Re: [PATCH 2/2] iotests: Test large write request to qcow2 file

[Eric Blake]

On 10/10/19 5:08 AM, Max Reitz wrote:
> Without HEAD^, the following happens when you attempt a large write
> request to a qcow2 file such that the number of bytes covered by all
> clusters involved in a single allocation will exceed INT_MAX:
>
> (A) handle_alloc_space() decides to fill the whole area with zeroes and
>      fails because bdrv_co_pwrite_zeroes() fails (the request is too
>      large).
>
> (B) If handle_alloc_space() does not do anything, but merge_cow()
>      decides that the requests can be merged, it will create a too long
>      IOV that later cannot be written.
>
> (C) Otherwise, all parts will be written separately, so those requests
>      will work.
>
> In either B or C, though, qcow2_alloc_cluster_link_l2() will have an
> overflow: We use an int (i) to iterate over nb_clusters, and then
> calculate the L2 entry based on "i << s->cluster_bits" -- which will
> overflow if the range covers more than INT_MAX bytes.  This then leads
> to image corruption because the L2 entry will be wrong (it will be
> recognized as a compressed cluster).
>
> Even if that were not the case, the .cow_end area would be empty
> (because handle_alloc() will cap avail_bytes and nb_bytes at INT_MAX, so
> their difference (which is the .cow_end size) will be 0).
>
> So this test checks that on such large requests, the image will not be
> corrupted.  Unfortunately, we cannot check whether COW will be handled
> correctly, because that data is discarded when it is written to null-co
> (but we have to use null-co, because writing 2 GB of data in a test is
> not quite reasonable).
>
> Signed-off-by: Max Reitz <address@hidden>
> ---
>   tests/qemu-iotests/268     | 83 ++++++++++++++++++++++++++++++++++++++
>   tests/qemu-iotests/268.out |  9 +++++
>   tests/qemu-iotests/group   |  1 +
>   3 files changed, 93 insertions(+)
>   create mode 100755 tests/qemu-iotests/268
>   create mode 100644 tests/qemu-iotests/268.out
>
> diff --git a/tests/qemu-iotests/268 b/tests/qemu-iotests/268
> new file mode 100755
> index 0000000000..b9a12b908c
> --- /dev/null
> +++ b/tests/qemu-iotests/268

> +# We want a null-co as the data file, because it allows us to quickly
> +# "write" 2G of data without using any space.
> +# (qemu-img create does not like it, though, because null-co does not
> +# support image creation.)
> +$QEMU_IMG amend -o data_file="json:{'driver':'null-co',,'size':'4294967296'}" \
> +    "$TEST_IMG"
> +

A bit awkward, but works.

> +# This gives us a range of:
> +#   2^31 - 512 + 768 - 1 = 2^31 + 255 > 2^31
> +# until the beginning of the end COW block.  (The total allocation
> +# size depends on the cluster size, but all that is important is that
> +# it exceeds INT_MAX.)
> +#
> +# 2^31 - 512 is the maximum request size.  We want this to result in a
> +# single allocation, and because the qcow2 driver splits allocations
> +# on L2 boundaries, we need large L2 tables; hence the cluster size of
> +# 2 MB.  (Anything from 256 kB should work, though, because then one L2
> +# table covers 8 GB.)
> +$QEMU_IO -c "write 768 $((2 ** 31 - 512))" "$TEST_IMG" | _filter_qemu_io

Yep, that causes the rounding issue that requires being able to handle > 
2G gracefully.

Reviewed-by: Eric Blake <address@hidden>

--
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org