Understanding VMFS volumes

Understanding VMFS volumes is an important element within VMware ESX environments. When storage issues surface we need to correctly evaluate the VMFS volume states and apply the appropriate corrective actions to remediate undesirable storage events. VMFS architecture is not publically available and this certainly adds to the challenge when we need to correct a volume configuration or change issue. So lets begin to look at the components of a VMFS from what I have been able to decrypt using direct analysis.

All VMFS volume partitions will have a partition ID value of fb. Running fdisk can identify any partitions that are flagged as VMFS as shown here.

[root@vh1 ]# fdisk -lu /dev/sdc

Disk /dev/sdc: 274.8 GB, 274877889536 bytes
255 heads, 63 sectors/track, 33418 cylinders, total 536870878 sectors
Units = sectors of 1 * 512 = 512 bytes

   Device Boot    Start       End    Blocks   Id  System
/dev/sdc1           128 536860169 268430021   fb  Unknown

What’s important to note here is the sector size = 512 and the starting/ending blocks.

Many VMFS volume configuration elements are visible in the /vmfs mount folder. Within the directory the are two subdirectories, the volumes directory and the devices directory. The volumes directory provisions the mount point location and the devices directory holds configuration elements. Within the devices directory the are several subdirectories of which I can explain the disks and lvm folders, the others are not known to me outside of theory only.

A key part of a VMFS volume is it’s UUID (aka Universally Unique Identifier) and as the name suggests it used to ensure uniqueness when more than one volume is in use. The UUID is generated on the initial ESX host that created the VMFS volume based on the UUID creation standards. You can determine which ESX host the initial VMFS volume was created on by referring to the last 6 bytes of the UUID. This value is the same as the last six bytes of the ESX host’s system UUID found in the /etc/vmware/esx.conf file.

By far one of the most critical elements on a VMFS volume is the GUID. The GUID is integral within the volume because it is used to form the vml path (aka virtual multipath link). The GUID is stored within the VMFS volume header and begins at address 0x10002E.

The format of the GUID can vary based on different implementations of SCSI transport protocols but generally you will see some obvious length variances of the vml path identifiers which stem from the use of T11 and T10 Standard SCSI address formats like EUI-64, and NAA 64. Regardless of those variables there are components outside of the GUID within the vml that we should take notice of. The vml construct contains references to the LUN and partition values and these are useful to know about. The following illustrates where these elements appear in some real examples.

When we issue an ls -l from the /vmfs/devices/disks directory the following info is observed.

vhhba#:Target:LUN:Partition -> vml:??_LUN_??_GUID:Partition

                       LUN     GUID                         PARTITION
                       ^       ^                            ^
vmhba0:1:0:0  -> vml.02000000005005076719d163d844544e313436
vmhba0:1:0:1  -> vml.02000000005005076719d163d844544e313436:1
vmhba32:1:3:0 -> vml.0200030000600144f07ed404000000496ff8cd0003434f4d535441
vmhba32:1:3:1 -> vml.0200030000600144f07ed404000000496ff8cd0003434f4d535441:1

As well the issuing ls -l on the /vmfs/volumes list the VMFS UUID’s and the link name which is what we see displayed in the GUI client. In this example we will follow the UUID shown in blue and the named ss2-cstar-zs0.2 volume.

ss2-cstar-zs0.2 -> 49716cd8-ebcbbf9a-6792-000d60d46e2e

Additionally we can use esxcfg-vmkhbadevs -m to list the vmhba, dev and UUID associations.

[root@vh1 ]#  esxcfg-vmhbadevs -m
vmhba0:1:0:1    /dev/sdd1                        48a3b0f3-736b896e-af8f-00025567144e
vmhba32:1:3:1   /dev/sdf1                        49716cd8-ebcbbf9a-6792-000d60d46e2e

As you can see we indeed have different GUID lengths in this example. We also can see that the vmhba device is linked to a vml construct and this is how the kernel defines paths to a visible SCSI LUN. The vml path hosts the LUN ID, GUID and partition number information and this is also stored in the volumes VMFS header. As well the header contains a UUID signature but this is not the VMFS UUID.

