Re: Block alignment of qcow2 compress driver

[Hanna Reitz]

On 28.01.22 12:07, Richard W.M. Jones wrote:
> The commands below set up a sparse RAM disk, with an allocated block
> at offset 32K and another one at offset 1M-32K.  Then it tries to copy
> this to a compressed qcow2 file using qemu-nbd + the qemu compress
> filter:
>
>    $ qemu-img create -f qcow2 output.qcow2 1M
>    $ qemu-nbd -t --image-opts 
> driver=compress,file.driver=qcow2,file.file.driver=file,file.file.filename=output.qcow2
>  & sleep 1
>    $ nbdkit -U - \
>             data '@32768 1*32768 @1015808 1*32768' \
>             --run 'nbdcopy $uri nbd://localhost -p'
>
> The nbdcopy command fails when zeroing the first 32K with:
>
>    nbd://localhost: nbd_aio_zero: request is unaligned: Invalid argument
>
> This is a bug in nbdcopy because it ignores the minimum block size
> being correctly declared by the compress filter:
>
>    $ nbdinfo nbd://localhost
>    protocol: newstyle-fixed without TLS
>    export="":
>         export-size: 1048576 (1M)
>         uri: nbd://localhost:10809/
>         contexts:
>    ...
>                 block_size_minimum: 65536          <----
>                 block_size_preferred: 65536
>                 block_size_maximum: 33554432
>
> The compress filter sets the minimum block size to the the same as the
> qcow2 cluster size here:
>
>    
> https://gitlab.com/qemu-project/qemu/-/blob/cfe63e46be0a1f8a7fd2fd5547222f8344a43279/block/filter-compress.c#L117
>
> I patched qemu to force this to 4K:
>
> -    bs->bl.request_alignment = bdi.cluster_size;
> +    //bs->bl.request_alignment = bdi.cluster_size;
> +    bs->bl.request_alignment = 4096;
>
> and the copy above works, and the output file is compressed!
>
> So my question is, does the compress filter in qemu really need to
> declare the large minimum block size?  I'm not especially concerned
> about efficiency, I'd prefer it just worked, and changing nbdcopy to
> understand block sizes is painful.
>
> Is it already adjustable at run time?  (I tried using --image-opts
> like compress.request_alignment=4096 but it seems like the filter
> doesn't support anything I could think of, and I don't know how to
> list the supported options.)

I’m kind of amazed this works because the qcow2 driver rejects unaligned 
compressed writes[1].

And if I apply your diff, I can see that, too:
$ ./qemu-img create -f qcow2 test.qcow2 64M
Formatting 'test.qcow2', fmt=qcow2 cluster_size=65536 extended_l2=off 
compression_type=zlib size=67108864 lazy_refcounts=off refcount_bits=16

$ ./qemu-io -c 'write 0 32k' --image-opts \
driver=compress,file.driver=qcow2,file.file.driver=file,file.file.filename=test.qcow2
write failed: Invalid argument

So I actually don’t know why it works for you.  OTOH, I don’t understand 
why the block size affects you over NBD, because I would have expected 
qemu to internally auto-align requests when they are not aligned (in 
bdrv_co_pwritev_part()).  Like, when I set the NBD block driver’s 
alignment to 512[2], the following still succeeds:

$ ./qemu-img create -f qcow2 test.qcow2 64M
Formatting 'test.qcow2', fmt=qcow2 cluster_size=65536 extended_l2=off 
compression_type=zlib size=67108864 lazy_refcounts=off refcount_bits=16

$ ./qemu-nbd --fork --image-opts \
driver=compress,file.driver=qcow2,file.file.driver=file,file.file.filename=test.qcow2

$ ./qemu-io -c 'write 0 32k' -f raw nbd://localhost
wrote 32768/32768 bytes at offset 0
32 KiB, 1 ops; 00.00 sec (9.034 MiB/sec and 289.0960 ops/sec)


So I wonder why your diff works on your end, and I also wonder whether 
we couldn’t just have the NBD server expose some custom alignment 
(because qemu should auto-align all requests the NBD server generates 
anyway).

Hanna

[1] https://gitlab.com/qemu-project/qemu/-/blob/master/block/qcow2.c#L4662
[2] https://gitlab.com/qemu-project/qemu/-/blob/master/block/nbd.c#L1918