapped in /dev/mapper.

The PV device is an md-raid RAID10 array, which seems fine, except for the confusion that it's not appearing in the pvs output.

I assume this is a problem with some internal tables being out of sync. How do I get things mapped correctly (wthout a reboot, which, I assume would fix it)?

Update:

A reboot did NOT fix the problem. I believe the issue is due to the configuration of the 'missing' PV (/dev/md99) as a RAID10 far-2 array built from a 750b disk (/dev/sdk) and a RAID0 array (/dev/md90) build from a 250GB disk (/dev/sdh) and a 500Gb disk (/dev/sdl). It seems from the output of pvscan -vvv that the lvm2 signature is found on /dev/sdh, but not on /dev/md99.

    Asking lvmetad for VG f1bpcw-oavs-1SlJ-0Gxf-4YZI-AiMD-WGAErL (name unknown)
  Setting response to OK
  Setting response to OK
  Setting name to b
  Setting metadata/format to lvm2
    Metadata cache has no info for vgname: "b"
  Setting id to AzKyTe-5Ut4-dxgq-txEc-7V9v-Bkm5-mOeMBN
  Setting format to lvm2
  Setting device to 2160
  Setting dev_size to 1464383488
  Setting label_sector to 1
    Opened /dev/sdh RO O_DIRECT
  /dev/sdh: size is 488397168 sectors
    /dev/sdh: block size is 4096 bytes
    /dev/sdh: physical block size is 512 bytes
    Closed /dev/sdh
  /dev/sdh: size is 488397168 sectors
    Opened /dev/sdh RO O_DIRECT
    /dev/sdh: block size is 4096 bytes
    /dev/sdh: physical block size is 512 bytes
    Closed /dev/sdh
    /dev/sdh: Skipping md component device
  No device found for PV AzKyTe-5Ut4-dxgq-txEc-7V9v-Bkm5-mOeMBN.
    Allocated VG b at 0x7fdeb00419f0.
  Couldn't find device with uuid AzKyTe-5Ut4-dxgq-txEc-7V9v-Bkm5-mOeMBN.
    Freeing VG b at 0x7fdeb00419f0.

The only reference to /dev/md99, which should be the PV, is when it's added to the device cache.

Update 2:

Stopping lvm2-lvmetad and repeating the pvscan confirms than the issue is that the system is confused about which PVs to use as it's finding 2 with the same UUID

    Using /dev/sdh
    Opened /dev/sdh RO O_DIRECT
    /dev/sdh: block size is 4096 bytes
    /dev/sdh: physical block size is 512 bytes
  /dev/sdh: lvm2 label detected at sector 1
  Found duplicate PV AzKyTe5Ut4dxgqtxEc7V9vBkm5mOeMBN: using /dev/sdh not /dev/md99
    /dev/sdh: PV header extension version 1 found
  Incorrect metadata area header checksum on /dev/sdh at offset 4096
    Closed /dev/sdh
    Opened /dev/sdh RO O_DIRECT
    /dev/sdh: block size is 4096 bytes
    /dev/sdh: physical block size is 512 bytes
  Incorrect metadata area header checksum on /dev/sdh at offset 4096
    Closed /dev/sdh
    Opened /dev/sdh RO O_DIRECT
    /dev/sdh: block size is 4096 bytes
    /dev/sdh: physical block size is 512 bytes
    Closed /dev/sdh
  Incorrect metadata area header checksum on /dev/sdh at offset 4096
    Telling lvmetad to store PV /dev/sdh (AzKyTe-5Ut4-dxgq-txEc-7V9v-Bkm5-mOeMBN)
  Setting response to OK

since this configuration was only meant to be temporary, I think I'd do better to rearrange my disk usage.

Unless anyone can tell me how to explicitly override the order in which pvscan views devices?

