4.1.0 bogus QCOW2 corruption reported after compress

[Toolybird]

Creating a compressed image then running `qemu-img check <..>.qcow2' on said image seems to report bogus corruption in some (but not all) cases:

Step 1.

# qemu-img info win7-base.qcow2
image: win7-base.qcow2
file format: qcow2
virtual size: 20 GiB (21474836480 bytes)
disk size: 12.2 GiB
cluster_size: 65536
Format specific information:
    compat: 1.1
    lazy refcounts: true
    refcount bits: 16
    corrupt: false

# qemu-img check win7-base.qcow2
No errors were found on the image.
327680/327680 = 100.00% allocated, 0.00% fragmented, 0.00% compressed clusters
Image end offset: 21478375424

Step 2.

# qemu-img convert -f qcow2 -O qcow2 -c win7-base.qcow2 test1-z.qcow2

Step 3.

# qemu-img info test1-z.qcow2
image: test1-z.qcow2
file format: qcow2
virtual size: 20 GiB (21474836480 bytes)
disk size: 5.78 GiB
cluster_size: 65536
Format specific information:
    compat: 1.1
    lazy refcounts: false
    refcount bits: 16
    corrupt: false

# qemu-img check test1-z.qcow2
ERROR cluster 1191 refcount=1 reference=2
ERROR cluster 1194 refcount=1 reference=4
ERROR cluster 1195 refcount=1 reference=7
ERROR cluster 1196 refcount=1 reference=7
ERROR cluster 1197 refcount=1 reference=6
ERROR cluster 1198 refcount=1 reference=4
ERROR cluster 1199 refcount=1 reference=4
ERROR cluster 1200 refcount=1 reference=5
ERROR cluster 1201 refcount=1 reference=3
<...> snip many errors
Leaked cluster 94847 refcount=3 reference=0
Leaked cluster 94848 refcount=3 reference=0
Leaked cluster 94849 refcount=11 reference=0
Leaked cluster 94850 refcount=14 reference=0

20503 errors were found on the image.
Data may be corrupted, or further writes to the image may corrupt it.

20503 leaked clusters were found on the image.
This means waste of disk space, but no harm to data.
197000/327680 = 60.12% allocated, 89.32% fragmented, 88.50% compressed clusters
Image end offset: 6216220672

The resultant image seems to work fine in a VM when used as a backing file.

Interestingly, if I substitute a qemu-img binary from qemu-4.0 then no errors are reported.

# /tmp/qemu-img check test1-z.qcow2
No errors were found on the image.
197000/327680 = 60.12% allocated, 89.32% fragmented, 88.50% compressed clusters
Image end offset: 6216220672

Is the image corrupted or not? I'm guessing not.

Just in case it matters, this is ext4 fs on rotational disk. Latest Arch Linux but self compiled 4.1.0 with recent QCOW2 corruption fixes added.

I haven't tried latest trunk but might do so if time permits.

[Toolybird (toolybird)]

Current trunk still displays the problem.

A git bisection between 4.0.0 and 4.1.0 revealed:

b6c246942b14d3e0dec46a6c5868ed84e7dbea19 is the first bad commit
commit b6c246942b14d3e0dec46a6c5868ed84e7dbea19
Author: Alberto Garcia <email address hidden>
Date: Fri May 10 19:22:54 2019 +0300

    qcow2: Define and use QCOW2_COMPRESSED_SECTOR_SIZE

    When an L2 table entry points to a compressed cluster the space used
    by the data is specified in 512-byte sectors. This size is independent
    from BDRV_SECTOR_SIZE and is specific to the qcow2 file format.

    The QCOW2_COMPRESSED_SECTOR_SIZE constant defined in this patch makes
    this explicit.

    Signed-off-by: Alberto Garcia <email address hidden>
    Signed-off-by: Kevin Wolf <email address hidden>

 block/qcow2-cluster.c | 5 +++--
 block/qcow2-refcount.c | 25 ++++++++++++++-----------
 block/qcow2.c | 3 ++-
 block/qcow2.h | 4 ++++
 4 files changed, 23 insertions(+), 14 deletions(-)

[Max Reitz (xanclic)]

There is definitely a bug in that patch, and that is that QCOW2_COMPRESSED_SECTOR_MASK is an unsigned int instead of a uint64_t (so the mask is too small).

It looks like the bug has existed in some places before that patch (because they use ~511 as a mask), but not in others.

This would explain why the bug is visible only for some images, namely for those with a compressed size of more than 4 GB, I presume.

And indeed, fixing QCOW2_COMPRESSED_SECTOR_MASK to be an unsigned long long fixes the bug. I’ll send a patch (but I’ll have to write a more simple and quicker test case first).

Max

[Max Reitz (xanclic)]

Forgot to say that I sent a patch:

https://lists.nongnu.org/archive/html/qemu-block/2019-10/msg01764.html

Thanks for reporting!

Max

[Toolybird (toolybird)]

Thank you Max.

Can confirm your patch fixes my issue (qemu-img check ...)

Not sure about those other code paths. I don't use internal snapshots and I'm not sure under which circumstances qcow2_free_any_clusters() gets exercised.

Just for good measure, with patch applied I created another >4GB compressed image then booted it a few times and all seems fine.

[Thomas Huth (th-huth)]

Fix has been included in QEMU v4.2:
https://git.qemu.org/?p=qemu.git;a=commitdiff;h=24552feb6ae2f615b76c2