If we use hexdump as illustrated below we can see these elements in the VMFS header directly.

[root@vh1 root]# hexdump -C -s 0x100000 -n 800 /dev/sdf1
00100000  0d d0 01 c0 03 00 00 00  10 00 00 00 02 16 03 00  |                | <- LUN ID
00100010  00 06 53 55 4e 20 20 20  20 20 43 4f 4d 53 54 41  |  SUN     COMSTA| <- Target Label
00100020  52 20 20 20 20 20 20 20  20 20 31 2e 30 20 60 01  |R         1.0 ` | <- LUN GUID
00100030  44 f0 7e d4 04 00 00 00  49 6f f8 cd 00 03 43 4f  |D ~     Io    CO|
00100040  4d 53 54 41 00 00 00 00  00 00 00 00 00 00 00 00  |MSTA            |
00100050  00 00 00 00 00 00 00 00  00 00 00 02 00 00 00 fc  |                | <- Volume Size
00100060  e9 ff 18 00 00 00 01 00  00 00 8f 01 00 00 8e 01  |                |
00100070  00 00 91 01 00 00 00 00  00 00 00 00 10 01 00 00  |                |
00100080  00 00 d8 6c 71 49 b0 aa  97 9b 6c 2f 00 0d 60 d4  |   lqI    l/  ` |
00100090  6e 2e 6e 89 19 fb a6 60  04 00 a7 ce 20 fb a6 60  |n n    `       `|
001000a0  04 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |
001000b0  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |
00100200  00 00 00 f0 18 00 00 00  90 01 00 00 00 00 00 00  |                |
00100210  01 00 00 00 34 39 37 31  36 63 64 38 2d 36 30 37  |    49716cd8-607| <- SEG UUID in ASCII
00100220  35 38 39 39 61 2d 61 64  31 63 2d 30 30 30 64 36  |5899a-ad1c-000d6|
00100230  30 64 34 36 65 32 65 00  00 00 00 00 00 00 00 00  |0d46e2e         |
00100240  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |
00100250  00 00 00 00 d8 6c 71 49  9a 89 75 60 1c ad 00 0d  |     lqI  u`    | <- SEG UUID
00100260  60 d4 6e 2e 01 00 00 00  e1 9c 19 fb a6 60 04 00  |` n          `  |
00100270  00 00 00 00 8f 01 00 00  00 00 00 00 00 00 00 00  |                |
00100280  8e 01 00 00 00 00 00 00  64 cc 20 fb a6 60 04 00  |        d    `  |
00100290  01 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |
001002a0  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |


In addition to the VMFS header block we have the hidden metadata files of the volume which you can list using ls -al. The vh.sf contains the UUID of the VMFS store and any member segments info. (I would presume the name vh stands for Volume Header … ;D) 



[root@vh1 ]# hexdump -C -s 0x200000 -n 256 /vmfs/volumes/49716cd8-ebcbbf9a-6792-000d60d46e2e/.vh.sf
00200000  5e f1 ab 2f 04 00 00 00  1f d8 6c 71 49 9a bf cb  |^   /  lqI      | <- VMFS UUID
00200010  eb 92 67 00 0d 60 d4 6e  2e 02 00 00 00 73 73 32  |  g  ` n     ss2| <- Volume Name
00200020  2d 63 73 74 61 72 2d 7a  73 30 2e 32 00 00 00 00  |-cstar-zs0.2    |
00200030  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |
00200090  00 00 00 00 00 00 00 00  00 00 00 00 00 00 02 00  |                |
002000a0  00 00 00 10 00 00 00 00  00 d8 6c 71 49 01 00 00  |          lqI   |
002000b0  00 d8 6c 71 49 9a 89 75  60 1c ad 00 0d 60 d4 6e  |  lqI  u`    ` n| <- SEG UUID
002000c0  2e 01 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |
002000d0  00 00 00 01 00 20 00 00  00 00 00 01 00 00 00 00  |                |
002000e0  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |

And of course we can not leave out the partition entry block data for the device.

hexdump -C -n 256 /dev/sdf1

00000000  fa b8 00 10 8e d0 bc 00  b0 b8 00 00 8e d8 8e c0  |                |
00000010  fb be 00 7c bf 00 06 b9  00 02 f3 a4 ea 21 06 00  |   |         !  |
00000020  00 be be 07 38 04 75 0b  83 c6 10 81 fe fe 07 75  |    8 u        u|
00000030  f3 eb 16 b4 02 b0 01 bb  00 7c b2 80 8a 74 01 8b  |         |   t  |
00000040  4c 02 cd 13 ea 00 7c 00  00 eb fe 00 00 00 00 00  |L     |         |
00000050  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |
000001b0  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 02  |                |
000001c0  03 00 fb fe ff ff 80 00  00 00 72 ef bf 5d 00 00  |          r  ]  | Type Start End
000001d0  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  |                |

With this detailed information it is possible to solve some common security issues with VMware stores like volume deletion and unintentional LUN ID changes.

Recently VMware added a some what useful command line tool named vmfs-undelete which exports metadata to a recovery log file which can restore vmdk block addresses in the event of deletion. It’s a simple tool and at present it’s experimental and unsupported and is not available on ESXi. The tool of course demands that you were proactive and ran it’s backup function in order to use it. Well I think this falls well short of what we need here. What if you have no previous backups of the VMFS configuration, so we really need to know what to look for and how to correct it and that’s exactly why I created this blog.  

The volume deletion event is quite easy to fix and thats simply because the VMFS volume header is not actually deleted. The partition block data is what gets trashed and you can just about get way with murder when it comes to recreating that part. Within the first 128 sectors is the peice we need to fix. One method is to create a store with the same storage volume and then block copy the partition to a file which can be block copied to the deleted device partition data blocks and this will fix the issue.

For example we create a new VMFS store on the same storage backing with the same LUN size as the original and it shows up as a LUN with a device name of /dev/sdd we can use esxcfg-vmhbadevs -m to find it if required

The deleted device name was /dev/sdc

We use the dd command to do a block copy from the new partition to a file or even directly in this case.

Remember to back it up first!

dd if=/dev/sdc of=/var/log/part-backup-sdc-1st.hex  bs=512 count=1

then issue

dd if=/dev/sdd of=/dev/sdc bs=512 count=1


dd if=/dev/sdd of=/var/log/part-backup-sdd.hex bs=512 count=1

dd if=/var/log/part-backup-sdd.hex of=/dev/sdc bs=512 count=1

I personally like using a file to perform this function as this becomes a future backup element which you can move to a safe location. The file can actually be edited with other utilities to provide more flexibility. e.g. hexedit etc. Addtitionally you could use fdisk to directly edit the partition table and provide the correct start and end addresses. This is something you should only do if you are well versed in it’s usage.

As as an additional level of protection we could even include making backups of the vh.sf metadata file and the VMFS header.

cp /vmfs/volumes/49716cd8-ebcbbf9a-6792-000d60d46e2e/.vh.sf /var/log/vh.sf.bu

dd if=/dev/sdc of=/var/log/vmfsheader-bu-sdc.hex bs=512 count=4096

This would grant the ability for support to examine the exact details of the VMFS configuration and potentially allow recovery from more complex issues.

One of the most annoying security events is when a VMFS LUN get’s changed inadvertently. If a VMFS volume LUN ID changes and is presented to an ESX host then the presented volume will be treated as a potential snapshot LUN. If this event occurs and the ESX servers advanced LVM parameter settings are at default the ESX host will not mount the volume. This behaviour is to prevent the possibility of corruption and downing the host since it can not determine which VM metadata inventory is correct.