StarNamer
  • 3,112
  • 1
  • 22
  • 33
  • Have you tried running another `pvscan` after the MD array comes up? Just in case it's a timing issue. – Bratchley May 17 '15 at 22:44
  • That would involve a reboot, which would probably fix it. I was hoping to just fix the current state, e.g.by restarting lvmetad (which didn't fix it). – StarNamer May 18 '15 at 10:40
  • `pvscan` is a command that you can run as root after it's booted. If this works we can work on resolving the timing issue. – Bratchley May 18 '15 at 11:30
  • I've been using LVM for about 5 years; I know what pvscan is. My point was that I wanted to avoid rebooting the machine since that will almost certainly fix the problem. What I wanted was some way to get LVM back to a consistent state without rebooting. I get the impression this isn't possible. Since LVs in the 'missing' VG can be mounted, I've taken the precaution of copying the data to a new set of LVs in a new VG and will probably reboot later this evening. I expect this will fix the issue, but won't provide info as to why it happened in the first place and if an online fix can be done. – StarNamer May 18 '15 at 15:12
  • @Bratchley, by the way, if you read my comment to derobert, you'd have realised that I've already tried a pvscan (several times). – StarNamer May 18 '15 at 15:16
  • Fine. Fix your own problem then. – Bratchley May 18 '15 at 17:07
  • @Bratchley: my point is that you were suggesting something I'd already tried. If you have a suggestion as to how I can get more detailed information, without simply rebooting, then I'd be happy to try it. Currently `pvscan -vvvv` (with or without lvm2-lvmetad running) doesn't reference the 'missing' RAID device except to add it to the device cache and note its aliases. `pvdisplay -vvvv /dev/md99` obviously finds the device, as it reports the correct UUID, VG, etc from libdm-config.c near the start, but then reports "Failed to find physical volume" at the end! – StarNamer May 18 '15 at 23:39

3 Answers3

3

The first thing to check are your filter and global_filter options in /etc/lvm/lvm.conf. Make sure you aren't filtering out the devices your PVs reside on.

The cache is set with the cache_dir option in the same file; on my Debian box it defaults to /run/lvm. The cache (if any) should be in that directory. If obtain_device_list_from_udev is set, I believe no cache is used.

Finally, check if use_lvmetad is set. If so, you may need to restart the LVM metadata daemon.

derobert
  • 107,579
  • 20
  • 231
  • 279
  • No filters and obtain_device_list_from_udev is 1. Stopping lvmetad and then doing a pvscan finda duplicate UUD of one of the component devices of the RAID device which make up the RAID10 aary, but the checksum is wrong. It then goes back to saying 'No device found...'. – StarNamer May 17 '15 at 18:31
1

The problem appears to be pvscan getting confused over seeing the same UUID on both a component device of the RAID array and the RAID array itself. I assume this is avoided normally by recognising that the device is a direct component. In my case, I had created a situation where the device was not directly a component of the RAID device which should be the PV.

My solution was to backup the LVs, force the array to be degraded and then reconfigure the disks so as not to use the multilevel RAID. Note that after another reboot the device lettering had changed. 500Gb = /dev/sdi, 250Gb = /dev/sdj, 750Gb = /dev/sdk

# mdadm /dev/md99 --fail --force /dev/md90
# mdadm /dev/md99 --remove failed
# mdadm --stop /dev/md90
# wipefs -a /dev/sdi /dev/sdj               # wipe components
# systemctl stop lvm2-lvmetad
# pvscan -vvv
# pvs
..... /dev/md99 is now correctly reported as the PV for VG b
# fdisk /dev/sdi
...... Create 2 partitions of equal size, i.e. 250Gb
# fdisk /dev/sdj
...... Create a single 250Gb patitiion
# mdadm /dev/md91 --create -lraid5 -n3 /dev/sdi1 /dev/sdj1 missing
# mdadm /dev/md92 --create -lraid1 -n2 /dev/sdi2 missing
# pvcreate /dev/md91 /dev/md92
# vgextend b /dev/md91 /dev/md92
# pvmove /dev/md99
# vgreduce b /dev/md99
# pvremove /dev/md99
# mdadm --stop /dev/md99
# wipefs -a /dev/sdk
# fdisk /dev/sdk
..... Create 3 250Gb partitions
# mdadm /dev/md91 --add /dev/sdk1
# mdadm /dev/md92 --add /dev/sdk2

Moral of the story:

Do not introduce too many levels of indirection into the filesystem!

StarNamer
  • 3,112
  • 1
  • 22
  • 33
0

As root, these commands will fix it:

pvscan --cache
pvscan
Jeff Schaller
  • 66,199
  • 35
  • 114
  • 250
user125971
  • 39
  • 4