[linux-lvm] access or interface to list of blocks that have, changed via C.O.W.?

Mark Woodward markw at mohawksoft.com
Thu Oct 4 10:17:46 UTC 2012

I was going to try to answer in line, but decided that it would be too 
much work. There are utilities to extract the exception table out of the 
LVM2 snapshot, and if you can code in almost any language, you can write 
your own. It is dead simple. You can google for ddsnap and zumastore to 
get the code. It old and not supported currently, but still works. I 
have my own snapshotting differential backup running on my servers on a 
cron. (sorry, block level)

Essentially the COW device is formatted in blocks the size of the chunk 
size. The first chunk only contains a header. The second chunk contains 
an array of maps. This chunk is followed by all the data for the array. 
If the snapshot grows beyond this, a new chunk with an array is written.

The format of the array is simple: old_address (The offset in the 
volume) followed by the new_address (the offset in the COW device). An 
array of all the "old_address" values is the changed block list. You 
don't even need to worry about the data if you can really get a file 
list by blocks.

BTW: I would be curious how you do that and what file system supports it?

On 10/04/2012 01:05 AM, Linda Walsh wrote:
> Mark Woodward wrote:
>> On 10/03/2012 10:52 PM, Linda Walsh wrote:
>>> Mark Woodward wrote:
>>>> There are a couple projects that do this. They are pretty much 
>>>> based on ddsnap. You can google it.
>>>> In LVM2 world, it is fairly trivial to do what you want to do.
>>> ---
>>>    I figured it was likely -- I as LVM2 has to to know what blocks
>>> change to make realtime snapshots.  I just am trying to figure out how
>>> to get a list of those blocks -- can I query some util and get the 
>>> blocks
>>> that are different at that point?   I was figuring on using that with
>>> a blockmap of the fs, to get files that have changed, as I'm wanting 
>>> to export
>>> the files for smb(win client ) usage.
>> Well, I can honestly say that you are doing it the hard way. If you 
>> are connecting to a Linux box through samba, you can log the file 
>> changes.
> ----
>    Changes can come in via samba or locally so logging through samba 
> wouldn't
> cut it.
>>>> (1) create a virtual disk.
>>>> (2) take the "old" snapshot.
>>>> (3) write to lvdisk
>>>> (4) take the "new" snapshot.
>>>> At this stage the COW device of the "old" snapshot has all the data 
>>>> that has changed up to and including the "new" snapshot. You can 
>>>> back that up. As a differential. Then delete the "old" snapshot. 
>>>> The "new" snapshot is now renamed to the old snapshot.
>>> ----
>>>    Now here's a confusion -- back it up as a differential?  Do you
>>> mean from a backup utility or going from some list of blocks that 
>>> have changed?
>> I was talking about backing up the raw block level device.
> ----
>    I'm not sure that would work for me -- as I'm planning on just storing
> the files that have changed for ~ a month and rotating them out...
> That's why I'm hoping to get the block numbers that have changed -- if 
> I can
> map those to 1 or more inodes -- I could just back them up rather than 
> walking
> a 5Million+ file tree.
>>>> Take the next "new" snapshot. The renamed "old" snapshot has the 
>>>> changes since the previous snapshot up to and including the latest 
>>>> "new" snapshot. Just repeat this process, and you can do 
>>>> incremental backups of your LVM disks.
>>> ----
>>>    I'm sorta already doing the above -- it's just that I'm doing my 
>>> 'diff'
>>> with 'rsync' and it's dog-slow.  100-120 minutes for ~800GB 
>>> resulting in
>>> about 2.5G of diff.  Then I shuffle that off to another static vol 
>>> sized for
>>> the content -- and the 'cp' usually takes about 60-70 seconds.
>>>    What's hurting me is that "incremental backup" by having to scan 
>>> the file
>>> system.
>> The file system is the hard way.
> ----
>    Yep... tell me about it...
>>>> The biggest issue with performance is the COW aspect of snapshots. 
>>>> I have found using 64K chunk sizes greatly increase performance by 
>>>> reducing COW to snapshots. The default size if 4K.
>>> ----
>>>    I didn't know it was that low as a default -- but am using 64K 
>>> already --
>>> as that's my RAID's 'chunksize' (I thought about experimenting with 
>>> larger sizes, but would like it to run in a reasonable time first.
>>>    Also I a relevant question 0-- when I do a dmsetup list, I see a 
>>> bunch of
>>> cow volumes for drives that I **had** snaps going from at one 
>>> point.  Seems like
>>> the COW volumes didn't go away when halted...though it looks like, 
>>> from the dates, that maybe they get cleaned up at a boot(?)
>>> I only have 1 snapshot going but I see 14 cow partitions....looking 
>>> like
>>> VG-Home (254, 3)
>>> VG-Home--2012.09.30--00.52.54   (254, 50)
>>> VG-Home--2012.09.30--00.52.54-cow       (254, 51)
>>> VG-Home--2012.10.01--04.58.11   (254, 52)
>>> VG-Home--2012.10.01--04.58.11-cow       (254, 53)
>>> VG-Home--2012.10.02--07.22.14   (254, 54)
>>> VG-Home--2012.10.02--07.22.14-cow       (254, 55)
>>> VG-Home--2012.10.03--09.08.27   (254, 56)
>>> VG-Home--2012.10.03--09.08.27-cow       (254, 57)
>>> VG-Home-real    (254, 2)
>>> So would those be the list of blocks that changed upto the point they
>>> were halted?
>>> Do I need to worry about those "cow" vols taking up space?
>> If they are active, not only are they taking up space, but they are 
>> also being updated with every write.
> ----
>    I doubt that -- their corresponding snapshot volumes are gone
> Odd that they cow volumesdon't go as well.
>    When I delete an active snap, I first remove the snapshot volume
> with dmsetup remove -- then I lvremove it.  Seems to work without me 
> getting
> warnings about removing an active volume -- I'd assume when I lvremove'd
> a snap, that'd be all I need to do...
>    FWIW -- only cow files are the ones since the last reboot. That's 
> why I
> wondered if a reboot cleaned out the spurious entries..
>    I try to only keep one active snapshot going at a time due to the 
> write
> penalty...

More information about the linux-lvm mailing list