If you are aware that the LUN ID has changed then the best course of action is to re-establish the correct LUN ID at the storage server first and rescan the affected vmhba’s. This is important because if you need to resignature the VMFS volume it will also require that the VM’s be imported back into inventory. Virtual Center logging and other various settings will be lost when this action is performed. This is a result of now having an incorrect UUID between the metadata, mount location and the vmx file UUID value.

If the storage change cannot be reverted back then a VMFS resignature method is the only option for reprovisioning a VMFS volume mount.

This is invoked by setting the LVM.DisallowSnapshotLun = 0 and LVM.EnableResignature = 1 and these should reverted back once the VMFS resignature operation is complete.



Tags: , , , , , , , , , , , , , , , , , ,

Site Contents: © 2009  Mike La Spina


  • Dai Nan says:

    Hi , I want to know is there anyway to restore vmdk file ? i delete the whole vm files using VC client , no vmdk-flat and nothing else, Is it possible to restore these files someway ? thank you very much

  • Hi,

    I’m sorry.
    You can only restore from a backup or use vmfs-undelete and both methods require a proactive function prior to a delete call.
    There are currently no tools that I am aware of which can recover the metadata and block allocation info after the security event occurs without some proactive steps prior to the incidents occurrence.



  • Arno says:

    Hi Mike, thanks for the info, it’s great. I have a question about the distributed file system nature of VMFS. In a VI environment, do you know if when a LUN gets detected as a snapshot for any reason by a host, say with no VMs running on it, is there a security mechanism in VMFS which would prevent say fdisk being run or vmkfstools -C to create a new VMFS on it when that VMFS is being accessed by other ESX hosts which have running VMs on it?

  • Hi Arno,

    The LUN will be at the mercy of the Administrator, fdisk will be able to write to it and you will be able to overwrite any VMFS header that is exposed to the host in question.



  • […] expliqué il y a longtemps par Mike La Spina sur son blog et lors d’une session dédiée au VMFS au VMworld 2007, il est possible de sauvegarder les […]

  • Raza says:

    This is really a very good Article, Thank Mike for sharing……………….
    All the best wishes for you….. 

  • Mike_Andreev says:

    we have a problem with vmdk file on vmfs volume – after hard reset (ESX server hang up) we could not acces vmdk file – when we try coping it on other host (storage) we got “invalid argument” message. We try remove lock from this vmdk file, but owner of file creator is dead, so how we could change UUID of vmdk file or remove lock from it ?

  • Hi Mike,

    Try editing the VM’s vmx file and set the following parm:
    Then see if it starts.

    As well I would try removing the VM from inventory and adding it back again. Another approach would be to connect your vCenter client to an ESX host directly which would avoid vCenter Server safety checks on the vmdk operations.

    It should not be necessary to change the uuid.



  • Mike_Andreev says:

    Mike, here more details of our problem:

    Host were was that storage is dead now 😉
    We install new ESX server and attach drive from dead ESX to new.
    Do with locked storage:
    vmkfstools -D 1C-SQL_1-flat.vmdk

    in /var/log/vmkernel we see –

    Lock [type 10c00001 offset 61200384 v 369, hb offset 3600384
    gen 11, mode 1, owner 4c36c4a5-981d4706-2cc6-0025900154ad mtime 1094]

    but old ESX host is gone….
    UUID of our ESX –
    Current /system/uuid = “4c45a358-adb8-8358-c665-0025900154ac”
    How we can unlock vmdk ?
    read file direct is impossible – we got “invalid argument”
    1) Could we change UUID of current disk on UUID which was in moment of died ? – it was 4c36c4a5-981d4706-2cc6-0025900154ad
    2) Could I directly open fb VMware VMFS, find place with locker owner and write zeros than ?
    3) Could we install new ESX and change it UUID on UUID which was on old ESX when file locked?


  • Hi Mike,

    The lock is not issued on the vmdk file directly. VMFS volumes use the metadata files to maintain cluster state info. Since the lock state is the result of an unfinished clustered file operation you need to issue a breaklock at the VMFS device level.

