[linux-lvm] converting a filesystem to a logical volume

[Heinz Mauelshagen]

> Bert Hubert writes:
> > I was wondering.. I think it should be fairly easy to convert a filesystem
> > into a logical volume - 'in place'.
> > 
> > Do people think that this is possible? If shifting a filesystem a bit is
> > enough, I'll start writing it.
> 
> Lennert Buytenhek and I were discussing this a bit, and it shouldn't be
> _too_ hard to do with the existing ext2resize code (although ext2resize
> isn't currently designed to move the start of a filesystem, only the end).
> This would be useful not only for LVM, but also for partition modification,
> which the parted folks are working on.  It may be that they even have
> some of this working already, but I haven't looked at their code yet.
> 
> What needs to be done is:
> 
> 0) figure out the number of disk blocks, N, needed at the start of a
>    partition for the LVM data structures

I can change pvcreate to provide this information
(without creating the volume for sure) and vgcreate/vgextend to wrap
the new locations of the filesystem.
There has to be some math because the LVM metatdat doesn't exactly fit
the start of a block group though.

Heinz


>    (N should be an even multiple
>    of 8 filesystem blocks to make this easy, as the block bitmaps are
>    packed into bytes and it would be extra work to re-map the block
>    bitmaps, although not impossible).
> 
>    if N < 3+num_inode_blocks+num_group_desc_blocks
>       this would be a lot of work, maybe you should backup/restore instead?
>       exit 1
> 
> 1) shrink existing filesystem by N blocks (easily done with existing code)
> 
> 2) mark data blocks, n, where n < N, for relocation
>    for each group in the ext2 filesystem
>       for each metadata block, m..num_metadata_blocks, in group
>          mark data block (m+N) for relocation
>    relocate all marked blocks (easily done with existing code)
> 
> Now for the new part for moving the start of the filesystem:
> 
> 3) for each group, g, in the ext2 filesystem
>       if group has superblock
>          copy superblock, s, to s+N
> 	 for each group descriptor in desc block d..d+num_group_desc_blocks
> 	    copy descriptor to block d+s and subtract N from
> 	       inode/block bitmap and inode table pointers
>       copy inode bitmap, i, to i+N
>       copy block bitmap g bytes [N/8,end) to new bitmap g [0,end-N/8)
>       copy block bitmap g+1 bytes* [0,N/8) to new bitmap g [end-N/8,end)
>          mark old metadata blocks from group g+1 as free in block bitmap g
>       mark new metadata blocks from group g as used in block bitmap g
> 
> * when g+1 is past the last group, copy zeros to the block bitmap, since you
>   freed N blocks at the end of the filesystem in step (1).
> 
>       for each inode in inode block i..i+num_inode_blocks
>          copy inode to block (i+N), subtract N from data block numbers
> 	 for each (single/double/triple) indirect block in old inode
> 	    allocate new free** block as indirect block, store in new inode
> 	    copy old indirect block to new indirect block and
> 	       subtract N from each data block number
> 	    mark old indirect block free in new block bitmap g
> 
> ** make sure the new indirect block is not one of the old metadata blocks
> 
> Basically, you have created a full new set of filesystem metadata while
> leaving the old metadata untouched.  You probably need less than 5% free
> space in the filesystem (enough for a copy of the metadata, which is fixed
> for a given filesystem, plus any indirect blocks, which depend on the size
> of the files in the filesystem, but usually there are not a huge number).
> 
> By making new copies of the indirect blocks, there is never a time when
> the old filesystem is not valid - this is a very good thing so you don't
> have problems if the program crashed, the machine crashed, you change your
> mind, etc.  You could always use the old filesystem until the time you
> overwrite the start of the old filesystem with LVM metadata.
> 
> If you are really interested in working on this, the ext2resize code is
> available via FTP or anonymous CVS from ext2resize.sourceforge.net.  Please
> feel free to ask me any questions, etc.
> 
> Cheers, Andreas
> -- 
> Andreas Dilger  \ "If a man ate a pound of pasta and a pound of antipasto,
>                  \  would they cancel out, leaving him still hungry?"
> http://www-mddsp.enel.ucalgary.ca/People/adilger/               -- Dogbert
> 

-- 

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

Systemmanagement TS                              T-Nova
                                                 Entwicklungszentrum Darmstadt
Heinz Mauelshagen                                Otto-Roehm-Strasse 71c
Senior Systems Engineer                          Postfach 10 05 41
                                                 64205 Darmstadt
mge at EZ-Darmstadt.Telekom.de                      Germany
                                                 +49 6151 886-425
                                                          FAX-386
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-