Re: [Qemu-devel] Possibly incorrect data sparsification by qemu-img

[Kevin Wolf]

Am 24.04.2019 um 08:40 hat Vladimir Sementsov-Ogievskiy geschrieben:
> 23.04.2019 18:08, Kevin Wolf wrote:
> > Am 23.04.2019 um 16:26 hat Martin Kletzander geschrieben:
> >> On Tue, Apr 23, 2019 at 02:12:18PM +0200, Kevin Wolf wrote:
> >>> Am 23.04.2019 um 13:30 hat Martin Kletzander geschrieben:
> >>>> Hi,
> >>>>
> >>>> I am using qemu-img with nbdkit to transfer a disk image and the update 
> >>>> it with
> >>>> extra data from newer snapshots.  The end image cannot be transferred 
> >>>> because
> >>>> the snapshots will be created later than the first transfer and we want 
> >>>> to save
> >>>> some time up front.  You might think of it as a continuous 
> >>>> synchronisation.  It
> >>>> looks something like this:
> >>>>
> >>>> I first transfer the whole image:
> >>>>
> >>>>   qemu-img convert -p $nbd disk.raw
> >>>>
> >>>> Where `$nbd` is something along the lines of 
> >>>> `nbd+unix:///?socket=nbdkit.sock`
> >>>>
> >>>> Then, after the next snapshot is created, I can update it thanks to the 
> >>>> `-n`
> >>>> parameter (the $nbd now points to the newer snapshot with unchanged data 
> >>>> looking
> >>>> like holes in the file):
> >>>>
> >>>>   qemu-img convert -p -n $nbd disk.raw
> >>>>
> >>>> This is fast and efficient as it uses block status nbd extension, so it 
> >>>> only
> >>>> transfers new data.
> >>>
> >>> This is an implementation detail. Don't rely on it. What you're doing is
> >>> abusing 'qemu-img convert', so problems like what you describe are to be
> >>> expected.
> >>>
> >>>> This can be done over and over again to keep the local
> >>>> `disk.raw` image up to date with the latest remote snapshot.
> >>>>
> >>>> However, when the guest OS zeroes some of the data and it gets written 
> >>>> into the
> >>>> snapshot, qemu-img scans for those zeros and does not write them to the
> >>>> destination image.  Checking the output of `qemu-img map --output=json 
> >>>> $nbd`
> >>>> shows that the zeroed data is properly marked as `data: true`.
> >>>>
> >>>> Using `-S 0` would write zeros even where the holes are, effectively 
> >>>> overwriting
> >>>> the data from the last snapshot even though they should not be changed.
> >>>>
> >>>> Having gone through some workarounds I would like there to be another 
> >>>> way.  I
> >>>> know this is far from the typical usage of qemu-img, but is this really 
> >>>> the
> >>>> expected behaviour or is this just something nobody really needed 
> >>>> before?  If it
> >>>> is the former, would it be possible to have a parameter that would 
> >>>> control this
> >>>> behaviour?  If the latter is the case, can that behaviour be changed so 
> >>>> that it
> >>>> properly replicates the data when `-n` parameter is used?
> >>>>
> >>>> Basically the only thing we need is to either:
> >>>>
> >>>> 1) write zeros where they actually are or
> >>>>
> >>>> 2) turn off explicit sparsification without requesting dense image 
> >>>> (basically
> >>>>     sparsify only the par that is reported as hole on the source) or
> >>>>
> >>>> 3) ideally, just FALLOC_FL_PUNCH_HOLE in places where source did report 
> >>>> data,
> >>>>     but qemu-img found they are all zeros (or source reported HOLE+ZERO 
> >>>> which, I
> >>>>     believe, is effectively the same)
> >>>>
> >>>> If you want to try this out, I found the easiest reproducible way is 
> >>>> using
> >>>> nbdkit's data plugin, which can simulate whatever source image you like.
> >>>
> >>> I think what you _really_ want is a commit block job. The problem is
> >>> just that you don't have a proper backing file chain, but just a bunch
> >>> of NBD connections.
> >>>
> >>> Can't you get an NBD connection that already provides the condensed form
> >>> of the whole snapshot chain directly at the source? If the NBD server
> >>> was QEMU, this would actually be easier than providing each snapshot
> >>> individually.
> >>>
> >>> If this isn't possible, I think you need to replicate the backing chain
> >>> on the destination instead of converting into the same image again and
> >>> again so that qemu-img knows that it must take existing data of the
> >>> backing file into consideration:
> >>>
> >>>     qemu-img convert -O qcow2 nbd://... base.qcow2
> >>>     qemu-img convert -O qcow2 -F qcow2 -B base.qcow2 nbd://... 
> >>> overlay1.qcow2
> >>>     qemu-img convert -O qcow2 -F qcow2 -B overlay1.qcow2 nbd://... 
> >>> overlay2.qcow2
> >>>     ...
> 
> Is it safe in general?
> 
> Qemu often consider rounding-up allocated ranges to be safe, or just
> consider unknown area as allocated.  And if this happen, we'll convert
> unallocated hole to allocated zeroes on target, which will break
> backing chain.
> 
> This way would be correct if on source under nbd server we have valid
> backing chain too, so in case of "rounding-up" we'll just read valid
> data from backing. But it is not the case (or sorry, if I
> misunderstood).

As I said, it depends on the NBD server providing the right block
allocation status - and this includes alignment etc. as well.

It's not a very nice solution because NBD doesn't actually do backing
fields, so we're relying on things that the spec doesn't talk about. But
as I understand, we don't have control over the server side, so it's
probably the best we can do under these conditions.

If the NBD server already took the backing chain into consideration, it
would indeed be much more reliable.

Kevin