    This operation is available using the vmkfstools as follows:

    # vmkfstools -B /vmfs/device/disk/vml.uuid


  • nosh says:

    hi mike…. lovely and really good learning post.. i would like to further deep dive in understanding how is the magic number/address of the header calculated. also i am interested in further using of Hexdump especially for vmfs….

  • Koen says:

    Hi Mike and Mike Andreev,

    I’m having the same problem as Mike Andreev has. The command ‘vmkfstools -B …’ didn’t help. Apparently there is corruption in the metadata on the disk. Is there any way to solve this?

    Mike Andreev,

    Were you able to resolve the problem?


  • […] La verdad no voy a describir VMFS ya que me tomaría un tiempo demasiado largo, les recomiendo leer este articulo http://blog.laspina.ca/ubiquitous/understanding-vmfs-volumes. […]

  • Ravi says:

    Hi Mike!

    We have problem with ESX 4.0 server. ESX 4.0 hand up with purple screen error, and VMs were not able to ping or connect. Unfortunately we reloaded with over wrting vmfs. Is there is any method to recover VMs.


  • Hi Ravi,

    It is possible to recover vmdks from a vmfs file system however it’s not normally done when the meta files are destroyed. Initializing a vmfs volume on install creates new meta files and overwrites the block allocation tables. So unless you made a copy of them it’s very unlikely that the vmdks are recoverable.


  • Michal says:

    I have the same problem as Mike Andreev and Koen. Were you able to solve this problem. Can you give me a hint please.



  • Hi Michal,

    If the command ‘vmkfstools -B …’ didn’t help then try cloning the vmdk file using the vmkfstools utility. Typically a lock is placed on a vmdk to allow safe writes from one cluster member so a read should be allowed. If that works you can rename the files and use the clone instead.


  • Michal says:

    Hi Mike,
    thanks for the hint, this was something I didn’t try.
    After I tried to clone it, I have got the following error:

    vmkfstools -i /vmfs/volumes/virtuals/Debian/Debian.vmdk /vmfs/volumes/datastore2/Debian/debian.vmdk
    Destination disk format: VMFS zeroedthick
    Failed to open ‘/vmfs/volumes/virtuals/Debian/Debian.vmdk’: Invalid argument (1441801).

    I’m starting to think, that there si a data problem with the files that were open during run of virtual machine…but I still hope, there is a possibility to recover the data.



  • Hi Michal,

    I think you need to try a cp of the Debian-flat.vmdk and Debian.vmdk files to your datastore2 volume from a ssh connection of the ESXi host. vmkfstools seems to be cranky.


  • Michal says:

    Hi Mike,
    I tried that, also tried to install new esxi and connect the datastore there. I always get and “Invalid argument” error message…cp, vi, scp copy…

    Don’t you know about some vmfs data recovery, something that can raw copy the data?


  • Michal,

    Thats not good. If cp returns invalid argument it means that the vmfs path is not recognized as a valid file system. VMFS is damaged. You may have luck using the open source java based vmfs product.
    You cannot run it on ESXi, you will need ESX so you can install it and java.


  • Steve says:


    I hope you can help. I had a Dell server with Perc4i raid. The system was setup Raid1 on the front (datastore1) and 6x36GB Raid5 for (datastore2). I had 2 drives go down in the datastore2 taking all of my VMs with it. I forced one of the drives back online and then did a rebuild on the 2nd drive which seems to be successful. During the drive problems I rebooted the ESX 3.5 server and it saved a new configuration that did not include datastore2. I have tried resignature on the datastore but ESX still does not see the volume. I can see the LUN and doing fdisk -lu /dev/sdb does show there is still a partition of fb type. The /vmfs/volumes folder only has mapping for datastore1. Rescanning using both the VMclient and commandline vmkfstools -s does not show the volume in the VMclient. Any ideas on my next step. At this point I just want to be able to mount the partition and recover the data off it. I have tried the java vmfs driver with Knoppix live cd and no luck. Any help is appreciated.

  • Hi Steve,

    It’s quite possible that the metadata is fubar however there is some promise since it does have an fb partition. I would try creating a USB ESXi 4 boot stick and try adding the store from that disk. Or you could throw in a scratch disk an install ESX and then bring in the Perc. If the volume has usable metadata ESX 4 will give you the option to resig it on the add function.


  • Cuttino says:

    Hi Mike,
    I delete .vmdk file from esx host,any way to recover this file? or any sugestion?

  • Sorry,
    There really is no easy way to recover it. You must have a previous copy of the allocation table to find the first block pointer for the vmdk file and unless you have that its a needle in the hay stack. Backend snapshots maybe…

  • […] wanted to check how the mounted vmfs looked like. There is actually a good blog about this “Understanding VMFS volumes“, from that […]

  • […] so in this case we would expect to see it start at 1MB + 64KB from the begginning of the disk. This blog post goes over the details of the […]

  • Al says:

    Hi Mike,

    you must be getting tired of responding to questions about recovering vmdk files.. sorry… You mentioned that you need previous copy of the allocation table otheriwse its a needle in the hay stack. Does this then make it impossible or just a really, really long/dificult process. if it’s too hard manually, know any 3rd party software that could pull it off?

    Many thanks

  • Hi,

    The issue is no different than any chains of allocation, you need to link everything back to the allocation tables. Thus you either have it backed up or you use some stiching software like http://www.diskinternals.com/vmfs-recovery/ .


  • […] both the same UUID and Datastore name as the previous re-signatured store. with a bit of help from Mike La Spina we were able to find that the first 20496K of the block device on the SAN contained the signature […]

  • Robert says:

    Hello Mike
    thank you for your interesting documentation
    I’m dealing with HEX editor on a RAW device (something went bad on this hdd).
    Since there were some partitions, I’d like to find the correct offset where the VMFS begins.
    To do so with an hypothetical NTFS partition, I’d seek for an NTFSB string at offset 0 of every sector.

    Allow me to call it that “NTFSB” string the ntfs’ signature., partition offset is 63 sectors before that signature.

    What should I search for the VMFS volume/partition? If any string or HEX sequence is available, and supposing I find it, how many sectors backward will be the partition/volume offset?

    Thank you


  • Hi Robert,

    Are you trying to recreate the VMFS partition data?
    If so the most important parts are the type and starting sector.
    The ending sector could be larger and still allow access.

    Searching for a NTFS signature is not necessary. Just recreate the start and type If the VMFS header was overwritten then the datastore will not be detected.


  • Robert says:

    Hello Mike,
    thank you for replying

    NTFS was just my example to explain what I mean to search for an partition signature.

    I don’t know what to look for (HEX or either string value) for the VMFS, nether do I know if when it has been found… if any offset should be calculated with reference to that signature 🙂

    Thank you for hinting


  • Robert says:

    Hello Mike, I did it 🙂
    Now I have an additional step 😉

    There are 5 VMs in this datastore. A couple of them have an curious structure, there are two vmdk inside the directory, like this
    VMNAME-flat.vmdk size some 50GB
    VMNAME-000001-delta.vmdk size some 5GB

    sounds like some kind of snapshot but needed, since the “delta” have the correct date while the “flat” is dated more than 1 year ago.

    Now the point is: is there a way or an tool that allow to assembly the two vmdks? Or will it be possible only attaching the entire directory to an working vmware server?

    In few words, I do use one of those tools that allows to “explode” the vmdk and turn/mount it the contained file system, for browsing and files copying purposes only.

    It works with the “flat”, but again the NTFS content is stopped at some date in the past.

    The question becomes, is it needed an computer with a running vmware? Or something could be done at vmdk files level?

    Thank you for hinting


  • Robert says:

    P.S. as you may figure out, the tool doesn’t work with the “delta”.

  • Robert says:

    by the way, thank you for your useful documents.


Leave a Reply

XHTML: You can use these tags:
<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>